Your AI project will inherit your data problems
Data & Report Integrity · Responsible AI Adoption & Governance
There's a version of the AI conversation happening in a lot of operations businesses right now, and it goes something like this: our reporting is messy, our systems don't talk to each other, half our numbers live in spreadsheets — so let's bring in AI to sort it out.
I understand the instinct. It's also backwards.
AI doesn't sit above your data problems. It sits downstream of them. Whatever is wrong with your data today — the untraceable numbers, the manual edits nobody logged, the records that quietly fall between systems — an AI model doesn't correct any of that. It consumes it, learns from it, and produces outputs with the same flaws, delivered faster and with more confidence than any spreadsheet ever managed.
That last part is what makes this genuinely risky rather than just inefficient. A messy spreadsheet looks messy. People instinctively treat it with suspicion, double-check it, ask around before acting on it. An AI-generated forecast or an automated recommendation doesn't look messy. It looks finished. It arrives with clean formatting and an implied authority, and the human instinct to double-check quietly switches off. The polish launders the problem.
So the failure mode isn't "the AI didn't work." It's worse: the AI worked exactly as designed, on inputs that were wrong, and everyone trusted the output because it came from the new system leadership just invested in. Wrong numbers used to travel at the speed of a weekly report. Now they travel at the speed of an API.
This is why, when a leadership team tells me they're ready to invest in AI, my first questions aren't about models or tools or vendors. They're the same questions I'd ask about a weekly report:
For the process you want AI to improve — can you trace its key numbers back to source systems?
Do you know how much of that data gets manually corrected before anyone sees it, and who corrects it?
If a record dropped out of the pipeline entirely, would anything tell you?
If those answers are solid, an AI initiative has a foundation. If they're not, the honest sequencing is: fix traceability first, then automate. It's less exciting than launching a pilot, and it's also the difference between an AI project that compounds value and one that compounds error.
There's a budget argument here too, not just a risk argument. Data integrity work is cheap relative to AI implementation. Finding out mid-project that your training data can't be trusted is the expensive version of the same discovery — you pay for the AI build, then you pay for the data cleanup anyway, then you pay again to rebuild confidence with the team that watched the first attempt stall. Sequencing isn't caution for its own sake. It's the cheaper path to the same destination.
None of this is an argument against AI in operations. It's an argument for order of operations. The companies getting real results aren't the ones that moved first — they're the ones whose data could support the weight of what they built on top of it.
We wrote a short playbook on exactly this sequencing — where to start, what to govern, and how to know whether it's working. It's a free download, no pitch attached.
And if you already suspect your data couldn't survive the three questions above, that's the gap our Report Integrity Assessment closes — a fixed-scope review of what your reports actually stand on.