A company runs its first internal AI pilot. The team asks the assistant a reasonable question: "What's the status of the Henderson project, and what did we quote them last time?"
The assistant answers with confidence. The project is at 60% completion, the last quote was $84,000, and the client had concerns about timeline.
None of it is true. There is no Henderson project at 60%. The $84,000 quote was for a different client. The "timeline concerns" came from a Slack message taken out of context.
The team's conclusion is usually wrong. They blame the model. They try a different one. Same result. They conclude AI "isn't ready yet" and shelve the project for a year.
The model wasn't the problem. It did exactly what it's designed to do: produce a plausible answer from the data it can access. The data it could access was fragmented across five tools, half of it duplicated, none of it labeled in a way a machine could read. So it filled the gaps with statistical guesses. That's what hallucination is.
Most companies thinking about AI in operations are running the wrong sequence. They're trying to add an intelligence layer on top of an operation that hasn't been digitized properly yet.
The data exists. That's not the issue. Every business that has been running for more than five years has enormous amounts of data. The issue is where it lives and how it's structured.
It lives in someone's email thread. In a spreadsheet on a salesperson's laptop. In a project tool that no one connected to the CRM. In an invoice that lives in QuickBooks but was never associated with the project it belongs to.
An AI model looking across that landscape sees fragments. It can't tell that the "Henderson" in your CRM is the same client as the "Henderson Industries" in your project tool and the "H. Industries LLC" in your accounting software. So it guesses. And the guess sounds confident because that's how language models work — they don't say "I don't know," they produce the most probable-sounding answer.
The Henderson example is the version everyone recognizes. There's a harder one that hits closer to the actual operation of a company that sells projects.
Try asking an AI assistant: "Which of our projects from last year were actually profitable, and what did the profitable ones have in common?"
To answer that, the AI needs to know which deal became which project, which invoices belonged to that project, which costs got logged against it, and what the original quote assumed. In most operations, none of those things are explicitly connected. The deal lives in the CRM. The project lives somewhere else. The invoices live in accounting. The costs live in a spreadsheet the operations director keeps. The original quote is a PDF in someone's email.
The AI can't reason across that gap. So it either refuses to answer or, worse, it produces a confident analysis that's invented. And the question you just asked — the one that should be the most basic question any company that sells projects can answer about itself — is unanswerable not because the data doesn't exist, but because nothing connects it.
The phrase gets used a lot in AI conversations, and it's worth being precise.
It doesn't mean clean. It doesn't mean tagged. It doesn't mean stored in a database instead of a spreadsheet.
It means: every important thing in your business has one place where it lives, one definition of what it is, and explicit relationships to the other things it connects to.
A client is one record, not five. A project is connected to the deal that originated it, to the invoices that came from it, and to the people who worked on it — explicitly, not through someone's memory. A conversation about a project lives attached to the project, not in someone's inbox.
When that's true, an AI can answer questions because the answer is traceable. When it's not true, the AI has to invent the connections that don't exist in the data — and invention is hallucination.
There's a tempting shortcut: just connect everything with APIs. Plug the CRM into the project tool into the accounting software into the AI layer.
This doesn't work for a reason that's easy to miss. Connecting tools moves data between them — it doesn't reconcile what that data means. If your CRM defines a "project" differently than your project tool does, and your accounting software has a third definition, an API just makes the confusion travel faster. You end up with an AI that has access to more contradictory data and produces hallucinations that sound more authoritative because they reference multiple sources.
The fix isn't more integration. It's deciding which system is the source of truth for which kind of information, and structuring the rest of the stack around that. That's an architectural decision, not a technical one. And it has to happen before AI can be useful.
If you want to know whether your operation is ready for AI without committing to anything, try this. Pick three questions you'd actually want an AI assistant to answer about your business:
1. "What was the gross margin on the Henderson project, and which costs came in over budget?"
2. "Which of our clients from last year had the most repeat business, and what did the first project we did for them look like?"
3. "Of the deals we lost last quarter, which ones came back to us later and which didn't?"
Now, instead of asking AI, ask yourself: where does the answer live? If your honest response to any of these is "I'd have to pull it from three tools and reconcile it manually," that's exactly where an AI will hallucinate. The systems you'd struggle to query are the systems no AI can query either.
That diagnostic is free. It also tells you, in 10 minutes, whether your operation is structured to benefit from AI or whether you'd be building intelligence on top of fragmentation.
There are two phases to using AI in operations, and they have to happen in order.
Phase one is structural. Get your operation into a state where every important thing has one home, one definition, and explicit relationships. This is the unglamorous work. It's also the work that determines whether phase two produces value or produces hallucinations.
Phase two is the AI layer. Once the structure exists, AI has something real to work with. It can summarize, recommend, draft, retrieve — and the output is grounded in operational truth instead of statistical guesses.
The companies skipping phase one are the ones writing posts about how AI "isn't ready yet." It's ready. Their operation isn't.
If you're running a company that sells projects — consulting, professional services, technology, architecture, any business where what you sell doesn't exist until you execute it — the structural work is more specific than "clean up your data." It's about how deals, projects, clients, and financial data relate to each other in your systems.
That architecture is what we build at SAP ASAP. Not as an AI implementation — as the foundation that has to exist before AI can do anything useful in your operation. You can see exactly how that foundation looks on the page for companies that sell projects, including how the data structure connects what your team is working on today to what AI can do with it tomorrow.