The conversation everyone is having about AI right now is at the wrong altitude. Vendors pitch capability. Boards ask about adoption rates. Trade press publishes breathless lists of "what every leader must know." Meanwhile the operator who has to write the check gets almost no help with the only question that matters: does this fit my business, in its current state, right now?

Not in the abstract. Not "AI is going to change everything." In your operation. With your data. With your team. In your market. This week. Yes or no. And if yes, where. Under what conditions. What has to be true first.

I sat on a panel last quarter with people from Mastercard, Boeing, and Google Cloud. The question I kept hearing from the operators in the room (the ones who would actually have to write checks) wasn't which tool should we buy? It was how do we know we're not getting played? That's the right question. The rest of this piece is a useful version of the answer.

What follows is seven questions. None of them is technical. All of them belong in the room before money moves. If you can't answer them clearly for your business, the next conversation isn't about AI. It's about everything underneath it.

Why this matters more than the hype suggests

Most "AI failures" aren't AI failures. They're misdiagnosis. The tool was wrong because the problem was wrong. Someone bought a capability that didn't address the actual constraint, and then spent six months trying to make the new thing solve the old thing it was never going to solve.

The cost isn't just the budget. It's the leadership time spent on the wrong fix while the actual problem keeps getting worse. It's also credibility. When the AI initiative doesn't deliver, the next operational change becomes harder to fund, harder to staff, and harder to believe.

The competitive panic is real. Your peers are doing things with AI. Some of those things are working. Some are theater. The panic itself isn't the problem. The unstructured response to it is.

Question 01What problem are we actually trying to solve, and is AI the right tool for it?

Most AI projects begin with the tool, not the problem. Someone saw a demo. Someone read an article. Someone's brother-in-law sold them on it at a wedding. That's not strategy. That's ambient pressure dressed up as a roadmap.

A useful filter: most operational pain falls into one of four categories. Only one of them genuinely responds to AI the way it's being sold.

A concrete example. A company replaces its first-line customer support with an AI agent. The agent handles inquiries faster. Complaint volume drops. Leadership declares victory.

What actually happened: the underlying return policy was unclear and inconsistent. The human agents had been quietly papering over it for years with judgment calls. The AI agent doesn't paper over anything. It applies the unclear policy faster and at higher volume. Six months later, the policy gap surfaces in court, in a viral post, in a regulatory complaint. The AI didn't help. It accelerated the failure to a more public stage.

Information problems respond to AI. Decision and motivation problems get worse with it.

Better question to start with: name one specific cost or constraint AI would relieve. Time. Money. Risk. Attention. If you can't say what gets smaller in concrete units, you don't have an AI problem yet. You have a curiosity. That's fine. Just don't fund it like a strategy.

Question 02Is our data clean, structured, and accessible enough for AI to add value?

AI without good data is autocomplete with confidence. It produces something that sounds right. It does not produce something that is right.

Most owners think they need more data. They almost always need cleaner data. "Clean" in plain terms means three things. Consistent: the same field names, the same formats, the same units. Current: not stale, not last quarter's snapshot pretending to be live. Complete on the dimensions you care about.

Here is the diagnostic I use. Pick one specific record. One customer, one project, one transaction. Ask your team to retrieve everything you know about that record in under thirty seconds. Notes from sales. Current contract terms. Support history. Last invoice. Stated preferences. Not the answer they think you want to hear. The actual answer.

If your team can't, AI will not help. AI will produce confident-sounding output from incomplete inputs, faster than before, and now you've got the same gaps wrapped in authoritative language.

A second test. Hand the same record question to two different people in your organization. Do you get the same answer? If not, AI doesn't fix the disagreement. It averages it and serves the average back to you as truth.

The work that closes this gap isn't glamorous. Data hygiene. Naming conventions. Retention policies. Process discipline. But this is the gate. Skip it and you're paying to amplify confusion.

Question 03What happens if the vendor changes pricing or shuts down?

Vendor risk is the silent cost of every cloud AI subscription. The platform you build a workflow around today may price you out next year, get acquired and re-priced, pivot to a different customer, or quietly degrade the version you depend on.

I've watched companies wire their customer-facing experience around a single AI provider, then receive a 4× pricing email at the end of a quarter with thirty days notice. The contract had no migration clause. The "savings" the AI was supposed to deliver were nominal compared to what it cost to rebuild the workflow on something else, under time pressure.

The honest framing isn't should we use this vendor? It's what does it cost to leave?

Three things to know before you sign:

None of these are exotic edge cases. Pricing changes happen. Acquisitions happen. Outages happen. The companies that handle them well planned for them at the contract stage. The ones that didn't end up making infrastructure decisions in crisis mode, which is where the most expensive mistakes live.

Question 04Who in our organization is accountable for the outputs AI produces?

"The AI got it wrong" is not a defense.

Not to your customer when an order ships incorrectly. Not to your auditor when financial categorization is off. Not to your board when a hiring tool produces a discrimination claim. Not to a regulator when something disclosed turns out to be wrong.

Accountability without authority is theatre. Authority without accountability is risk.

The fix isn't we'll be careful. The fix is naming a human (by role, ideally by name) who reviews and signs off on AI-driven outputs in each domain where they matter. Not generally. Specifically. Marketing copy is one domain. Customer-facing automated replies are another. Pricing recommendations are a third. Each one needs an owner.

And that human needs the authority to override the AI, slow it down, or shut it off when their judgment says so. Without that authority, the role is a fig leaf.

I've seen the failure mode where a tool was bought by marketing, deployed in customer service, queried by sales, and accountable to nobody. By the time something obviously broken happened, three departments were pointing at each other and at the tool. None of them owned the result. The customer didn't care which department's name was on the org chart.

Question 05How will we know if the AI is making the right decisions, and who checks?

This is the question most often skipped, because the honest answer is uncomfortable.

"It seems to be working" is not measurement. It's the absence of measurement, dressed up as reassurance.

A measurement protocol has three parts. First, a signal you actually look at: a metric, a sample, a complaint pattern, a comparison to a known-correct answer. Second, a frequency on the calendar (weekly, monthly, every release), not when something feels off. Third, a person whose job description includes reading that signal and acting on what it says.

Without all three, AI drift is invisible until it is expensive.

Drift is the dirty secret of deployed AI. The thing that worked at launch can degrade slowly. The input distribution shifts. The model gets quietly retrained by the vendor. The underlying engine gets swapped. You only notice when a customer complaint reaches the right ear, or when a regulator does.

A pragmatic check that any business can run: pull twenty AI outputs at random every month. Have a human evaluate them against a clear standard you wrote down at deployment. Track the score. When it dips, investigate before you ship more on top of it.

This isn't optional infrastructure for using AI seriously. It's the cost of using it at all. If you can't afford the measurement layer, you can't afford the AI.

Question 06What are we replacing, and what are the failure modes of the replacement?

Every AI deployment replaces something. A process. A person. A step. A judgment call.

The thing being replaced has known failure modes. You've lived with them for years. You know what a bad day looks like. You know who notices first and what the fallback is.

The replacement has unknown failure modes. They will emerge under load, in edge cases, with input shapes you didn't test. Knowable risks beat unknown ones, even when the knowable risks look higher in absolute terms on the comparison spreadsheet.

An example I've seen play out more than once. A company replaces the human triage step in support with an AI categorizer. The human was 90% accurate on the test set. The AI is 94%. It looks like a clean upgrade.

What they didn't catch: the human's 10% of errors were random and obvious. Caught quickly, corrected easily. The AI's 6% are systematic. Concentrated in non-English-speaking customers and one specific issue category. Aggregate accuracy looks better. Customer experience for a specific segment got dramatically worse, and nobody saw it for months because the dashboard rolled up to a single number.

Accuracy on the average is not accuracy on the customer who is paying you.

Better practice: write down what failure looks like for the new system before you deploy it. What does a bad day look like? Who notices first? What's the fallback? What metric, broken out by which segment, would tell you something is going sideways before the customers do?

Question 07Are we buying AI because it solves a problem, or because we feel behind?

This is the ugly question. It's also the most useful one.

Most AI investment in mid-market companies right now is not strategy. It's a response to peer signaling. Everyone we know is doing something with AI, we should too.

Buying AI to look modern is buying compliance with a peer story, not capability. The expense is real. The operational return is wishful.

A test I'd offer any executive considering an AI investment: would you buy this exact tool for this exact use case if your last three operational initiatives had been wildly successful and you had a track record of clean execution? Or are you buying it partly because you need a "yes, we're doing something with AI" answer at the next board meeting, investor call, or industry event?

If it's the latter (and for a lot of companies right now, it is), the right move isn't to skip AI. The right move is to be honest about what you're buying. You're buying an answer to a question, not capability. Sometimes that's worth paying for. Just name it correctly so you don't blame the tool when it doesn't do what it was never going to do.

The peer pressure is real. The pressure is information. Your peers are also figuring this out, and what they're doing is data. But it's not a strategy. Confusing the two is what buys you a tool you don't need at a price you didn't budget for, with no plan for what happens after the pilot.

Operations first. Tools second.

AI is good at amplifying what you have. Speed, consistency, scale. If what you have is clean, AI amplifies clean. If what you have is broken, AI amplifies broken. Faster, at more volume, with more confidence, in more places.

The companies I see winning with AI right now aren't the ones who adopted earliest. They're the ones who built their operations clean enough that AI could plug in without exposing chaos. Their data was ordered before they bought the tool. Their decisions had owners before they automated them. Their measurements existed before they needed to evaluate whether the new thing was working.

This isn't anti-AI. AI is going to reshape categories of work I haven't even imagined yet. The question isn't whether to engage. The question is whether your business is in a state where engaging produces leverage, or whether it produces noise.

The seven questions in this piece aren't a test you pass or fail. They're a flashlight. They show you which parts of your operation are ready for AI to amplify, and which parts are going to get amplified into a problem you didn't have yesterday.

If you can't answer them clearly, the next conversation isn't about AI. It's about everything underneath it.