AI Adoption the Missing Piece: What Data? In Which System?

For over two years since I started using AI, I’ve been reading everything I can find on AI adoption. The major outlets, the consulting reports, the trade press, the LinkedIn thought leadership. The advice converges on a familiar list. Lock down cybersecurity. Get your data ready. Don’t let confidential information into free tools. Write a usage policy. Define real problems before you reach for a tool. But one piece keeps catching my attention, what data?

It’s reasonable advice. Most of it is correct. But I keep noticing something missing.

The advice talks about data the way a weather report talks about water. Generically. As a category. “Make sure your data is clean.” “Build a strong data foundation.” “Data quality is essential.”

What data? In which system? I’m serious. Read the next ten articles you see on AI adoption and count how many of them name a single business application. I’ll wait.

The gap isn’t an oversight. It’s structural.

The major business publications write for an audience that spans every industry and every tech stack. McKinsey doesn’t know if you’re running Sage, NetSuite, or Acumatica. Harvard Business Review doesn’t know if your CRM is Creatio, Salesforce, HubSpot, or something custom built fifteen years ago. The Economist isn’t going to write about Sage Intacct. So, the writing stays at the level where everyone can nod along. Data foundation. Data quality. Data readiness.

That abstraction is doing real work. It lets a finance leader and a manufacturing leader, and a nonprofit leader read the same article and each get something out of it. Fair enough.

But there is a second conversation happening, and it’s narrower than you’d expect. Vendor white papers, ERP and CRM trade press, and integration consulting content do name the systems. They cover product features, agent connectivity, and integration architecture.

What they mostly don’t cover is the operational work underneath. How do you actually assess whether a chart of accounts is structured well enough for an AI agent to draw correct conclusions from it?

How do you reconcile customer records across sales, service, and finance in a way that holds up when an agent reads them? How do you evaluate whether an HRIS reflects the organization that actually exists today? That guidance is hard to find in either track.

The strategy publications I’ve seen don’t get specific enough to ask the questions. The implementation publications get specific about the wrong things.

Here is what I see consistently. The data inside a business doesn’t sit in a generic pool. Not in any LLM, unless of course someone accidentally put it there. It sits in specific systems, each with its own logic, its own conventions, and its own accumulated history.

The financials live in an ERP. The customer relationships live in a CRM. The employee records live in an HRIS. The work specific to the industry, whether that’s project accounting in construction, donor management in nonprofits, patient tracking in healthcare, or production scheduling in manufacturing, lives in a line of business system that often nobody outside that industry has heard of.

Each of those systems has a data model that reflects decisions made years or sometimes decades ago. Each carries customizations layered on by people who may no longer be at the company. Each has fields that were repurposed when the original use case changed. Each has integrations to other systems that work most of the time and fail in ways that nobody has fully documented.

This is the actual substrate. Not “data” in the abstract. A chart of accounts somebody designed in 2014. A customer record schema that was reasonable when the company sold three products and now strains under thirty. An HRIS that still reflects an org structure from before the last reorganization. And it’s true. I see it every week.

When the strategy article says “make sure your data is ready,” this is what it’s quietly asking you to do. Make sure the chart of accounts is clean. Make sure the customer records are consistent across sales, service, and finance. Make sure the HRIS reflects the company that exists today, not the one that existed three years ago. Make sure the line of business system that runs the actual work of the company has data a reasonable person, or a reasonable model, could draw correct conclusions from.

That is a much bigger ask than the abstract version makes it sound. I see it so often when my clients were about to upgrade their legacy system to modern technology. “Oh, my data is garbage.”

The shift is from one question to several. Instead of asking “is our data ready for AI,” try asking the questions one or two or three levels down.

Is our chart of accounts structured well enough that an AI agent reading it would draw the right conclusions about how the business actually operates? Or has it accumulated enough inconsistencies over the years that it would mislead a careful human reader, let alone a model?

Are the customer records in the CRM consistent across the people who touch them? Does sales see the same customer that service sees, that finance sees, that the executive team sees in the dashboard? Or are there four versions of the truth, reconciled informally by whoever happens to be in the meeting?

Does the HRIS reflect the organization that exists today? The reporting lines, the roles, the responsibilities, the people? Or is it carrying forward a structure from before the last reorganization, with workarounds that everyone in HR understands and nobody outside HR could decode?

Are the customizations made to the line of business system documented anywhere a person joining the company today could find them? Anywhere a model could find them?

These questions are concrete. They are answerable. They are also, in my experience, where most companies get stuck long before AI enters the picture. The work of getting ready for AI turns out to be, in large part, the work of getting the underlying systems into the shape they should have been in already.

That isn’t a flaw in the AI conversation. It’s a feature of it. The pressure to adopt AI is forcing a level of attention to the systems of record that probably should have happened years ago, and would have paid off years ago, and didn’t because there was never a forcing function.

What changed is where AI sits.

When AI was a chatbot you visited in a separate window, generic data advice was enough. You copied something in, you read what came out, you decided what to do with it. The systems that ran the business were not really part of the loop.

That arrangement is ending. Microsoft Copilot is embedded in Dynamics. Sage Copilot is rolling out across Sage Intacct and other Sage products. NetSuite has Oracle’s Fusion AI built into its core. Salesforce has Einstein across sales and service. Workday is embedding agents in HR workflows. In February, OpenAI announced its Frontier platform with McKinsey, BCG, Accenture, and Capgemini, explicitly positioning it as a layer that lets agents act across CRM, HR platforms, and ticketing tools.

The pattern is unambiguous. Every major business application vendor and every major AI player is racing to put agents inside the systems of record. Can you imagine the impact of that? Agents acting directly inside the systems where the work happens. Not in a window off to the side. Inside.

That changes the stakes of the questions above. The condition of your chart of accounts, your customer records, your HRIS, and your line of business system used to matter mostly to the people who used those systems daily. Now it matters to anyone trying to deploy an agent against them.