AI, Architecture, and Ontologies

One of the hottest topics in technology right now is the role of semantic architecture in AI workflow integrations. The short story version of it is this: your organization wants to use AI to better retrieve, organize, and understand the data it has.

Think of this as an AI use case in which the AI is a layer of logic on top of your organizational data. That way, the AI will figure out the path through the confusing data landscape and just give your teams the best responses based on its synthesis. Sounds great!

Mirage #1: AI Will Navigate the Data Perfectly

But, as a technologist, data analyst, or digital manager, you know that your org data higgledy-piggledy (technical term). It’s currently held in multiple, fragmented systems, all of which are built on datasets that differ logically and semantically. The promise of AI is that this confusion won’t matter anymore. 

But that promise is a mirage. Data logic matters because AIs and humans work in logical structures. If the logical structure isn’t there, the AI can’t do great work. 

Even in the best situation, where everyone who works in the org holds similar understandings of the organizational purpose and works well together well, data and disciplines differ widely from each other.  

The big problem is that in almost no organization at any scale actually has that best situation. People do not have a unified vision of what orgs do, nor do they work well together all the time. Layer on top of that the years separating different leaders’ tenure and a reasonable evolution in organizational focus, and a complete breakdown in logic across datasets becomes almost an inevitability.

As an organically emerging hack, organizations end up with people specializing in how certain systems work, aka accruing institutional knowledge, and these people don’t tell anyone how these systems work because they’re not incentivized to. No one actually cares how internal systems work: they just want the system data presented to them in a way that’s useful to their specific needs. 

Mirage 2: The RAG Model Promise

So what’s an organization in 2025 to do? Train a few special AIs, of course! Throw together RAG models that specialize in your specialized datasets, bring them to a unified chat interface and suddenly everyone at the company can see across the system data. That will get those individuals specializing in how these esoteric systems work out of their silos and working together! 

Well, that’s the idea anyway. 

The problem is that training a RAG model on poorly built datasets leads to model outputs that are based on partial or deeply flawed inputs. Garbage in; garbage out. The best answer is to solve the root problem, not layer another technology on top of it.

But that is hard. Solving the root problem here means unpacking how datasets are structured, working with those system specialists to do so, and rebuilding them in a way that is congruent with the logic implemented in other datasets. 

Bring Digital Management to the Problem

That’s not a technical problem: that’s a human one. That means ICs, managers, and C-suite all have to want to do this backend work. That means that we have to convince our boards that the data foundations needed to implement lightening-fast, lean, and smart AI workflows are worth spending two or three fiscal quarters on. That means convincing teams to keep track of how long they spend analyzing data now and then produce diff data after AI system implementation. 

All of this is absolutely worth it, and I’ll go into why and how in Part Two: Semantic Solutions next week. But the short answer is to accept that AI tooling and workflow integration isn’t a magic bullet, and that investment in a few weeks of foundational, human-centered logic work is well-worth the investment.  

Human-center your data structures, and your AI workflow implementation will take off. Set your AI up for success by doing the hard, intelligent, foundational work. If you do, you will springboard into the AI era in just a few weeks.

Read next week to find out how we at Ishmael Interactive approached this problem in a 12,000 person organization, and book a call with us to learn more about how to do it in your organization.