You can read the paper John published that spurred the conversation here:
https://www.linkedin.com/feed/update/urn:li:activity:7511570544782356480/
In this episode Shane Gibson sits down with data modeling veteran’s John Giles and Remco Broekmans
The conversation explores the premise of democratising data modeling through AI, centered around a paper published by John regarding where AI can and cannot add real value. The group highlights the critical importance of conceptual and context data modeling, stressing that business terms and definitions must be established upfront before looking at physical source tables. They strongly critique the anti-pattern of placing humans “at the end of the line” merely to rubber-stamp automated outputs, arguing instead for an “AI-in-the-loop” approach where technology serves strictly as an assistant under human direction.
The discussion also dives deep into complex real-world scenarios where automated tools fail—specifically when physical data or source systems do not yet exist. John shares extreme edge cases, such as establishing a brand-new government department from scratch based on a vague Act of Parliament, unifying over 70 distinct healthcare organisations into a single model, and modeling evolving visions like the Australian wool industry. Remco breaks down his multi-agent ELM framework, explaining how introducing deliberate manual friction points between LLM tasks keeps human modelers in control and why model-storming workshops, human trust, and organisational culture are far more vital than push-button automation.
Whether you are curious about conceptual modeling techniques, how to effectively integrate “AI in the loop” while avoiding push-button modeling pitfalls, or strategies for aligning business definitions across complex organisations, this episode is packed with valuable insights for data professionals.
Listen
Listen on all good podcast hosts or over at:
Subscribe: Apple Podcast | Spotify | Google Podcast | Amazon Audible | TuneIn | iHeartRadio | PlayerFM | Listen Notes | Podchaser | Deezer | Podcast Addict |
Tired of vague data requests and endless requirement meetings? The Information Product Canvas helps you get clarity in 30 minutes or less?
Google NotebookLM Mindmap
Google NoteBookLM Briefing
Executive Summary
This briefing document synthesizes key insights, arguments, and practical considerations from an industry discussion featuring John, Remco and Shane.
The central discussion focuses on whether Artificial Intelligence (AI) can democratize data modeling by autonomously generating the business definitions, context models, ontologies, and taxonomies it requires to process enterprise data effectively. While AI—specifically Large Language Models (LLMs)—excels at token prediction, parsing text, comparing frameworks, and acting as an “assistant,” the panel concludes that fully automated, “push-button” data modeling is an anti-pattern.
The primary takeaways include:
The “AI in the Loop” Paradigm: The popular concept of “Human in the Loop” often degenerates into an anti-pattern where AI performs the work and an uninformed human merely clicks “approve.” A more effective operational model is “AI in the loop,” where AI acts strictly as an assistant with explicit human control points between workflow handoffs.
Upfront Conceptual Modeling vs. Physical Reverse Engineering: Defining business concepts early—independent of physical source data structures—is critical. Attempting to generate data models solely by scanning physical tables or source systems replicates existing technical debt, vendor-enforced assumptions, and siloed terminology without reflecting true business semantics.
Edge Cases Beyond AI Capabilities: AI cannot construct valid conceptual models in scenarios where source data does not exist (e.g., brand-new entities established by legislation), where knowledge is evolving or vision-driven, or where data is strictly confidential (e.g., M&A, military, or stealth commercial products).
Methodological Convergence: Methodologies vary—from top-down generalized patterns (”Data Town Plans”) to bottom-up event matrix analysis—but all successful data modeling requires capturing business exceptions, accounting for organizational culture, and building human trust across operational silos.
Key Themes & Detailed Analysis
1. The Core Thesis: Can AI Generate the Definitions It Needs?
AI tools perform significantly better when supplied with clear, business-centric definitions of core concepts (e.g., customer, employee, product, order). However, relying on AI to autonomously derive these business definitions introduces a circular dependency.
[ Unstructured / Structured Enterprise Data ]
│
▼
[ AI Processing & Interpretation ]
│
┌──────────────┴──────────────┐
▼ ▼
Needs Business-Centric Attempts Autonomous
Definitions/Ontologies Generation of Concepts
│ │
└──────────────┬──────────────┘
▼
[ Risk: Hallucination & Amplification
of Physical Table / Silo Biases ]
The Business Context Problem: If AI is forced to guess business meanings from source systems alone, it restructures physical technical debt into new formats (e.g., Data Vault, Iceberg tables, or dbt streams) without determining whether the outcome is meaningful to business stakeholders.
Semantic Ambiguity: Identical terms often carry conflicting definitions across departments. For example, in a healthcare setting, a bed is defined as a place by hospital operational management (occupancy tracking) but as a thing/asset by housekeeping and maintenance teams. AI scanning a database cannot resolve these contextual divergences without human direction.
2. Operational Models: “Human in the Loop” vs. “AI in the Loop”
The panel critiques standard implementations of AI-driven automation, contrasting superficial approval chains with structured human facilitation.
Anti-Pattern: "Human at the End of the Line"
[ Data Input ] ──> [ Full AI Modeling Generation ] ──> [ Human "Approve" Button ] ──> [ Deployment ]
(Low Context/Understanding)
Preferred Pattern: "AI in the Loop / Assistant"
[ Human Context ] ──> [ AI Agent Task ] ──> [ Manual Review Gate ] ──> [ Next Iteration/Model ]
The “Human at the End of the Line” Anti-Pattern: A key flaw in modern AI integration occurs when an LLM performs 99% of the modeling work and presents the output to a human at the final stage. The human, lacking insight into the AI’s underlying logic, signs off without comprehending the structure.
Intentional Friction and Multi-Agent Workflows: To maintain control, practitioners introduce manual friction points between LLM processing stages. For example, Remco’s ELM Data Modeling Framework uses six distinct AI agent services (e.g., discovering core business concepts, identifying relationships/definitions, reviewing gaps against business cases) where data output from one stage must be manually evaluated and transferred to the next.
Comparative LLM Consolidation: Because LLMs are non-biased predictors, running identical business transcripts or workshop notes through multiple models (e.g., Claude, DeepSeek, ChatGPT, Gemini, Mistral) and applying a “compare and consolidate” service highlights variations and uncovers overlooked edge cases.
3. Upfront Conceptual Modeling vs. Physical Source Dependency
A central thesis of the discussion is that data modeling must occur before looking at physical databases or source systems.
+-----------------------------------------------------------------------------------+
| RECOMMENDED WORKFLOW |
| |
| 1. Scope & Boundary ──> 2. SME Interviews ──> 3. Business Event Matrix |
| │ |
| ▼ |
| 6. Physical Schema <── 5. Walk the Map <── 4. Concept/Context Model |
| Implementation (Validate Questions) (Nouns, Verbs, Definitions) |
+-----------------------------------------------------------------------------------+
Definitions Early and Well: Following principles from library sciences, establishing definitions upfront drives downstream physical design, BI metric layers, and integration architectures.
The Failure of Automated Source Harvesting: Historical attempts—from 2010-era ETL auto-modeling tools to modern dbt “blobocoder” or reverse-engineered semantic layers—focus on physical tables. Scanning existing software schemas mirrors past system constraints rather than future business requirements.
Conceptual Abstraction: Tools like Lawrence Corr’s BEAM framework and Shane’s Business Event Matrix guide interviews with Subject Matter Experts (SMEs). Capturing raw narrative transcripts allows teams to re-query the original human text repeatedly across different design phases, preventing premature data flattening.
4. Edge Cases: Critical Scenarios Beyond AI Capabilities
The panel outlined specific real-world scenarios where AI cannot generate data models due to the absence, secrecy, or novelty of underlying data.
Scenario Type
Real-World Example (from Transcript)
Why AI/LLMs Fail in This Context
Zero Prior IT Systems / Greenfield Entity
Brand-new Australian government department; Employee #1 had an Act of Parliament and a budget, but no real estate, staff, or existing IT databases.
There is no existing physical or text data for AI to scan. The conceptual model must be synthesized from vague, contradictory legislative text and expert interviews to benchmark vendor candidate software.
Mass Mergers & Complex Consolidation
Amalgamation of 73 separate health practitioner organizations (dentists, chiropractors, nurses) across multiple Australian states into a unified entity.
No unified software or dataset exists. Human consensus must reconcile disparate operational terminology and organizational structures.
Highly Guarded / Stealth Business Initiatives
Stealth product launches, commercial M&A strategies, or military/security domain modeling.
Strategic data cannot be exposed to commercial AI models due to strict security boundaries. Requirements rely on trusted human-to-human relationships.
Evolving / Visionary Innovation
A PhD student developing a novel concept to revolutionize global processing in the Australian wool industry.
AI predicts token patterns based on existing knowledge combinations. Novel conceptual paradigms must first be conceived by humans before AI can process or extend them.
The “Peanut Butter Floor” Analogy
To illustrate AI’s fundamental limitation with novel concepts, Remco referenced an art installation featuring a floor made entirely of peanut butter:
AI understands that a floor is a structural surface you stand on. It understands that peanut butter is food or an ingredient. AI would never independently combine these two concepts into a floor made of peanut butter because it only operates on established probabilities. Only after a human introduces the novel concept can the AI synthesize variations (e.g., floors made of honey or flour).
5. Data Modeling Methodologies: Top-Down, Bottom-Up, and Patterns
Data modeling professionals employ divergent strategies that ultimately converge in the middle.
TOP-DOWN APPROACH (John) BOTTOM-UP APPROACH (Remco/Shane)
┌───────────────────────────────┐ ┌───────────────────────────────┐
│ High-Level "Data Town Plan" │ │ Business Events & Transcripts │
│ Generalized Patterns: │ │ Identify Nouns & Verbs │
│ Location, Party, Asset, Role │ │ Extract Discrete Concepts │
└──────────────┬────────────────┘ └──────────────┬────────────────┘
│ │
└───────────────────┬──────────────────────┘
▼
[ CONVERGENCE IN THE MIDDLE ]
- Validate Core Concepts
- Flag Enterprise vs. Local Scope
- Run Business Question Walkthroughs
Top-Down (”Data Town Plan”): John begins with highly generalized, universal patterns (e.g., Party, Location, Asset, Resource, Task). From this broad macro-view, the modeler drills down into specific domain details through SME interviews, enriching the central enterprise model while flagging hyper-specific nuances that belong only to localized sub-models.
Bottom-Up (Event-Driven Matrix): Remco and Shane start by evaluating specific business processes, business events, transcripts, nouns, and verbs. They extract discrete concepts and cluster them into logical categories (e.g., locations, people, physical assets).
Universal Patterns (Silverston / Hay): Universal patterns provide a major jump-start—potentially covering 50% to 75% of a standard business model off the shelf. However, blindly applying industry reference models or rigid vendor templates (e.g., ERP systems) without customizing or simplifying them introduces unnecessary complexity and operational friction.
Framework Diversity: Modeler skill relies on mastering multiple approaches—such as Fact-Based Oriented Modeling (FCOAM), BEAM, ELM, Inmon, Kimball, or Data Vault—and selecting the appropriate tool based on the client’s problem space. As referenced from Joe Reis’s Mixed Model Arts, no single modeling paradigm fits every scenario.
6. Human Factors: Trust, Culture, and Exceptions
A recurring theme is that data modeling is primarily a social, communicative discipline rather than a purely technical one.
Communication Models over Technical Diagrams: A data model serves fundamentally as a communication bridge between stakeholders. Static, multi-million-dollar enterprise models generated in isolation fail because they lack business alignment and stakeholder buy-in.
Cross-Silo Workshops: Human interactions reveal operational exceptions that algorithms ignore.
Social Services Case: Bringing siloed social workers (drug addiction, homelessness, domestic violence, unemployment) into a joint workshop revealed how an individual’s challenge cascades across domains, leading to an ongoing, board-level data integration initiative.
Regional Operational Exceptions: In a social housing organization, regional managers in Auckland optimized for high-volume intake pipelines, whereas managers in Northland bypassed standard staging protocols to immediately house families sleeping in cars. Scanning database logs alone would brand the Northland process as an anomaly rather than a necessary, value-driven adaptation.
Organizational Culture: Standard reference models enforced by major ERP vendors often fail because they ignore the unique culture and operational realities of the adopting enterprise.
Tired of vague data requests and endless requirement meetings? The Information Product Canvas helps you get clarity in 30 minutes or less?
Transcript
(00:00:03.104):
Excellent.
(00:00:03.784):
Welcome, John.
(00:00:04.844):
Welcome, Remco.
(00:00:05.885):
Thanks for coming on the show.
(00:00:07.325):
Today, we’re going to have a bit of a live chat about this idea
(00:00:11.526):
of democratizing data modeling through AI,
(00:00:14.766):
and it’s driven really around a paper that John published on LinkedIn
(00:00:19.287):
with his thoughts around where AI can and potentially can’t help.
(00:00:25.409):
So I wanted to get John on,
(00:00:26.369):
and then obviously, Remco,
(00:00:27.269):
you’ve been So I thought I’d get you on and then between the three of us,
(00:00:34.037):
we’ll probably have similar but divergent views on this thing.
(00:00:38.979):
So before we kick off though,
(00:00:40.420):
for the very few people out there who don’t know who you are, John,
(00:00:42.881):
do you want to give a bit of background to the audience of you and your background?
(00:00:48.084):
Okay, so you probably get from my accent, I’m an Aussie, got into IT in the late 1960s.
(00:00:57.639):
It’s, what, 57 years I’ve been in IT.
(00:01:01.480):
Basically,
(00:01:02.580):
a long story short,
(00:01:03.720):
started as a developer,
(00:01:05.400):
moved into boring things like project management,
(00:01:09.741):
started to focus on data modelling in the 90s and have seen an increased interest,
(00:01:17.883):
not just in data modelling for specific solutions,
(00:01:21.363):
here we’ve got a relational database,
(00:01:22.784):
how do we design it,
(00:01:24.004):
but more and more at the enterprise level, integration, data sharing.
(00:01:29.471):
So there we go.
(00:01:30.512):
That’s my quick intro, Shane.
(00:01:33.634):
Thanks.
(00:01:34.395):
Remco, what about you?
(00:01:35.375):
Well, then I’m going to do a very quick intro.
(00:01:37.797):
I’m a little bit younger.
(00:01:39.619):
I’m Dutch, different accent.
(00:01:42.665):
More or less the same kind of story after my fifth career when I went into BI
(00:01:48.729):
and data warehousing.
(00:01:49.629):
So I did the whole development.
(00:01:51.270):
I did a little bit of project management and do a lot of data modeling and now currently in AI.
(00:01:56.454):
So basically the same kind of story.
(00:01:58.735):
John worked only.
(00:02:00.917):
Kudos to you, John.
(00:02:02.118):
You know everything.
(00:02:03.239):
You’ve been the old dude over here with all the knowledge.
(00:02:06.641):
Yeah.
(00:02:08.042):
I’m just new kid on the block, I guess.
(00:02:10.874):
Wow.
(00:02:11.114):
If you’re in UK, then I must be like a baby.
(00:02:14.575):
Yes, you are.
(00:02:15.796):
Okay.
(00:02:16.096):
Thank you.
(00:02:16.616):
All right.
(00:02:16.996):
So, John, why don’t you give us a bit of a summary around the paper that you wrote?
(00:02:22.158):
What was the core thesis, effectively, or the core part of your argument?
(00:02:27.119):
Okay.
(00:02:27.339):
Well, we are all seeing a lot of chatter about AI.
(00:02:34.742):
Questions are being asked.
(00:02:36.444):
Is it going to threaten our jobs?
(00:02:39.326):
Is it going to replace us?
(00:02:42.349):
Discussions on human in the loop.
(00:02:44.491):
And on that,
(00:02:45.552):
I’ve seen a comment that what human in the loop does not mean is AI does all
(00:02:50.876):
the work.
(00:02:51.917):
And at the end, a human hits the approve button, has no idea what they’re approving.
(00:03:02.124):
One of the many areas that AI seems to be Really getting a lot
(00:03:07.351):
of traction is looking at the data of an enterprise,
(00:03:10.833):
structured data, unstructured data.
(00:03:13.555):
And what people are saying,
(00:03:15.717):
plenty of spokespeople are saying this, Shane,
(00:03:19.279):
is that AI, if it has to guess at the meaning of data, it will do the best it can.
(00:03:29.707):
But if you can help AI,
(00:03:31.970):
By giving it definitions in business terms,
(00:03:35.611):
what does this business mean by a customer, by an employee, by a product?
(00:03:41.973):
What does the business mean by these terms?
(00:03:45.174):
If you can help AI to help you by giving business centric definitions,
(00:03:53.817):
then you get better results.
(00:03:56.018):
Now, just to confuse the matters,
(00:04:00.026):
People talk of glossaries,
(00:04:01.767):
of dictionaries, of ontologies, taxonomies, context, semantics.
(00:04:08.469):
I can probably think of some more names.
(00:04:11.690):
But for the sake of discussion,
(00:04:13.851):
I think we’re all agreed that what we need is definitions of the business
(00:04:21.773):
that humans are involved in and agree with and that AI can understand.
(00:04:30.067):
So if we have a general feeling that business definitions are helpful, here’s the real question.
(00:04:40.214):
If AI needs that,
(00:04:42.696):
can AI go one step further and generate the business definitions that it needs?
(00:04:51.687):
So that’s the heart of the topic.
(00:04:54.830):
Yeah, so I think, you know, we hear a lot about human in the loop.
(00:04:58.454):
And what we actually see is human at the end of the line.
(00:05:02.339):
You know, the human layer, like you said.
(00:05:04.642):
Oh, that looks good.
(00:05:05.483):
I don’t understand it.
(00:05:06.544):
Carry on, right?
(00:05:07.365):
Good.
(00:05:07.945):
Done it.
(00:05:08.326):
So, yeah, I think that’s an anti-pattern.
(00:05:11.089):
I think that one’s going to get washed out.
(00:05:13.433):
I think the other thing we see a lot of is what I call the BI semantic layer or
(00:05:18.358):
the BI metrics layer, right?
(00:05:19.959):
It’s the definitions happen at the end of the process and they’re based on physical tables.
(00:05:25.465):
And so one of the things that I picked up in your article,
(00:05:28.107):
and we’ll kind of talk about some of the edge cases you had
(00:05:30.369):
because you’ve done stuff that I don’t think either Remco or I have,
(00:05:33.832):
and they kind of brought home those edge cases.
(00:05:37.148):
One of the keys for me in terms of conceptual or concept models or ontologies
(00:05:42.749):
and taxonomies is that we’re doing the definitions up front before we get to any of
(00:05:47.871):
the physical data in a source system.
(00:05:49.951):
So I just want to jump over to Renko because you’ve been working on AI tooling in this space.
(00:05:55.833):
Is that what you see that actually getting AI to help us at the beginning of
(00:05:59.834):
the process before we look at any physical tables,
(00:06:02.514):
any actual data is where some of the value is?
(00:06:06.107):
Well,
(00:06:06.747):
I would definitely say that the whole ILM approach that I’m a big fan of, right,
(00:06:12.632):
using business input to come up with a data model is key to any successful using AI,
(00:06:19.476):
yes or no.
(00:06:20.677):
It also reminds me back around 2010-ish something.
(00:06:28.142):
I was involved 2012, around that time, with the Boulder BI Brain Trust around Claudie Imhoff.
(00:06:37.487):
And what’s cool, we had vendors coming in and a lot of, especially the ETL vendors,
(00:06:44.596):
They said, you don’t need to model, we do the modeling for you.
(00:06:47.758):
Or a tool like Clickset, you don’t need to model, we do the modeling for you.
(00:06:52.022):
And it always comes up with, yeah,
(00:06:54.064):
we’re looking at the source system and basically restructure the source
(00:06:56.867):
into something which is actually the same,
(00:06:59.549):
but then in a different way,
(00:07:01.231):
which is not reflecting to the business in any way,
(00:07:03.953):
especially not over multiple systems integration.
(00:07:07.757):
So our approach has always been and my approach has always been looking from
(00:07:11.138):
a different perspective,
(00:07:12.138):
looking for, okay, what does the business mean on this process?
(00:07:16.260):
What they’re talking about in their own term and what is the definition of it?
(00:07:20.401):
And then find out, okay, what are the sources that match to that one?
(00:07:24.643):
And what I see now in the AI age, right, people are doing the same stuff.
(00:07:29.544):
We have AI, and indeed, let’s have AI look at the source systems and just create a model.
(00:07:34.988):
And either it’s going to be a data fault model or iceberg tables or DBT streams.
(00:07:42.793):
Whatever they can do, they just push the button and, oh, there it is.
(00:07:47.556):
Without thinking about, is this really something which is useful for the business?
(00:07:53.279):
And yes,
(00:07:54.460):
I do believe that the first step is always to say,
(00:07:58.669):
Look at what the business is telling you.
(00:08:00.610):
Get the definition of the business.
(00:08:02.090):
What do you mean with a customer?
(00:08:03.551):
Do we mean the customer, someone purchasing or someone is contacting us?
(00:08:09.753):
Or is it the internal person placing a call with our help desk?
(00:08:14.275):
What is the customer?
(00:08:16.296):
All that kind of stuff.
(00:08:17.976):
And I found out that if you give the right information to AI,
(00:08:23.118):
which I described in my book in Rati, the ELM data modeling framework, right?
(00:08:28.720):
The right amount of information, it is capable of helping and assisting.
(00:08:33.404):
And there’s a big reason.
(00:08:35.525):
Remember that you were in the loop, we just pushed for agreement.
(00:08:41.229):
I have six different services,
(00:08:42.931):
six different, let’s say, agents, and they each do a different task.
(00:08:47.394):
And they’re not streamlined.
(00:08:49.155):
You manually have to copy and paste the output from one as input in the other one.
(00:08:53.899):
And it’s just not that I don’t want to say, okay, we can do it easier.
(00:08:58.440):
I can, but I want the human to be really looking at what are you doing?
(00:09:03.325):
What part are you in?
(00:09:04.366):
What are you approving?
(00:09:05.948):
Because I think that’s more important than just push the button.
(00:09:09.811):
Because I don’t believe in push button automation or push button data modeling.
(00:09:16.389):
Yeah, I think Johnny Winter’s on the sub-sex side of this,
(00:09:19.492):
and he’s saying he believes in AI in the loop is his favorite phrase.
(00:09:22.836):
And actually, that’s probably a better way of phrasing
(00:09:25.579):
it, right?
(00:09:25.959):
I like AI in the loop for sure.
(00:09:27.741):
That’s also, I’m calling AI, for me, it’s an assistant.
(00:09:32.205):
That’s the only thing it’s doing.
(00:09:33.186):
It’s helping.
(00:09:34.614):
Yeah, and I agree.
(00:09:36.335):
As we’re all experimenting,
(00:09:37.556):
what I’m doing is I’m putting manual friction points in between all the work that AI
(00:09:42.880):
or the gene AI,
(00:09:45.622):
LLMs are doing for me because I want to be able to be in control of those handoffs
(00:09:49.485):
because that’s when I’m in the loop.
(00:09:51.546):
That’s when I’m reviewing what it’s doing and seeing what the next stage is.
(00:09:54.769):
But if we go back to that idea of
(00:09:58.201):
If we look at the library sciences and our friends that are very strong
(00:10:02.368):
on ontologies and taxonomies and thesauri,
(00:10:06.097):
You know,
(00:10:06.317):
what they say is definitions early and definitions done well
(00:10:09.998):
then drive everything else.
(00:10:11.839):
And, John,
(00:10:12.639):
one of the,
(00:10:12.979):
you know,
(00:10:13.479):
I kind of believe that,
(00:10:14.519):
but we’ve seen the whole DBT wave where,
(00:10:16.820):
you know, there is, you know, they called a blobocoder model.
(00:10:19.481):
That got me annoyed.
(00:10:21.121):
Even before there,
(00:10:21.741):
like you said, Remco,
(00:10:23.602):
we saw data warehouse automation tools that created source-specific data vaults,
(00:10:28.183):
data warehouses, whatever data modeling technique they were using.
(00:10:32.344):
So just automating it, you know, for speed but not for value.
(00:10:36.425):
But John,
(00:10:37.106):
in that paper that you wrote,
(00:10:38.227):
you gave us some really interesting edge cases that probably neither Remco
(00:10:42.769):
or I had ever thought about.
(00:10:45.711):
So where actually the physical data was not actually accessible to you,
(00:10:51.575):
but you had to come up with a model.
(00:10:53.496):
Do you want to talk through that briefly?
(00:10:55.677):
Yes, thank you, Shane.
(00:10:58.159):
I’ll make a statement first, though,
(00:11:00.480):
and perhaps this is one of the things that I’d like people who listen to this
(00:11:04.262):
to go away with.
(00:11:06.064):
And
(00:11:08.406):
can we call it a context model for the sake of a name, conceptual model, whatever.
(00:11:15.512):
One of these definitions and it is hugely powerful for analytics.
(00:11:24.599):
And I don’t care whether it’s an Inmon style,
(00:11:28.182):
a Kimball style, Data Vault and there’s other approaches out there as well.
(00:11:34.197):
Hugely powerful where you have existing sources of data and you want to integrate.
(00:11:40.901):
But a heck of a lot of the cases I’m seeing are where you need this concept model
(00:11:49.447):
for a business.
(00:11:50.828):
And I’m going to take a few extreme ones.
(00:11:53.229):
I worked for a guy who headed up a brand new government department
(00:12:01.255):
and he was employee number one.
(00:12:03.768):
He had an act of parliament and a budget and there were no existing IT systems to look at.
(00:12:11.353):
There were no employees that you could interview.
(00:12:14.496):
You couldn’t go to the office of this new department because
(00:12:20.920):
the real estate hadn’t been acquired yet.
(00:12:23.982):
All they had was an act of parliament,
(00:12:27.145):
which was terribly vague, imprecise and contradictory and a budget.
(00:12:34.258):
And he had nine months to create an entire department with staff,
(00:12:39.582):
HR salaries, CFOs, et cetera, et cetera, starting with nothing.
(00:12:47.247):
So what he asked me to do was to work
(00:12:49.608):
with some subject matter experts who’d helped shape the legislation and come up with
(00:12:56.433):
a concept model,
(00:12:58.294):
which we could then go And use it as a benchmark for evaluating candidate software.
(00:13:05.311):
So in that scenario, there was no data to analyze.
(00:13:09.914):
There was no relational tables.
(00:13:11.876):
They didn’t exist.
(00:13:15.659):
So you can use the concept model for benchmarking software for designing new bespoke things.
(00:13:24.325):
Another example was, again, Act of Parliament here in Australia.
(00:13:30.952):
Where we had over 70 separate organisations and I’ll give some examples,
(00:13:38.499):
the dentists in my state of Victoria,
(00:13:41.962):
the chiropractors in Victoria,
(00:13:44.624):
the nurses in Victoria and then the same for other states.
(00:13:48.607):
So we had 70 organisations covering multiple states
(00:13:53.251):
and multiple health practitioners and Anybody, Shane,
(00:13:59.098):
who’s been involved in mergers and acquisition where you’ve got to get two companies
(00:14:04.121):
to become one,
(00:14:05.542):
imagine melding 73 organisations into one and they had no software
(00:14:14.047):
for this unified company.
(00:14:17.990):
And again,
(00:14:19.391):
a business centric view of what the future would look like When
(00:14:23.937):
the future did not exist.
(00:14:26.078):
So I’m just please Shane and Remco throwing in some examples where you can’t go
(00:14:31.302):
to existing data because the existing data doesn’t even exist.
(00:14:37.525):
So you just want to pick up on that idea of subject matter experts, right?
(00:14:40.527):
So as you both know,
(00:14:41.688):
Juha and I are writing a book around modeling business concepts and we’re trying
(00:14:45.730):
to create a simple guide for people that don’t model.
(00:14:48.914):
To understand a relatively simple process to start off with, right?
(00:14:52.937):
So a beginner’s guide to a degree.
(00:14:55.598):
And when we come down to the core, the core is interview subject matter.
(00:14:59.561):
Well, set a scope, right?
(00:15:00.762):
Set a boundary of the work you’re going to start off with.
(00:15:03.284):
Don’t boil the ocean.
(00:15:04.965):
Identify some subject matter experts.
(00:15:07.386):
Talk to them.
(00:15:08.907):
Take what they tell you.
(00:15:09.968):
Identify the nouns, which are the concepts, the things we care about.
(00:15:13.581):
Go and define what they mean.
(00:15:16.364):
So when we see subscription,
(00:15:18.105):
customer,
(00:15:18.646):
supplier, sales order, those kind of things, get a definition we can all agree.
(00:15:23.290):
Identify the relationships that happen between those concepts.
(00:15:27.634):
So customer places an order in a store.
(00:15:30.757):
And then potentially go into a little bit more detail around rules or cardinality
(00:15:35.362):
if we find some value in that.
(00:15:37.379):
But if I come back,
(00:15:38.540):
that idea of a concept model being just a bunch of things,
(00:15:42.544):
definitions of those things,
(00:15:44.065):
and relationship between those things,
(00:15:46.447):
do you both agree that really that is the core that we’re talking
(00:15:50.571):
about actually creating?
(00:15:52.012):
And there’s different ways to do it,
(00:15:53.493):
but that’s the value, right, of just defining those relatively simple definitions?
(00:15:59.058):
I totally agree, Shane.
(00:16:00.760):
And I think it’s fun because I’m doing now my main work,
(00:16:05.366):
apart from training and writing and stuff,
(00:16:07.648):
it’s indeed consulting coaching on data modeling.
(00:16:10.552):
I don’t think I really saw a data source system for the last six, seven, eight years.
(00:16:18.361):
I never...
(00:16:19.434):
I ask for sometimes examples coming out of the actual data because it’s helping me
(00:16:23.978):
to enforce relationships, right?
(00:16:26.139):
Is it really just a relationship?
(00:16:28.741):
But basically,
(00:16:30.283):
for a concept or a context model,
(00:16:32.705):
you don’t need any source system because they are only confusing.
(00:16:36.888):
Because again,
(00:16:38.029):
terminology used in any single source system can use the same term
(00:16:43.553):
for completely different
(00:16:47.541):
I work for a healthcare organization where we talked about a bad
(00:16:52.005):
and my famous example, what is a bad?
(00:16:54.507):
And I have my categorization,
(00:16:56.129):
is it a place, is it a thing, a person, event or something different?
(00:17:00.693):
And I had half of the works of saying it’s a place and the other half says it’s a thing.
(00:17:05.538):
I found out that the definition for bad,
(00:17:07.900):
just the same word, there were two different definitions.
(00:17:11.753):
And if I would focus on source systems,
(00:17:14.234):
I wouldn’t have found any information because that’s just how it works, right?
(00:17:18.056):
You see a bed and it’s connected to all the other stuff.
(00:17:21.557):
But from a conceptual perspective,
(00:17:24.258):
contextual perspective, they’re completely different and different used.
(00:17:28.400):
By the way,
(00:17:28.780):
they’re a place for the whole hospital organization,
(00:17:34.042):
maintaining occupants and et cetera.
(00:17:37.884):
And they are a thing for maintenance and housekeeping.
(00:17:43.316):
So just for the people of a strong Kiwi or Australian accent, that’s B-E-D as in bed.
(00:17:50.428):
Oh, sorry, sorry, sorry.
(00:17:52.050):
That’s right, because every time you say it, I think bet, B-E-T.
(00:17:55.813):
So let’s play devil’s advocate, right?
(00:17:57.574):
So why don’t I just go and define my scope,
(00:18:00.776):
find the subject matter expert,
(00:18:02.197):
interview them,
(00:18:03.979):
get a transcript,
(00:18:05.320):
put it into my favorite LLM tool and say to it, create a concept model.
(00:18:10.287):
And it’s a language model, right?
(00:18:12.368):
It’s going to find the nouns.
(00:18:13.668):
It’s going to find the verbs.
(00:18:14.969):
It’s going to probably make up a definition, but we can argue about that.
(00:18:18.930):
I mean, John, what’s wrong with that, right?
(00:18:20.651):
What’s wrong
(00:18:21.111):
with me just shortcutting all this really what looks like complex hard work
(00:18:26.033):
that experts do that really these tools can do now, right?
(00:18:29.774):
That’s my devil’s advocate approach for this one.
(00:18:33.956):
Okay, so I want to take, if I may, please, Shane, two extremes.
(00:18:37.637):
One extreme is where you’ve got An organisation where the source systems reflect
(00:18:44.523):
the business,
(00:18:45.984):
they’re well defined,
(00:18:49.706):
people know and love it and it’s a great start
(00:18:55.009):
and it may actually take you almost all the way.
(00:18:58.972):
So to use an LLM to look at text or existing data structure where it is Stable
(00:19:09.655):
and agreed, why not?
(00:19:12.757):
Yeah, let’s do it.
(00:19:15.138):
We talk of some edge cases.
(00:19:17.880):
So one example,
(00:19:20.642):
I worked for an organisation that said we’ve got two,
(00:19:26.665):
I’ll say secret initiatives and these are not spy intelligence military things.
(00:19:33.090):
This was a commercial organisation and
(00:19:37.402):
Lots of organizations want to develop new products, but they keep the products under wrap.
(00:19:42.906):
Or even tighter is if they are intending to acquire another organization
(00:19:49.370):
and completely change the emphasis of their organization,
(00:19:56.874):
then sometimes you may not be able to expose to AI the information it needs Because
(00:20:07.774):
that information is closely guarded by a handful of people and you’ve got
(00:20:14.678):
to build trust with those people.
(00:20:18.821):
So please, Shane, I’m taking two extremes.
(00:20:21.583):
One where maybe AI can do most, if not all, of the work for you.
(00:20:28.367):
And if not, why not?
(00:20:29.788):
Yeah, let it shine.
(00:20:32.750):
But what I’m saying is that At the other extreme for a variety of reasons
(00:20:38.622):
and I’ve done some work for organisations in let’s say
(00:20:41.782):
the military security type world where the information is tightly guarded
(00:20:52.505):
or commercial situations where it’s hidden.
(00:20:55.845):
I’m going to throw in another one Shane where I’ve come across and I’m happy
(00:21:01.286):
to share some examples of this.
(00:21:04.433):
But where the,
(00:21:07.215):
and I’ll share one example,
(00:21:08.696):
Australia wool industry,
(00:21:11.198):
there was a PhD student who had a vision
(00:21:15.282):
for completely revolutionising Australia’s wool industry.
(00:21:21.026):
And he had to interact with stakeholders,
(00:21:24.529):
including internationally, woolen mills in China and Spain, et cetera, et cetera.
(00:21:29.753):
And there was nothing That AI could look at to capture his vision.
(00:21:38.108):
And yet we had four months to develop the software.
(00:21:44.312):
This is the days when I was actually hands on the keyboard doing software development.
(00:21:50.556):
But my responsibility,
(00:21:52.598):
it was not a conceptual model,
(00:21:56.460):
it was a physical model,
(00:21:58.522):
but based on the concepts which were still evolving and There is no way,
(00:22:05.959):
if AI had existed at the time,
(00:22:08.180):
that it would have been able to get hold of the knowledge it needed because
(00:22:12.802):
the knowledge was still evolving.
(00:22:14.903):
Had to have people.
(00:22:18.905):
John, sorry, Shane,
(00:22:20.065):
to interrupt,
(00:22:20.526):
but I think you hit the nail on the head,
(00:22:26.568):
I guess, with Your example on the wall, we have a conceptual artist.
(00:22:32.687):
Well, we had one in the Netherlands.
(00:22:34.910):
He passed away just recently.
(00:22:37.293):
And one of his ideas,
(00:22:38.455):
a famous one is for a long time, is an entire floor made out of peanut butter.
(00:22:46.773):
And it’s conceptual, it’s great, it’s smelly, but then again, it’s fun.
(00:22:53.359):
But I talked to friends as well on this part,
(00:22:55.621):
I said AI could never come up with using peanut butter for floor.
(00:23:01.147):
And a very simple reason, AI exists about what is given to it.
(00:23:06.409):
It knows that a floor is something constructional where you can stand on,
(00:23:11.631):
it’s holding stuff together.
(00:23:14.352):
That’s what its understanding of a floor is.
(00:23:16.712):
Its understanding of peanut butter is you can cook with it,
(00:23:19.693):
you can eat it, you can feed birth with it, whatever.
(00:23:23.555):
It wouldn’t make the combination.
(00:23:25.375):
Only after you show it, Oh,
(00:23:27.873):
by the way,
(00:23:28.453):
there’s a concept of a floor made out of peanut butter and it’s art, whatever.
(00:23:33.636):
Then it can do some other kind of combination, right?
(00:23:38.278):
It might come up with a floor out of honey or flower baths or whatever.
(00:23:45.141):
But it first needs to be fed with the idea, like the concept, right?
(00:23:49.183):
Something crazy new, it has to be fed with it.
(00:23:52.404):
And then it can work with that part.
(00:23:55.905):
So it can go on.
(00:23:58.913):
And I think that’s very important to understand because, indeed,
(00:24:02.334):
what we’re saying,
(00:24:03.015):
right, AI is just a tool very good in prediction of what would be the next token.
(00:24:10.058):
So let’s
(00:24:11.118):
just get kind of into the messy middle, right?
(00:24:13.800):
So, John,
(00:24:15.120):
I’ve never been lucky enough to go into an organization that had a source system
(00:24:19.242):
that actually matched their business model or documentation that ever survived
(00:24:24.024):
the first scan of it.
(00:24:26.506):
Two out of six.
(00:24:29.213):
You still have time, Shane.
(00:24:30.235):
You still have time.
(00:24:30.796):
You’re young.
(00:24:34.116):
And so, yes,
(00:24:35.497):
harvesting from documentation is useful for an LLM,
(00:24:38.619):
but also there’s the risk that it stays out of date, right?
(00:24:41.881):
Or nowadays, actually, even worse, it was written by an LLM.
(00:24:45.002):
So let’s go, okay,
(00:24:48.604):
there’s these edge cases,
(00:24:49.965):
the ones you described, John, where actually there is nothing to look at, right?
(00:24:53.527):
It is purely talking to a human because we’re at the inception stage, really, right?
(00:24:58.130):
We’re kind of building this thing incredibly early in the business model for the organization.
(00:25:03.221):
And then let’s stay away from the physical table stuff, right?
(00:25:06.382):
The bi-semantic metric layer and the reverse engineering of
(00:25:10.784):
the physical data structures because I think LLMs can do that.
(00:25:14.946):
I don’t think we should, right?
(00:25:16.466):
I think it’s an input to conceptual modeling or creating concept or context model.
(00:25:21.008):
I don’t think it’s the answer, right?
(00:25:22.769):
It’s not soup to nuts.
(00:25:24.229):
And I want to kind of focus on that middle ground, right?
(00:25:28.540):
So Remco,
(00:25:29.461):
you know,
(00:25:29.862):
one of the things that I started off doing was I would interview people,
(00:25:33.045):
subject matter experts, and then you and I both did this.
(00:25:37.671):
And then what we found is we were translating what we heard into a bunch
(00:25:41.816):
of artifacts that then went into the concept model.
(00:25:46.015):
And so I’m not a great fan for ad hoc, right?
(00:25:49.238):
It’s just not the way my brain works.
(00:25:50.559):
So I need a set of patterns to run things in a semi-repeatable way for me to be able
(00:25:54.502):
to do something.
(00:25:55.502):
So that’s why I was always a great fan of Lawrence Kaur and his Beam stuff
(00:25:58.605):
because I kind of got it right and I could use it and run it.
(00:26:02.484):
And so what I found now when I’m interviewing people for identifying concepts,
(00:26:06.607):
relationships, and definitions, I use a thing called a business event matrix, right?
(00:26:10.591):
I actually bring it up on the screen and I use it as an artifact to guide the conversation.
(00:26:15.035):
So what I find is by asking them questions and getting them to tell me stories
(00:26:18.979):
and typing it into that, right?
(00:26:20.420):
They can see what I’m typing and then I get instant feedback.
(00:26:23.644):
But what I found was I’m now bringing an abstraction layer in.
(00:26:27.799):
So they’re telling me stories and I’m already summarizing those stories into
(00:26:31.621):
an abstracted layer and I lose the value of being able to go back to
(00:26:36.003):
that transcript effectively and ask it different questions and reuse it many,
(00:26:40.845):
many times for different parts of the design and build focus.
(00:26:44.427):
With the Elm stuff, you’re doing the same, right?
(00:26:46.448):
It’s like the original Elm thing was talk a bunch of effectively pattern templates
(00:26:51.950):
to abstract out in a way that makes sense and then use that to create
(00:26:56.432):
The concept model, which in itself is an abstraction.
(00:26:59.755):
Is that what you’re still doing?
(00:27:05.219):
The whole framework I developed using LLM.
(00:27:09.041):
Basically, the first two parts of that one is discovering what we call the core-based concept.
(00:27:14.265):
Basically,
(00:27:14.726):
what are the terminology and discovering the MBR of the relationships
(00:27:18.768):
and how they relate to each other, including giving a definition.
(00:27:23.058):
So those are my two first services.
(00:27:26.004):
I still got back from my,
(00:27:28.107):
because it’s still following kind of the same kind of pattern rights and
(00:27:31.433):
the same kind of templates.
(00:27:34.109):
After each part, you will get questions back.
(00:27:36.991):
So my LLM gives me a transcript out of a workshop or out of interviews or basic documentation.
(00:27:44.497):
I’ll feed it and it gives me already back.
(00:27:47.359):
Oh, by the way,
(00:27:48.339):
these are questions you have to ask in the next workshop or more like suggestions.
(00:27:52.522):
So that is already in the first two, I call the services, right?
(00:27:56.265):
They’re already giving you back.
(00:27:58.066):
But then I have something and I think this is interesting for you as well to say,
(00:28:03.010):
I have a review service to say, this was your original input.
(00:28:07.333):
This is what we found so far.
(00:28:10.075):
Check the output with the business case.
(00:28:12.836):
How confident are you that it’s still correct?
(00:28:15.138):
What are the gaps?
(00:28:16.018):
What are the still open points?
(00:28:17.819):
So you get a whole bunch of rich information to follow into the next workshop.
(00:28:22.780):
And that’s really, really helpful.
(00:28:24.581):
I still think that after a workshop,
(00:28:27.623):
when I’m going to fill in the templates,
(00:28:29.584):
I’m still giving it back to business and say, this is what we found so far.
(00:28:34.486):
Please make sure that we found the right definition or what you think that is
(00:28:38.989):
a customer, what do you think it’s an order?
(00:28:41.470):
Make sure that the definition is correct.
(00:28:43.671):
Please fill it in.
(00:28:44.932):
If you have some additional information,
(00:28:46.593):
just add it to the template because it’s going to help me.
(00:28:51.728):
I also have a service which actually can say during a workshop,
(00:28:57.632):
you have now your contextual model, your concept model.
(00:29:00.434):
Can this question,
(00:29:01.815):
the business is going to ask,
(00:29:03.076):
is this question be answered with the model we came up so far?
(00:29:08.779):
If not, do we need to add it?
(00:29:10.981):
So it also has the option to want to figure out.
(00:29:14.787):
What are the questions we’re having?
(00:29:16.988):
Is it a fit?
(00:29:17.809):
But also,
(00:29:18.569):
and that’s a little bit different input,
(00:29:21.671):
basically say, oh, by the way, you forgot this core business concept.
(00:29:25.853):
Oh, you forgot this relationship.
(00:29:27.574):
Add it to the model during the workshop itself.
(00:29:31.636):
And I think this is very important.
(00:29:33.397):
That is not replacing interviewing SMEs or workshops.
(00:29:39.420):
It’s not saying, oh, here’s the documentation about our organization.
(00:29:42.902):
Go and have fun.
(00:29:44.940):
No, we really still need to do the whole workshop thing.
(00:29:47.962):
And yes, it can help.
(00:29:49.083):
You can do some stuff in advance, but you still need to have that contact with the business.
(00:29:53.866):
There is no way that you can just push the button,
(00:29:57.269):
give some business input,
(00:29:58.249):
push the button, and voila, there’s a perfect conceptual model.
(00:30:03.112):
I’m glad you said business questions because actually step nine
(00:30:06.194):
of our book is gather the business questions and step 10 is walk the map,
(00:30:10.958):
which is the bit we’re still trying to write is actually...
(00:30:13.391):
It’s already done.
(00:30:14.332):
It’s already done.
(00:30:15.093):
Yeah.
(00:30:16.434):
So one of the things that I struggled with when I kind of came into the data world,
(00:30:19.917):
right, was,
(00:30:20.697):
you know,
(00:30:20.938):
this is many decades ago,
(00:30:23.119):
is there were a bunch of experts who knew the issues, right?
(00:30:26.162):
They’d been doing it for years,
(00:30:28.104):
and they’d gone through, they’d learned it, and they were experts.
(00:30:31.927):
And, you know,
(00:30:33.068):
when we’re doing, say,
(00:30:34.149):
a physical model,
(00:30:35.530):
I could put,
(00:30:36.251):
you know, in my consulting company, I could put Bob, right, and do the work.
(00:30:42.243):
And the outcome was always when we kind of built it and delivered it,
(00:30:46.268):
it was very disappointing.
(00:30:48.091):
And I could get Kate,
(00:30:49.152):
and I could get her to do the exact same work,
(00:30:51.655):
and she’s an expert, and then somehow the model just worked, right?
(00:30:55.340):
When we implemented it, it answered every business question, the customer was really happy.
(00:30:59.584):
And I got to the stage where I was like, well, that person’s an expert and that person’s not.
(00:31:04.626):
And I wanted to codify that.
(00:31:06.107):
And so what I was spending my time is trying to figure out what I do unconsciously
(00:31:10.508):
and then how do I do it conscious.
(00:31:11.949):
And one of the things we do as experts is we write books.
(00:31:15.770):
So, John, you’ve written a couple of books around data modeling with your expertise.
(00:31:20.452):
And my question is, as an expert, there’s a bunch of steps that you take.
(00:31:25.808):
And they probably are similar to the ones that Renko just mentioned,
(00:31:30.993):
but you do it as a human,
(00:31:33.336):
you probably do some of them unconsciously and some of them consciously.
(00:31:37.440):
And so while we still need to treat it as AI in the loop,
(00:31:41.584):
I still see value in getting AI tooling to help people that aren’t experts.
(00:31:49.187):
To understand the process and do some of it.
(00:31:51.608):
And now maybe it still needs experts at the end to review it,
(00:31:54.848):
to ask those gnarly questions, right?
(00:31:57.669):
Those two concepts look the same.
(00:32:02.030):
Why we treated them separately?
(00:32:03.530):
Now, there’s probably a good reason, right?
(00:32:04.771):
But that challenge, that review process is really valuable.
(00:32:07.911):
What’s your thoughts around that, right?
(00:32:09.212):
Like of taking your brain and making effectively a second brain of how to do this.
(00:32:14.753):
There’s got to be value there, right?
(00:32:16.872):
Okay, so I’m thinking of a whole heap of things.
(00:32:20.313):
First, I’ll mention data modeling patterns.
(00:32:22.855):
David Hay, Len Silverstone have published books on proven data model patterns that work.
(00:32:31.218):
Len calls them universal because he suggests you walk into any organization in
(00:32:37.381):
the world and you’ll be able to probably pick 50% off the shelf.
(00:32:43.778):
They’ve got employees, they’ve got customers, they’ve got products, etc.
(00:32:48.119):
Then he says,
(00:32:48.559):
but you’re not walking into any company,
(00:32:50.299):
you’re walking into a company in an industry and that industry has some patterns.
(00:32:57.341):
And please, Shane, I am not talking about industry data models, big expensive things.
(00:33:04.722):
But he’s suggesting that you might be able to get to 75% mark
(00:33:11.303):
Taking existing patterns,
(00:33:13.645):
challenging them,
(00:33:15.507):
extending them where they need extension and actually doing the opposite.
(00:33:18.649):
Remco and I have spoken about this,
(00:33:20.631):
dumbing them down to get rid of unnecessary complexity and using them as
(00:33:27.916):
a starting point.
(00:33:29.798):
The second thing about talking with people,
(00:33:33.321):
I worked for a not-for-profit with some delightful people.
(00:33:39.271):
And some of them had specialised in helping people with drug and alcohol addiction,
(00:33:43.873):
homelessness, unemployment, domestic violence, etc.
(00:33:48.716):
And the important thing was not to interview them separately,
(00:33:53.458):
but to get them into a room together.
(00:33:56.019):
And these wonderful people had worked in silos.
(00:34:01.722):
I’m not talking about data silos.
(00:34:04.563):
I’m talking about work practice silos.
(00:34:07.386):
And you get the situation, Shane, where somebody says, I’m hungry, can I have a meal voucher?
(00:34:12.971):
How come you’re hungry?
(00:34:14.152):
Well, I’m unemployed.
(00:34:15.493):
How come you’re unemployed?
(00:34:16.995):
I have an anger problem and I was abusive to my boss and I’m homeless
(00:34:23.440):
because I was also abusive to my wife, et cetera, et cetera.
(00:34:26.823):
So it’s a complex thing.
(00:34:29.826):
Got these people together.
(00:34:31.685):
And it was meant to be a data workshop,
(00:34:34.887):
but it was a person to person,
(00:34:37.188):
human to human trust based thing where we talked about issues.
(00:34:41.591):
It was so successful.
(00:34:43.412):
I got the strangest contract I have ever had.
(00:34:47.434):
I had a contract for one and a half hours per month for the next three years to sit
(00:34:53.698):
in on their monthly board meetings to bring a data integration,
(00:34:59.121):
a data sharing perspective To the organization.
(00:35:03.139):
That’s all about humans.
(00:35:04.359):
That’s not technology, Shane.
(00:35:07.881):
Yeah.
(00:35:08.741):
Yeah.
(00:35:10.502):
I remember a project we did years ago.
(00:35:12.583):
We were running the beam workshop for a housing organization.
(00:35:15.384):
So Housing New Zealand back there, which is the government owned social housing.
(00:35:20.612):
The organisational model at the time was regional.
(00:35:23.894):
There was a regional office.
(00:35:25.935):
They spent a fortune building standardised processes.
(00:35:32.097):
If you needed a house,
(00:35:33.077):
you’d come in and you’d be an early assessment and then you’d go away and
(00:35:36.419):
then you’d get an appointment and you’d come in and there’d be a deeper assessment
(00:35:39.900):
and you’d go away and put on the waitlist.
(00:35:41.961):
There was this pipeline of process effectively.
(00:35:45.462):
We were running the workshops.
(00:35:47.443):
And we’ve done what you said, right?
(00:35:48.525):
We brought all the regional managers together into the room.
(00:35:52.769):
And so everybody was explaining the process to us and giving us the data
(00:35:56.773):
because it was the beam process.
(00:35:59.435):
And then the regional manager for Northland,
(00:36:01.577):
which is at the top of New Zealand, They said, yeah, that’s not what we do.
(00:36:05.856):
And everybody’s like, what?
(00:36:07.337):
And they said, well,
(00:36:08.798):
that may work when you’re in Auckland because Auckland is the biggest city
(00:36:11.319):
and you’ve got a massive problem with thousands of people turning up and you’ve got
(00:36:14.781):
to optimise for the problem of too many people turning up.
(00:36:18.963):
But in Northland,
(00:36:19.563):
the problem we’ve got is a person will turn up and say,
(00:36:22.684):
I’m sleeping in my car with my six kids and winter’s coming.
(00:36:27.387):
And We had the capacity to take them through the whole process without having
(00:36:31.623):
to send them away and let them sleep in a car for three days
(00:36:34.905):
because we were too busy.
(00:36:36.506):
So that optimized the process for the value to their customer effectively.
(00:36:41.167):
Auckland probably couldn’t do that,
(00:36:43.068):
although they should probably look at it right
(00:36:44.669):
because I think they could have changed the constraints.
(00:36:47.990):
But we would never have known if we didn’t have everybody in the room because
(00:36:51.972):
if we had created a concept model,
(00:36:54.902):
And send it to the person in Northland,
(00:36:57.365):
they probably would have gone, yeah, yeah, whatever.
(00:37:00.147):
I’ve got five more people living in a car that I need to get them and their kids into housing.
(00:37:04.331):
So whatever corporate, yeah, yeah, it’s all good, right?
(00:37:06.933):
Because I’m not going to follow it anyway.
(00:37:09.195):
So I think that’s the first thing.
(00:37:10.817):
And then I think also if we gave all that transcripts effectively,
(00:37:15.201):
we interviewed everybody differently and we gave it to an LLM.
(00:37:18.293):
It’s going to optimize for the thing it sees the most.
(00:37:20.996):
Unless I told it, tell me where there’s variation.
(00:37:25.821):
So I think that’s the other thing when we use AI in the loop for this process.
(00:37:28.644):
We should be saying, you know, what’s the golden path?
(00:37:31.187):
What’s the happy path from what you can see?
(00:37:33.949):
And then What are the exceptions?
(00:37:36.055):
Because again,
(00:37:37.277):
my guess is between the three of us,
(00:37:40.181):
when we’re running those workshops, we’re looking for the exceptions.
(00:37:43.846):
Like Remco said,
(00:37:45.529):
bed is a physical piece of metal that gets moved around the hospital and bed is
(00:37:49.575):
a We
(00:37:57.961):
know those are the horrible ones,
(00:37:59.543):
the definition of active customer and marketing versus risk.
(00:38:04.366):
We know those are the ones that actually we’re going to spend an inordinate amount
(00:38:07.649):
of time on definitions that count.
(00:38:09.830):
So we look for them and I think LLMs don’t unless we tell it to.
(00:38:14.454):
And that’s Shane,
(00:38:15.675):
that’s exactly I have another,
(00:38:19.424):
well, of course, my framework consists of a whole bunch of services.
(00:38:22.565):
One of them is to compare and consolidate.
(00:38:25.306):
There are two reasons for that.
(00:38:26.667):
One, the main reason is always ask different LLMs to create, to do the whole process, right?
(00:38:33.789):
Find the business concept,
(00:38:34.850):
find the relationships, the definition, blah, blah, all that stuff.
(00:38:38.573):
But if you’re asking Claude or Quinn or Deepseek or Chet or Gemini or Mistral,
(00:38:46.345):
whatever, you will get different variations, slight variations.
(00:38:50.991):
So that’s compare and consolidate.
(00:38:53.455):
It’s to say,
(00:38:54.396):
I have all the different outputs,
(00:38:55.677):
all different models from each one of the family members,
(00:38:58.620):
compare them and where are the differences and consolidate.
(00:39:02.404):
But you could do the same kind of stuff saying,
(00:39:05.067):
I have different departments or different regions that are supposed to do
(00:39:10.292):
the same stuff.
(00:39:11.493):
If you talk to each of the regions,
(00:39:12.795):
what is your process, how it looks like, feed it into the LLM.
(00:39:17.318):
One by one, you get slightly different differences.
(00:39:22.401):
That one service is just there.
(00:39:24.503):
And that’s why an LLM is very good at.
(00:39:26.264):
It’s just because it is not biased.
(00:39:29.246):
It’s not saying, yeah, they’re talking rubbish or whatever.
(00:39:32.068):
They’re just saying, this is what we have.
(00:39:34.529):
This is what I found.
(00:39:35.390):
And this is which direction it’s going.
(00:39:37.891):
So you could have that comparison.
(00:39:39.973):
You could make use of the power of an LLM to do that.
(00:39:44.056):
Because I think it’s important what you said, right?
(00:39:47.559):
Hey, we talked to the buyers in the northern part.
(00:39:49.540):
They have a completely different process in Auckland and et cetera.
(00:39:53.523):
And you hear it as a person and you understand what it means.
(00:39:58.187):
But if you can say to an LM, this is the output of one, this is the output of another.
(00:40:02.790):
They’re supposed to be the same process, but they are different.
(00:40:05.493):
Tell me, what are the differences on all the different levels?
(00:40:09.760):
And I think that’s where you can use the power.
(00:40:12.821):
I’m not saying that humans should be replaced
(00:40:15.922):
because one very important thing I think we haven’t touched yet is
(00:40:20.263):
that it gives you some idea of a data model.
(00:40:24.244):
Let’s say, for me, I think it gives you some idea.
(00:40:27.485):
But you need to understand data models and data modeling as a person before implementing.
(00:40:32.646):
You need to understand what is the model given back to me.
(00:40:36.747):
Is it indeed a correct data model?
(00:40:38.668):
Yes or no.
(00:40:40.084):
You still need
(00:40:40.784):
to have that.
(00:40:42.165):
Yeah, I agree.
(00:40:43.646):
For me, the idea of a concept map,
(00:40:45.307):
a visualization of the concept model,
(00:40:48.088):
the nodes and links,
(00:40:49.249):
the concepts and the relationships, allows me to scan it really quickly.
(00:40:53.228):
And so if I get two of them,
(00:40:54.669):
I can put them side by side and visually my brain,
(00:40:56.871):
you know, it’s pattern recognition, right?
(00:40:58.432):
We look at the map and we go, they look different.
(00:41:00.734):
Why?
(00:41:01.234):
And I’ve never actually tested whether putting the diagram,
(00:41:04.757):
those maps back into an LLM give us a better response to comparison versus the text.
(00:41:09.481):
But the other thing is, it’s about this cookie cutter stuff.
(00:41:12.023):
So I want to go back to that, John,
(00:41:13.244):
because,
(00:41:14.124):
you know,
(00:41:14.545):
we’ve all been around for the $5 million enterprise data model diagram from IBM
(00:41:19.388):
that never really got implemented and then There’s been lots of industry models
(00:41:25.019):
and standard models over many waves of technology in the data space.
(00:41:29.422):
I mean, Databricks now have just come out with out of the box models again.
(00:41:35.025):
They’ve now got ontology snippets, so small definitions.
(00:41:40.348):
And I always come back to,
(00:41:41.408):
I think it was in your first book,
(00:41:42.609):
you told a story about the fire service and vehicles and what is a vehicle.
(00:41:49.133):
And the fact that they could change the engine, the body, the thing at the back of it.
(00:41:53.536):
It’s like Lego.
(00:41:54.517):
Basically, nothing survived when they reconfigured a vehicle, right?
(00:41:59.721):
And I thought about that and I thought, well,
(00:42:02.062):
if I knew there was a model that you had for a fire service
(00:42:07.406):
that can constantly reorganize their vehicles, then, yeah,
(00:42:13.950):
I can pick that up and it would probably fit to an organization that looked like
(00:42:17.533):
that.
(00:42:18.408):
But I have to know it exists.
(00:42:19.789):
I have to know that it fits for that organization.
(00:42:22.371):
If I’m, I don’t know,
(00:42:24.273):
if I’m a taxi company and I’ve got vans and trailers and I just swap a trailer on
(00:42:29.217):
a van, it’s probably not the same level of complexity.
(00:42:32.379):
I probably don’t need the same model that you created for the fire service.
(00:42:36.643):
And so I can see LLMs be useful as pointers.
(00:42:41.624):
And then I think, well, maybe it can explain to me where that model has valuing could fit.
(00:42:49.590):
But I’m not convinced, right?
(00:42:50.670):
Because,
(00:42:50.950):
you know,
(00:42:51.331):
we talked about the edge where we have no physical data
(00:42:55.434):
and we have no business model.
(00:42:56.694):
So actually, we’re starting from scratch.
(00:42:59.016):
We talked about the other extreme of just point everything at
(00:43:02.058):
a physical data structure and hope we get something valuable.
(00:43:05.390):
I see those really complex use cases as being back to that edge of needing a human
(00:43:12.015):
to bring their expertise because it’s not like everything else, right?
(00:43:15.598):
And therefore,
(00:43:16.178):
it needs that creativity and that craft to say,
(00:43:19.521):
it’s going to be a peanut butter floor,
(00:43:21.803):
not a wooden floor for whatever weird reason.
(00:43:23.724):
Thank you, Renko.
(00:43:24.244):
I’m never going to forget that one.
(00:43:25.545):
John, what do you think?
(00:43:26.246):
Do you think that’s where it is?
(00:43:27.567):
It’s like there’s useful,
(00:43:29.048):
but there’s also danger if we kind of cookie cutter it as our only way of working.
(00:43:35.395):
Absolutely.
(00:43:36.036):
So I love patterns.
(00:43:37.937):
Remco, you and I have had chats about patterns.
(00:43:42.360):
And Remco,
(00:43:43.220):
you have correctly pointed out there are dangers in unthinking,
(00:43:49.084):
unchallenged embracing of patterns.
(00:43:53.630):
There are highly generalised patterns and I’m in 100% agreement.
(00:44:02.713):
There are values in patterns,
(00:44:04.514):
they need to be understood,
(00:44:05.634):
they need to be considered for applicability,
(00:44:08.975):
but perhaps if I can make this perhaps one of the main points
(00:44:13.917):
from this session is how you
(00:44:19.812):
Handle divergent views within an enterprise.
(00:44:24.115):
So my approach and Shane and Remco, it’s...
(00:44:28.539):
Okay, I’ll set the scene for those listening in.
(00:44:31.361):
This is a gross oversimplification and I’m going to focus Remco on you.
(00:44:37.105):
I think, Shane, you’re probably aligned with Remco.
(00:44:41.268):
Remco, you take events, as do you, Shane.
(00:44:46.155):
And processes and you identify things within that discrete process.
(00:44:51.800):
So it’s identifying the specifics and from there you, Remco, cluster them.
(00:44:59.247):
Here are all the things to do with locations.
(00:45:02.449):
Here are all the things to do with people, persons.
(00:45:05.112):
Here are all the things to do with things, physical assets, etc.
(00:45:09.416):
So you cluster them and you look for The generalizations,
(00:45:15.538):
not at a ridiculous level, at a useful level that is applicable to the business.
(00:45:20.541):
Okay.
(00:45:21.342):
I start at the generalizations and I say, you’re familiar with Google Maps.
(00:45:28.326):
Let’s talk about the locations that are relevant to your organization.
(00:45:33.829):
You’re familiar with The address book in your smartphone,
(00:45:38.047):
let’s talk about people and organisations and the roles they fill.
(00:45:42.029):
Let’s talk about the tasks people perform.
(00:45:44.290):
So I start as a conversation starter with highly generalised patterns,
(00:45:50.993):
location, asset, product,
(00:45:55.335):
resource, party, people hate, love, party, whatever.
(00:45:59.977):
So I start with them and I discover the specifics So we start at different ends,
(00:46:06.301):
but I think the three of us and more are actually on the same page.
(00:46:12.445):
I am in danger if I don’t stop and take specific events and processes and challenge my model.
(00:46:20.489):
So I get to that, but I get to it later.
(00:46:24.112):
Remco, you start there and look at the generalizations.
(00:46:27.413):
Here’s the key point I want to make.
(00:46:30.544):
With the way I approach,
(00:46:32.505):
I try to come up with what I call a data town plan,
(00:46:35.087):
high level, soft, fuzzy, imprecise, but it gives a perspective of an organisation.
(00:46:44.613):
And then I drill down into,
(00:46:47.515):
I don’t care whether it’s the Auckland region and the Northern region,
(00:46:51.958):
I don’t care whether it’s regional based,
(00:46:54.619):
business unit based, process based, pain point based, I drill down into specifics.
(00:47:01.138):
And then here’s my really strong point.
(00:47:05.480):
I will discover things when I talk to this bunch of SMEs that enrich and improve
(00:47:15.444):
the central model.
(00:47:17.725):
And I will discover bits that should never go in the central model.
(00:47:22.106):
They are too specific, too much related just to that particular group of people.
(00:47:29.826):
And I capture them in the same model but I flag things that are
(00:47:37.271):
specific for
(00:47:38.492):
a small group and things that can contribute to the whole and
(00:47:43.335):
that way I can show respect to the individual stakeholders to say I’ve heard you,
(00:47:49.719):
I’ve captured what you need and this part of your world important to you,
(00:47:55.743):
this part of your world important to everybody else And I get agreement,
(00:48:00.908):
but I do not try to force everything from everybody into one consolidated model.
(00:48:12.620):
And I think that’s important, John.
(00:48:16.183):
I think I totally agree that we basically start from different angles and end up at
(00:48:21.989):
the same level for sure.
(00:48:24.952):
One thing about that standard model reference models story,
(00:48:29.377):
I work for consulting for a brewery,
(00:48:32.460):
one of the biggest one in the Netherlands, also worldwide.
(00:48:35.663):
Green Bottle, whatever.
(00:48:38.085):
And they are implementing a big German ERP system.
(00:48:43.048):
And it says,
(00:48:44.088):
this is how it works because that German company said,
(00:48:48.431):
if you’re a brewery, this is how you work.
(00:48:50.312):
These are your processes, adjust to what we are saying.
(00:48:54.255):
Almost like the Borg, resistance is futile, you will be assimilated.
(00:48:59.050):
And in the end of happening,
(00:49:00.371):
what was wrong,
(00:49:01.212):
I think on all standard models and reference models,
(00:49:04.495):
is you’re forgetting the culture of the organization you want to implement or work
(00:49:08.638):
for.
(00:49:09.459):
And the culture is so much different.
(00:49:11.301):
You cannot change the process if you’re not taking into account what the culture of
(00:49:15.104):
the organization is.
(00:49:17.134):
And I think this is where it definitely goes wrong.
(00:49:19.736):
And I totally agree.
(00:49:21.757):
I’m always starting up indeed from the more low level.
(00:49:24.900):
I’m just writing down all the nouns and the verbs,
(00:49:27.502):
et cetera,
(00:49:27.842):
et cetera,
(00:49:28.603):
and categorize them on a low level and then see who is a customer,
(00:49:32.886):
a kund and a client.
(00:49:34.407):
Are they the same?
(00:49:35.167):
Yes or no?
(00:49:35.628):
Because we’re all categorized in the same bucket.
(00:49:37.689):
They’re all persons.
(00:49:39.722):
I get it when you say, oh, we have parties and looking down to parties.
(00:49:43.944):
By the way, I do love a party, but that’s a different kind of party.
(00:49:47.586):
But do you mean it’s an organization or is it a natural person?
(00:49:53.509):
Is it something which is yours?
(00:49:55.810):
So you’re top down,
(00:49:58.371):
right from your data town plan,
(00:50:00.312):
going down to a certain level and I’m going the other direction and we meet in
(00:50:04.234):
the middle.
(00:50:05.694):
Yeah, I think data people love to argue, right?
(00:50:08.938):
If you want to argue semantics, get two data people in a room.
(00:50:12.502):
And I’ve always struggled with the idea of an ontology because to me it seemed like
(00:50:16.507):
a bunch of concepts, definitions and relationships.
(00:50:20.081):
Yeah, I see the library sciences people saying,
(00:50:21.923):
well, no, it’s not a, you know, ontologies are different to concept models.
(00:50:25.206):
And, you know, I’m reading a book on it at the moment.
(00:50:27.207):
I’ve come to the realization it’s not the thing that’s created at the end.
(00:50:30.770):
It’s the path that they use to get there.
(00:50:33.112):
They are far more formal in the way they define things to get to concepts,
(00:50:39.098):
relationships and definitions.
(00:50:42.118):
And it’s the same working with Juha on the book.
(00:50:45.241):
I can see some of your patterns, John, and the way he models the concept model.
(00:50:50.565):
I can see some of the ways that Remco and I kind of do it,
(00:50:53.688):
but even Remco and I take a different path slightly.
(00:50:56.830):
I started off very strong on the seven Ws from Lawrence, you know, the who does what’s.
(00:51:00.992):
I’ve kind of lost them now because they didn’t add as much value,
(00:51:03.494):
but I will look at a core business process, right?
(00:51:06.135):
A multi-part concept relationship, a custom orders product in the store from an employee.
(00:51:10.937):
And then that’s the story I’m looking for in the subject matter expert when I’m talking to them.
(00:51:16.800):
And from there, I decompose that into the concepts and the relationships.
(00:51:21.603):
But what it means is I can translate, right?
(00:51:23.284):
I can look at your models, John, and I can see party.
(00:51:25.905):
Right.
(00:51:26.247):
And I can go, well,
(00:51:27.958):
I would never draw a concept map with a party box,
(00:51:31.400):
but that’s just my opinion on the slight abstraction that I use for concept maps.
(00:51:38.945):
But as soon as I see yours, I understand it, right?
(00:51:40.787):
I know that when I see party,
(00:51:42.168):
I’m going to see potentially people and organizations in there because I know
(00:51:46.411):
that how you do it.
(00:51:47.451):
And so I think that’s part of the danger with LLMs,
(00:51:50.453):
especially ones that have been the harnesses, right?
(00:51:52.995):
The policies, the skills, the processes, the ways of working that are embedded in this Right.
(00:52:17.210):
Because that is the way that you work.
(00:52:18.851):
And if I don’t want to work that way, then I need to iterate your harness.
(00:52:22.713):
So what we’re seeing is,
(00:52:23.693):
you know,
(00:52:24.054):
we saw enterprise data models before as a quick cheat,
(00:52:26.775):
supposedly, and some of it was valuable, right?
(00:52:29.597):
It gives us starters, but it doesn’t give us the answer.
(00:52:32.398):
As more and more people are delivering these harnesses,
(00:52:34.919):
these frameworks for LLMs,
(00:52:36.340):
we’ve got to be really clear that we’re adopting their language,
(00:52:39.022):
their way of working with no ability,
(00:52:42.376):
And we may want to change it, but we need to know to change it.
(00:52:44.978):
And that takes increased levels of expertise.
(00:52:49.680):
So they’re valuable, but they’re also dangerous.
(00:52:52.762):
One of the things I did recently,
(00:52:54.323):
which was highly valuable for me,
(00:52:56.044):
and still my brain hurts very,
(00:52:57.625):
very much, I attended an FCOAM case talk class by Marco Wobber.
(00:53:04.669):
Effect-based oriented modeling,
(00:53:06.411):
which is a completely different way of looking at data models than what I’m used to.
(00:53:12.758):
And at the same token, there are so much similarities.
(00:53:17.243):
But still, I’m trying to grasp my head around it.
(00:53:20.126):
And one of the things I know,
(00:53:21.327):
and I think that’s from the LLM perspective,
(00:53:25.672):
I will create Not now because I need to really grasp it,
(00:53:30.101):
but one of my services from Discovery and Connection Point will be completely based
(00:53:34.706):
on FCLAM.
(00:53:36.188):
And a very simple reason why, then you have another approach to do the same kind of stuff.
(00:53:41.954):
Because it’s, by the way, it’s already incorporated in this tool.
(00:53:45.717):
But I want to do it a little bit different.
(00:53:48.698):
But I think there is a big use for data modelers all over the world.
(00:53:53.680):
And it’s the same,
(00:53:54.380):
I think,
(00:53:54.680):
Joe writes in his latest book, Mixed Model Arts, is saying there is no one approach.
(00:54:01.502):
You need to know more than just one approach and feel comfortable with a couple
(00:54:06.464):
of them,
(00:54:06.804):
at least,
(00:54:07.224):
and know the other ones,
(00:54:08.604):
just to come up with the best idea for the organization you’re hired for,
(00:54:13.406):
you’re working for.
(00:54:15.725):
A quick comment on that.
(00:54:16.966):
I’ve seen a number of people voice opinions such as there are four types of data model.
(00:54:24.852):
There’s Enmon Kimble, Linstead and One Big Table.
(00:54:31.096):
And I sit back and say they are four variations on the analytics space.
(00:54:36.921):
Yep, yep.
(00:54:37.840):
But there are lots more, and Joe Rees has addressed that.
(00:54:43.671):
I would say, by the way, there are two models, useful one and not useful one.
(00:54:51.621):
You know, definition of a model is going to be as bad as definition of a data product.
(00:54:55.122):
All right, we’ve got five more minutes to go, John.
(00:54:56.783):
So we covered a lot.
(00:54:58.603):
We went all over the place, which we always do, which is why I love doing these sessions.
(00:55:01.324):
But coming back to the article that you wrote, right?
(00:55:04.305):
Anything we’ve missed?
(00:55:04.986):
Anything we haven’t covered that was kind of part of the core thesis that you wrote
(00:55:08.527):
in that article?
(00:55:10.207):
Okay.
(00:55:11.428):
I think we’ve covered it.
(00:55:13.809):
I’ve had some notes here.
(00:55:16.169):
But basically that,
(00:55:20.411):
you know,
(00:55:22.080):
The article talks about different levels of visibility of business knowledge
(00:55:28.322):
and I’m just saying that some of those more tightly guarded bits
(00:55:34.525):
of knowledge may be reluctantly or never exposed to AI and that’s
(00:55:42.247):
where I believe that’s the human to human contact,
(00:55:48.130):
building trust, So a quick story and I’ll make this my last one.
(00:55:54.385):
I worked for an organisation and I thought I see an opportunity here.
(00:56:00.547):
So this isn’t IT responding to declared requirements.
(00:56:06.009):
It’s IT seeing technology that might deliver something they hadn’t thought of.
(00:56:11.891):
So I had a relationship with the CEO and I went into the CEO’s office and I said,
(00:56:18.547):
Just playing with this possibility.
(00:56:20.709):
And she said, who have you been talking to?
(00:56:23.331):
And I said, nobody.
(00:56:25.092):
And it was one of those closely guarded secrets for a new product, new line of business.
(00:56:32.358):
And so delightful interaction.
(00:56:36.261):
I’m just saying that when you get the trust of management and have
(00:56:42.225):
that open relationship and Remco, I salute you for the work you’re doing with AI.
(00:56:51.490):
It’s an assistant,
(00:56:53.353):
but call me old fashioned,
(00:56:55.035):
there is still a need for human trust, human respect, human interaction.
(00:57:00.916):
100% Joe, 100% agree.
(00:57:04.810):
I’m very happy that the whole process I’m describing or pattern, Shane,
(00:57:09.093):
if you like to,
(00:57:10.414):
is that all the services I described in my latest book, right,
(00:57:14.476):
they also can be done perfectly fine without any AI because the work is the same.
(00:57:20.660):
And I do agree, right,
(00:57:22.281):
you cannot do something,
(00:57:23.802):
data modeling is not sitting in an ivory tower and do the work.
(00:57:28.206):
It’s talking to people.
(00:57:29.947):
That’s why I liked what Lawrence Kaur was doing, right?
(00:57:34.248):
And I definitely nicked some idea about model storming workshops.
(00:57:39.330):
Get people in a room, get to talk to people, get that trust.
(00:57:42.851):
And that’s the only way you get a data model which is fit for your organization.
(00:57:47.578):
I agree.
(00:57:49.138):
Do I want to sit in a cupboard for six months building
(00:57:51.979):
an enterprise data model diagram that nobody will ever use?
(00:57:55.500):
Hell no.
(00:57:55.920):
Do I want to go and talk to people and get great stories about their business
(00:57:59.221):
and how it works and what they do and then put that into a simple concept map
(00:58:04.102):
that will actually
(00:58:06.022):
help the next piece of work?
(00:58:07.303):
Oh yes, it’s fun.
(00:58:08.760):
And explain that concept map,
(00:58:10.642):
explain the data model,
(00:58:11.902):
show it to the business and tell you,
(00:58:13.864):
you told me this, this is what it come up to, does it match what you told me?
(00:58:18.227):
Because a data model is a communication model.
(00:58:22.607):
So I think putting it back to what Johnny said,
(00:58:25.530):
actually, it’s AI in the loop for this one, right?
(00:58:29.232):
It’s there to assist humans who have done it before to do it maybe
(00:58:35.598):
a little bit faster or a little bit easier.
(00:58:38.860):
And it’s there to help humans who don’t do it very often to understand the steps
(00:58:43.804):
that they should take.
(00:58:46.059):
But it’s like Lego, right?
(00:58:47.360):
There’s different ways in and there’s different ways out.
(00:58:49.462):
You know,
(00:58:49.783):
the real core of a concept model is there’s a bunch of concepts,
(00:58:52.946):
things, there’s a bunch of definitions, and there’s a bunch of relationships.
(00:58:56.650):
That’s the simple part, right?
(00:58:58.391):
Everything else is just complexity,
(00:59:00.333):
complexity of process,
(00:59:01.354):
complexity of organization, Complexity of the person doing the work.
(00:59:04.438):
So I think the key is, you know, sharing is caring.
(00:59:08.743):
The ability to take the knowledge of people like yourselves and share that out there.
(00:59:12.668):
And that’s,
(00:59:13.169):
you know,
(00:59:13.429):
again, thank you, John, for writing the books and writing the articles, right?
(00:59:17.735):
It helps lots of people and it also helps the rest of us think We sit back
(00:59:21.138):
and we read it and we go, is that what I do?
(00:59:24.859):
Like fact-based modeling.
(00:59:27.240):
What we’ve done in the companion app for the book is we now have examples
(00:59:31.121):
where we’re taking both the concepts, the nouns, and example data.
(00:59:36.863):
Customer shame or customer one, two, three.
(00:59:40.084):
Places order, ABC one, two, four.
(00:59:43.805):
Just seeing the data example from the organization next to
(00:59:48.729):
the words gives me more context, right?
(00:59:50.770):
I have more understanding of what that sentence means.
(00:59:54.894):
And so will the LLM, right?
(00:59:57.055):
And so it’s useful.
(00:59:57.756):
So, yeah,
(00:59:59.197):
mixed morals,
(01:00:00.218):
mixing things together,
(01:00:01.138):
all these patterns that we do to find ones that work for you.
(01:00:03.720):
But it’s not an out-of-the-box solution, right?
(01:00:05.762):
It doesn’t remove the human problem.
(01:00:07.912):
Thank you for having us, Shane.
(01:00:27.322):
And I wish you the best, you and you, with the new book.
(01:00:31.144):
Thank you, Shane.
(01:00:31.965):
So do we do.
(01:00:32.625):
You know what writing’s like or both of you do, right?
(01:00:35.006):
I’ll see you in a couple of weeks anyway, Shane.
(01:00:37.628):
You will, yeah, yeah.
(01:00:38.888):
You won’t see the book in a couple of weeks, but we’ll see us.
(01:00:43.231):
We’ll catch up then.
(01:00:44.271):
Excellent.
(01:00:44.651):
All right.
(01:00:45.192):
Catch everybody later.
(01:00:46.372):
See you soon.
(01:00:47.553):
Bye.
«oo»
Information Product Canvas Workshop
“That’s not what I wanted.”
“But that’s what you asked for.”
Shane runs a 6 hour workshop on the Information Product Canvas, a repeatable way to gather data requirements in 30 minutes. The approach he actually uses with stakeholders, not a textbook version that falls apart on contact with messy data and vague requirements.
He runs it regularly.
Latest dates : Gather data requirements in 30 minutes using a shared language
Workshop details : https://maven.com/agiledataguides/info-prod
