Key takeaways
- "We want AI on our data" names a solution. The sharpest test of whether a real problem sits behind it is whether the executive who owns that problem will pay for it.
- A chat box on company data mostly saves the wait for an analyst, which is hard to fund. Practitioners see the value in agents that take over a step of work inside a process.
- Before an AI reads company data, the few numbers it uses need one agreed definition each, built for that one use case rather than for the whole company.
- Six questions, with who pays as the gate: whose problem it is, who pays to move which business number, whether a report would do, which numbers need defining, a small pilot, the next case.
- If the owner won't commit a meaningful budget, or named people's time where budgets aren't held, stop before data work. Passing that gate shows the problem matters, and the pilot shows whether the agent works.
In this article
Sooner or later someone in the management meeting says it: “We want AI on our data.” It sounds like a plan, and practitioners say it’s a common request reaching data teams today.
The trouble is that the sentence describes a solution and says nothing about the problem it’s supposed to solve.
People whose job is putting AI on top of company data describe the same split. Start from the technology and you usually get a nice demo that nobody uses two weeks later. Start from the problem and you get a project with an owner, a budget and a way to tell whether it worked.
And the sharpest way to tell which one is in front of you is to ask who’s willing to pay for it. It’s the second of the six questions below, and it decides whether the rest are worth asking.
Two terms worth knowing
Chat with your data is a box where a manager asks a question in plain language, say “how many new customers did we get last quarter?”, and the system turns it into a query on the company’s database and returns the answer.
A semantic layer is an agreed business dictionary that sits between the data and the people who use it. It sets what counts as revenue, who counts as an active customer and how each measure is calculated. Without one, an AI system can calculate the same measure three different ways and give back three different answers.
Why a chat box on the data is a weak business case
A chat box mostly saves the wait for an analyst’s answer. That’s convenient, but as practitioners put it, it’s hard to get a senior executive to sign off on hundreds of thousands of dollars so answers arrive a week sooner.
There’s a human reason too. Practitioners find that many managers prefer an analysis from an analyst they trust, who brings context, a story and judgment, and who can take a follow-up question.
Even when the chat box is accurate, it doesn’t replace that trust, and those managers tend to drift back to their old habits after a few days.
Where the value shows up
The projects practitioners see move forward are agents doing work inside a business process. An agent is an AI system that carries out several actions in a row to finish a task, rather than answering one question and stopping. For how that changes adoption, see an agent gets adopted differently from an app.
Two examples practitioners describe make the difference concrete. At a bank where closing the monthly books takes weeks of work from a large team, the CFO sets a target of cutting that time by 90% and is ready to fund it from the finance budget.
At a company that runs clinical trials for drug makers, leadership maps every stage of the service and chooses which stages to automate, because its clients want faster and cheaper service.
The bank’s case shows what practitioners look for before anything is built: a senior executive with a defined problem, a target in numbers and a budget to pay for it. The clinical-trials case shows the other half: map the whole process first, then choose which stages to automate. In both, the plan is for an agent to take over part of the work in the process itself, which is more than shortening a wait, and both start from the process, the same place redesigning the work before choosing the tool starts.
Why the data people will ask about definitions first
When a business unit builds an agent and asks for “just access to the data,” the people who run the data see the risk straight away. They know what direct access tends to produce: an agent that reads raw tables, works out its own version of revenue, and runs up computing costs while it does.
At that point, agreeing what the numbers mean stops being an infrastructure project that’s hard to fund and becomes a condition for something a business unit already wants. It applies once an AI is reading the data, and it doesn’t need to cover the whole company, which would be a huge project. It needs to cover the one use case in front of you.
A good answer from the data side comes in business terms: what it’ll cost, how long it’ll take, and a quick start. When that answer is slow, practitioners say, the business unit tends to try building the thing alone.
Six questions, from the request to the next case
The second question is the gate, and question 3 can still stop the project before the pilot, which is where the real cost starts.
1. Whose problem is it, and is the request real? Ask the manager making the request what they’re trying to achieve, what’s urgent and what they’d put their name on. If what they’re asking for is a chat box, check whether they use the dashboards that already exist, because if those go unopened, a chat box probably won’t change the habit either.
Start here, because the gate in question 2 needs a person to put its question to.
2. Who pays, and what’s the one number? This is the gate. A success measure is a number that says in advance what’s supposed to change. The sharpest test of whether the problem is real is whether the executive who owns it puts a meaningful amount from their own budget behind it, the way the bank’s CFO is ready to fund a faster close.
A manager who’s excited but won’t pay is usually telling you the problem isn’t urgent. Where managers don’t hold budgets, look for a different sign of commitment: key people, named, whose time is set aside for the project. If there’s no meaningful budget, or no named people’s time where budgets aren’t held, stop here.
Passing the gate tells you the problem matters enough to the person who owns it to pay for it. It doesn’t tell you the agent will work, and finding that out is what the pilot in question 5 is for.
Money comes before any work on the data because asking takes one conversation, while every later question takes other people’s time.
3. Is there data that can answer it, and would something simpler do? Check that data exists that can answer the problem at all. If there’s more than one place in the process an agent could start, start where the data is tidiest, so results come quickly.
Sometimes a scheduled report by email is enough. And if the job is narrow enough, a handful of fixed queries written in advance can do it with no AI model and no business dictionary, in which case there’s no agent to build and the remaining questions don’t apply. Check this before the pilot, because a report that does the job saves building one.
4. Are the numbers this agent will use defined? The few measures it will touch each get one written definition, agreed with the business owner, before the agent reads a table. Everything else in the company’s data can wait.
Settle them before the pilot, or nobody can trust its numbers.
5. Is the pilot small, and built with tools you already have? Don’t start by buying a product. A small agent built in Microsoft Copilot, OpenAI’s ChatGPT, Anthropic’s Claude or whatever the company has already approved shows where the hard part is, whether that’s building it, keeping it running or checking its answers, and only then do you know whether a dedicated product is needed.
The pilot sits this late because it’s the first step that costs weeks of work and real money.
Six to eight weeks is a reasonable target. One week won’t be enough, and a project that runs for several quarters tends to lose its business owner along the way. At the end, the owner looks at the number agreed in question 2 and should have seen enough of a result to say whether this is the direction.
Put limits on it while it runs. Give the agent read-only access to only the tables it needs, have an analyst check its output before anyone decides anything on it, and limit the kinds of requests it takes at first, so trust builds step by step.
6. What’s the next case? After a successful pilot, go to the next use case, perhaps with three or four running side by side, and send each one back through the gate with its own business owner, just as the first one went. Jumping from one to a hundred is the move to avoid.
Where this doesn’t fit
Smaller companies with nobody who owns the data. What decides it is whether anyone is responsible for the company’s data and whether that data is organized. Without both, most of these questions don’t apply yet. It’s better to work with an outside firm or specialist and start small than to hand a big project to one person early in their data career.
Organizations where managers don’t hold budgets. The money test in question 2 needs a substitute there, and the natural one is key people’s time, named and set aside for the project.
Personal experiments on sample data only. If the company doesn’t allow these tools yet, a personal subscription to Anthropic’s Claude, OpenAI’s ChatGPT or xAI’s Grok at around $20 to $30 a month is a cheap way for someone to learn what can be built. No customer data or company data goes into it.
What to do this week
- Pick one manager who has asked for AI on the data and ask what’s slowing them down today. If they want a chat box, check whether they use the existing dashboards.
- Agree one success measure, in numbers, and ask whose budget pays for it and from which line. If there’s no meaningful budget, or no named people’s time where budgets aren’t held, stop here.
- If it clears, ask the data people whether the data exists and whether a scheduled report would do.
- Ask which numbers the agent would use, and have each one defined.
- Only then, ask whoever runs the data to build a small version in a tool the company already approved, with read-only access to only the tables it needs and an analyst checking its output.
The request will keep arriving in the same words. In the projects practitioners describe as moving forward, an agent does real work in a process, and someone who owns the problem, like the bank’s CFO, has a number and the budget to back it. And in those projects, agreeing what the numbers mean finally has a reason behind it.
Questions this article gets
Is a chat-with-your-data tool worth paying for?
On its own, it's usually a weak business case. A chat box mostly saves the wait for an analyst's answer, and people whose job is putting AI on company data find that many managers drift back to their old habits after a few days, because they prefer an analyst who brings context and judgment. The projects that tend to move forward are agents that take over a step of work inside a process, backed by an executive who owns the problem and is willing to pay for it.
How long should a first AI pilot on company data run?
Six to eight weeks is a reasonable target. One week isn't enough to learn where the hard part is, whether that's building the agent, keeping it running or checking its answers, and a project spread over several quarters tends to lose its business owner along the way. At the end, the owner should be able to look at the agreed number and say whether this is the direction.
What access should an AI agent get to company data during a pilot?
Read-only access, and only to the tables it needs. An analyst checks its output before anyone makes a decision on it, and at first it takes a limited set of request types so trust builds step by step. Open, direct access tends to produce wrong numbers and high computing costs.
