A hospital IT director and a finance analyst at a mid-size manufacturing company have almost nothing in common professionally. But last year, both of them told me the same thing within a few weeks of each other: the AI tool they adopted didn’t save time at first. It created more work, because nobody had cleaned up the mess underneath it.
That’s the part vendors don’t put in the demo.
The Myth of Plug-and-Play AI
Every AI tool on the market right now promises minimal setup. Type a prompt, get an answer, done. And for simple use cases, that’s mostly true. But the moment AI touches something with real structure underneath it, whether that’s patient records or a ten-year-old financial model with broken formulas, the tool is only as good as what it’s reading.
I’ve watched a clinic try to roll out an AI scheduling assistant that kept recommending appointment slots that didn’t exist, because the underlying calendar system had three overlapping data sources nobody had reconciled. The AI wasn’t wrong. It was accurately reflecting a mess. Same story on the spreadsheet side. Analysts feed AI tools a workbook with hardcoded values disguised as formulas, and the AI produces confident, wrong output because it has no way of knowing the numbers were never live in the first place.
Healthcare’s Version of the Problem Is Higher Stakes
In healthcare, the stakes aren’t just accuracy. They’re legal exposure.
A physical therapy practice I spoke with wanted to let patients text questions to their care team and get AI-assisted triage on response priority. Reasonable idea. Except most consumer messaging platforms retain data in ways that violate basic privacy requirements the moment a patient mentions a diagnosis or medication. The practice ended up switching to a HIPAA-compliant messaging app specifically because it needed encrypted storage and access logging built into the same layer where the AI was reading and summarizing messages. Bolting AI onto a messaging tool that wasn’t designed for protected health information doesn’t just create risk. It creates risk you often don’t discover until an audit, which is the worst possible time to discover it.
This is the pattern across healthcare AI generally. The AI layer gets attention because it’s exciting. The infrastructure underneath, the part that determines whether the whole thing is even legal to run, gets treated as an afterthought until someone in compliance asks a question nobody can answer.
Spreadsheet Users Are Running Into a Quieter Version of the Same Thing
Finance and operations teams don’t have HIPAA to worry about, but they have their own version of institutional risk: decisions made on bad numbers that nobody catches until the quarter’s already closed.
Excel AI tools have gotten genuinely good at flagging anomalies, writing formulas, and summarizing trends across tabs that would take a human twenty minutes to scan manually. A retail operations manager I know uses one to catch inventory discrepancies across regional spreadsheets that used to require a Friday afternoon of manual cross-checking. That’s real value. But it only works because her team spent months standardizing how those spreadsheets were structured in the first place. AI reading twelve different formatting conventions across twelve regional offices would produce twelve different kinds of nonsense.
The uncomfortable truth is that AI rewards organizations that already had reasonably clean systems, and punishes the ones that didn’t, often by amplifying existing errors faster than a human ever could.
What Both Groups Eventually Learn
Give it six months, and the healthcare IT director and the finance analyst arrive at the same conclusion from opposite directions: the AI itself was never the hard part. Getting the underlying system, whether that’s a communication platform or a spreadsheet workflow, into a state where AI could trust what it was looking at, was the actual project.
Nobody budgets time for that part. They budget for the AI subscription and assume the rest takes care of itself.
It doesn’t. And the organizations getting real value out of these tools right now aren’t the ones with the flashiest AI features. They’re the ones that did the unglamorous work of fixing their data and their compliance posture first, so that when the AI finally showed up, it had something solid to stand on.
