Five questions to ask before buying another AI tool

Every team we work with has more AI tools than it can name. Some arrived through a considered decision. Most arrived because someone saw a demo, a competitor mentioned it, or a free trial quietly converted into a line item. The result is a stack nobody fully understands and a budget nobody can defend.

This is tool sprawl, and it has real costs. There is the obvious one: licence fees adding up for software that nobody opens. But there are quieter ones too. People context-switch between tools that partially overlap, so nothing gets used well. Data ends up in systems that were never properly vetted. The mental load of managing a fragmented set of tools falls on whoever is already most stretched. And when something goes wrong, nobody can trace which tool touched what.

If you have ever looked at your organisation's software subscriptions and felt vague dread, this post is for you. These five questions take about thirty minutes to answer together. They will not eliminate every bad purchase, but they will stop the most common ones: the reflex buy, the trial-that-never-ended, and the tool that duplicates something you already have.

1. What specific job does this do that we can't already do?

Not "what can it do." What job, for whom, that your existing tools genuinely cannot handle?

This distinction matters because AI tools are marketed at the level of capability: it summarises, it drafts, it analyses. Almost every tool can do those things in some form. The question that actually tests fit is whether the tool solves a specific, named problem for a specific person or team in a way that your current stack genuinely does not.

When you skip this question, you buy overlap. A sixteen-person agency once brought in a new AI writing tool because the demo was convincing and a competitor had mentioned it in a podcast. Three months later, their operations manager noticed they were paying for four tools that could all draft marketing copy. None of them was being used consistently, because nobody had decided which one was the actual answer. The new tool added cost and confusion without adding capability. Running this question before the purchase would have taken five minutes and saved the equivalent of a month's worth of licence fees.

2. Who owns it after the trial ends?

A tool with no owner is a tool that drifts. Someone has to be accountable for whether it is used, whether it is actually working, and whether it is still worth the cost in six months. If you cannot name that person before you sign up, the purchase is already half-dead.

Ownership does not have to mean a dedicated administrator. It means one person whose job it is to notice if the tool stops being used, to decide whether to renew when the reminder arrives, and to field questions from the rest of the team. Without that, tools persist on the subscription list long after they have stopped earning their place, because nobody is specifically responsible for making a call.

A twelve-person consultancy bought a new AI proposal tool after seeing it at a conference. Nobody asked who owned it. The trial ran for six months, quietly converting to a paid subscription when the trial period ended. It was not until the firm's accountant flagged the recurring charge during a quarterly review that anyone noticed nobody on the team had used it in four months. Naming an owner at the point of purchase would have prevented a write-off that embarrassed two senior partners.

3. What does it touch?

Where does your data go, and what would it mean if that vendor had a breach or changed their terms of service?

You do not need a legal review for every tool. You do need to know whether you are feeding client-confidential material, personal data, or commercially sensitive information into a system you have not checked. Many AI tools process your inputs on their servers to improve their models. Some store data you upload. Some share it with third parties under terms that are buried in their privacy policy. None of this is automatically disqualifying, but it is information you need before you start using the tool, not after.

The specific scenario to think through: which of your team's actual workflows would this tool touch, and what category of data would flow through it? A tool used only to rephrase internal announcements carries very different risk to one integrated into a client-facing proposal process. When a team at a professional services firm started using an AI tool to draft client documents, they assumed it worked like a word processor. Six months in, a review of the vendor's terms revealed that uploaded documents were retained and used for model training. The firm had been feeding client-confidential material into a system with no data processing agreement. The remediation cost more in time and legal fees than the tool had saved.

4. How will we know if it worked?

Decide the signal before you buy, not after.

Pick something concrete: adoption rate across the team, time saved on a named task, reduction in a specific type of error, volume of output produced per week. It does not need to be elaborate. It does need to be decided in advance, because a tool you cannot evaluate is a tool you will renew out of inertia. And inertia is expensive.

The trap is setting vague success criteria like "improves productivity" or "helps with writing." These cannot be measured, which means the renewal conversation will default to whether people feel like the tool is useful, which is a much easier bar to clear than whether it is actually earning its licence cost. A small property management company introduced an AI tool for responding to tenant enquiries. They never set a target. At renewal, the operations lead thought it was probably saving time. Nobody could say how much. They renewed. Two renewals later, the company hired a VA to handle the same work, at lower total cost, and cancelled the tool. A simple baseline established at the start would have surfaced that outcome months earlier.

5. What does it replace?

New tools should retire old ones, or at least old habits. If nothing comes out when something goes in, you are not adopting AI. You are accumulating it.

Every addition to a stack has a hidden cost: the time required to learn it, the overhead of managing another subscription, the mental weight of knowing it exists and deciding when to use it. That cost only makes sense if the tool is replacing something with a higher cost. Explicitly asking what this tool replaces forces the conversation that most teams avoid: not just what we are adding, but what we are stopping. If the answer is "nothing," that is worth sitting with before the card goes down.

The accumulation pattern looks like this: a marketing team adopts an AI image tool, keeps their existing stock photo subscription, and does not retire their old design tool because some people still prefer it. Eighteen months later they are paying for three overlapping visual content solutions and using each one inconsistently. The stack only stays manageable if buying something is also a decision to stop doing something else.

The goal isn't fewer tools for its own sake. It's a stack where every item earns its place and someone can say why.

Running the five questions as a team

These five questions work best when they are asked together, out loud, before the purchase is made. Not in a committee. Not as a procurement process. In a single thirty-minute conversation with the people who would actually use the tool and the person who holds the budget.

The format is simple: open a shared document, write the five questions at the top, and spend six minutes on each one. Whoever proposed the tool answers first. Everyone else pushes back or adds nuance. By the end, you will either have five solid answers that justify moving forward, or you will have found the one question nobody can answer, which is usually the one that would have caused the problem.

Writing the answers down matters. It creates a record for the renewal conversation. It means the person who owns the tool knows what they committed to. And it means that if the tool gets cancelled in three months, you have a clear reason that the team agreed on, not a vague sense that it did not quite work out.

The thirty minutes will feel unnecessary for a tool that seems obvious. It is most valuable precisely in those cases.

What to do next

Pick one AI tool currently in your organisation's stack that nobody consciously evaluated. It might be the one that arrived via a free trial that converted. It might be the one a team member bought on a card and expensed. It might be the one that has been on the subscription list so long that nobody remembers the original reason.

Run it through the five questions this week. Write the answers in a shared document. Then make a decision by end of Friday: it stays because it earns its place, or it goes because it does not.

The same discipline applies when the purchase is help rather than software. Before hiring anyone, work out whether the gap calls for AI training, coaching or consultancy, and name what your firm is short of first.

One tool, five questions, one decision. That is the whole exercise. If the answer is to cancel, cancel and note what you are saving. If the answer is to keep it, note what success looks like so the next renewal conversation has something to work with.

The stack you have right now was built one unexamined decision at a time. The way to improve it is one examined decision at a time, starting with the one you run through this week.