There is a pattern that repeats itself. An AI vendor delivers a flawless demo, talks about operational efficiency, cites success stories from companies no one can verify, and proposes an eight-week pilot. The CFO approves the budget. Six months later, the project is on hold, the team isn't using it, and no one quite knows what went wrong.
The problem is almost never the technology. It's that the right questions weren't asked before signing.
This article compiles twelve questions that an experienced CFO should raise in any AI vendor evaluation process. These are not technical questions. They are questions about business, risk, and control.
1. What is the expected return, and over what timeframe?
Not in vague terms. In euros, hours, or percentage points. If the vendor cannot articulate an ROI hypothesis with reasonable ranges for your industry and company size, the proposal isn't ready. An honest range — for example, "between 15 and 40 monthly hours recovered in the reporting function, depending on the volume of data sources" — is more useful than a promise of transformation without figures.
2. What happens if the project doesn't work?
An uncomfortable question, but a necessary one. Are there exit clauses? Does the vendor assume any risk? What defines success, and who measures it? A serious vendor has clear answers. One that avoids the question will likely avoid accountability as well.
3. How does it integrate with the systems we already have?
ERP, CRM, spreadsheets, internal databases. Integration is where most AI projects run into trouble. Ask which connectors exist, what requires custom development, and who maintains those integrations when something changes in the source system.
4. Who on the internal team needs to be involved, and how much time will it require?
If the answer is "minimal involvement on your part," be skeptical. Every AI project that works in production requires time from the team: to validate outputs, to correct workflows, to learn how to operate what has been built. The question isn't whether there will be an internal burden, but how much and when.
5. What happens when the agent makes a mistake?
AI systems make mistakes. What matters is how they are detected, who corrects them, and what impact they have on the business. Ask about supervision mechanisms, confidence thresholds, and human escalation workflows. If the vendor doesn't have a structured answer, the operational risk is yours.
6. How is system performance measured in production?
Dashboards, alerts, activity logs. An AI system without observability is a black box. You need to know whether the agent is functioning, how many requests it processes, how many it escalates to humans, and at what error rate. If that isn't included in the proposal, request it explicitly.
7. Does company data leave our infrastructure?
Privacy, regulatory compliance, data sovereignty. The GDPR imposes specific obligations. Ask where data is processed, whether it is used to train external models, and what contractual guarantees exist. This is not a technical question — it is a question of legal risk.
8. What happens if we want to change vendors in twelve months?
Technological dependency is a real risk. If the system is built on the vendor's proprietary tools without accessible documentation, the cost of exit can be prohibitive. Ask about the portability of workflows, code documentation, and the intellectual property ownership of what is built.
9. How many projects similar to ours have you taken to production?
Not pilots. Not demos. Projects in production, with real users, at companies of comparable size and sector. If the vendor has experience in industrial manufacturing or in companies with ERP systems and complex manual processes, that is relevant. If their references are limited to sectors unrelated to yours, the adaptation risk is higher.
10. How do you ensure team adoption?
Technology that no one uses generates no ROI. Ask what the project includes in terms of training, ongoing support, and change management. A well-built agent that the team avoids using — out of distrust or lack of training — is an expense, not an investment.
11. What does the long-term pricing model look like?
Many vendors offer attractive entry-level prices that scale in unpredictable ways with usage volume, number of agents, or additional integrations. Request a cost projection at twelve and twenty-four months based on the usage assumptions you expect. Compare it against the projected ROI.
12. What remains in our company when the contract ends?
Documentation, code, workflows, internal knowledge. If the answer is "nothing, because everything lives on our platform," the dependency is total. A client-oriented vendor leaves installed capacity behind: a team that knows how to operate and extend what was built, not a system that only functions while the subscription is being paid.
Why these questions matter now
The AI market for mid-size enterprises is growing rapidly, and with it the number of vendors promising results they cannot sustain. The difference between a solid proposal and an empty promise isn't always visible in the demo. It becomes visible in the answers to these twelve questions.
A vendor that responds with clarity, with data, and without evasion is in a position to take on a real project. One that generalizes, deflects, or promises without evidence probably isn't.
These questions don't guarantee the success of an AI project. They do reduce the risk of approving one that should never have been approved.
How to use them in your next evaluation process
Bring these questions to your next meeting with a vendor. Not as an interrogation, but as a filter for seriousness. The answers will tell you more about the vendor's maturity than any sales presentation.
If you are assessing whether AI makes sense for your company right now — which processes to address first, what return is reasonable to expect, what risks to consider — we can work through that analysis together in a fifteen-minute call.
It's not a sales call. It's a diagnostic: what makes sense to automate, what doesn't, and why.
[Request a free diagnostic →]
