You have received an AI audit proposal running to dozens of pages. It talks about maturity, transformation and governance. It does not say what you will hold in your hands six weeks from now, or how you will check that the work was done. Yet the audit is the step that keeps you from launching a project into thin air. You still need to know what it should produce, and what it should not cost.
Key Takeaways
- An AI audit is not a status report. It is a documented investment decision, with use cases kept and use cases explicitly dropped.
- Five deliverables carry the weight: use case map, data inventory, risk analysis, costing, and a roadmap with stop conditions.
- An audit is billed on time spent. An invoice indexed on the future project budget is a conflict of interest, not a price.
- A free audit, or one bundled into an implementation offer, is there to sell the solution, not to inform the decision.
- For a mid-sized company, expect a few weeks and a four or five figure budget, to be weighed against the cost of a project launched on the wrong use case.
An AI audit is meant to decide, not to describe
A status report answers "where do we stand?". An audit answers "what do we do, in what order, with what budget, and what will make us stop?". The difference is not semantic: it determines what you can do with the spend.
The starting point is rarely the right one. Companies that fail with AI do not lack tools. They picked a use case because it was visible, not because it was profitable. In a July 2024 note, Gartner estimated that at least 30% of generative AI projects would be abandoned after the proof of concept before the end of 2025, for four reasons: poor data quality, inadequate risk controls, escalating costs and unclear business value.
Keep the list, not the number. Those four causes can be checked before the first euro of implementation, not after. That is precisely the work of an audit.
The three decisions that must come out of it
An audit is judged on what it settles. Reading the document, you should be able to answer three questions without calling anyone.
- Which use cases we launch, which we drop, and why.
- What the company spends once, and what it spends every month.
- Who owns the system once it runs, and who can stop it.
If the document leaves those three answers open, it is not an audit. It is a study.
The five deliverables that carry the weight
A useful audit fits in thirty pages, not a hundred. Volume is not the point; the nature of the artefacts is. Here is what a serious provider hands over, and the concrete evidence that must come with it.
| Deliverable | Question it answers | Expected evidence |
|---|---|---|
| Use case map | Where AI creates value here, and where it does not | List ranked by estimated annual gain, effort and risk, with dropped cases and their reasons |
| Data and access inventory | Whether our data really supports this use case | Source list, observed quality, owner, usage rights |
| Risk analysis | What happens if the system is wrong or leaks | Risk register, personal data handling, place of human control |
| Costing | What each option costs, once and recurring | Ranges per scenario, licences, model usage, hosting, maintenance |
| Roadmap | What to start with, and under what conditions to stop | Milestone sequence, quantified exit criteria, what stays in-house |
The data inventory is the part most often skipped
No use case can be judged without looking at the data behind it. A serious audit lists the sources, their actual quality, their owner, usage rights and update frequency. Without that page, the project stalls in month three, at the moment of connecting the system to the real customer file, the one with duplicates, empty fields and three naming conventions.
That page also has a political virtue. It makes visible what the company cannot yet do, before the budget is committed.
The costing fits on one page
Four lines are enough: build cost, recurring cost (licences, model usage, hosting), internal cost (team time, training), exit cost. An audit that only costs the build phase sets up a disappointment. The recurring spend of an agent in production often exceeds, over two years, the build spend.
The costing should stay an accepted range, with the assumptions that produce it. A single figure accurate to the cent on a project that does not exist yet is an invented figure.
What an AI audit must not deliver
Some documents look like an audit and are not. The signs are easy to spot before signing.
- A trends report. The AI market does not need explaining to an executive who reads the press. Thirty pages on the state of the art are thirty pages you pay for with no decision inside.
- A maturity score without a scale. "3.2 out of 5 on data maturity" means nothing if the scale is not provided and nobody knows what a 4 is worth.
- A tool recommendation without a test. An audit that names a product without confronting it with your documents shifts the risk onto you.
- A promised pilot. The audit prepares the pilot; it does not deliver it. Blending the two makes you pay for implementation at diagnostic prices, and removes the moment when you could still say no.
- An invoice indexed on the future project. A percentage of the implementation budget rewards the expected conclusion, not the right one.
The last point deserves its own paragraph. An audit whose price depends on the project it will trigger is no longer a diagnosis. It is a sales quote with different formatting.
What an AI audit must not cost
An audit is billed on time spent, through a day rate and a number of days announced in advance. That is the only structure that makes the price verifiable.
To get a sense of scale, the 2026 study by Jobbers, which compiles the Malt, Hays and Free-Work barometers, puts the average day rate for IT freelancers in France at 520 EUR, all profiles combined. At that rate, ten days represent 5,200 EUR. An audit for a mid-sized company usually falls between five and fifteen days depending on how many departments are involved, which places the spend in a four or five figure range. AI specialists bill above that average, and that is normal: you are paying for a rare skill, not for slides.
As an internal comparison, this spend should stay below what a pilot launched on the wrong use case would cost: three months of team time, licences, and a result nobody dares to stop.
What inflates the invoice without improving the decision
Four items keep showing up in audit quotes, and none of them makes the diagnosis better.
- Widening the scope to the whole company instead of one department.
- Multiplying interviews up to forty, when eight well chosen ones are enough.
- Producing a benchmark of forty tools.
- Adding an awareness phase billed at consulting rates.
Conversely, what should lower the price: a single-department scope, one question per meeting, dated deliverables, and an advisor who already knows your industry.
Three warning signs on an audit quote
A free audit, or one offered because you sign the implementation, does not tell you whether the project is worth it. It tells you the provider already has the answer. A price expressed as a percentage of the future project has the same effect. An open-ended duration, with no milestone or delivery date, prevents any control.
In all three cases, the question to ask is the same: what happens if the audit concludes that nothing should be launched? If the answer is about the rest of the engagement, the answer is a bad one.
The questions to ask before signing
Eight questions are enough to sort proposals. They fit on one page and should be asked in writing, before signing.
- Which precise deliverables do we receive, and on what date?
- How many days are billed, and at what day rate?
- Who performs the audit, and will that person attend the interviews?
- Which decisions must the document settle?
- How will use cases be ranked, and on which internal figures?
- What happens if the conclusion is to launch nothing?
- Who owns the documents produced, and in what format?
- What must we provide, and how much of our time will it take?
The sixth question is the most discriminating. A provider who accepts a negative conclusion without renegotiating the contract is working for you.
FAQ
How long does an AI audit take for a mid-sized company?
Two to four weeks for a scope covering one or two departments, with grouped interviews and accessible documents. Beyond six weeks, value drops: the data collected ages and the company has already started deciding without the report.
Can an AI audit be run in-house?
Yes, when three conditions are met: someone has the real time to do it, that person has no stake in the outcome, and the company knows how to cost recurring licence and usage spend. An in-house audit rarely falls short on business knowledge, often on arbitration: it is hard to say no to your own director.
Is an audit needed before launching a pilot?
A light audit is enough when the pilot is small, bounded and reversible. It becomes essential as soon as the pilot touches customer data, a critical process or a commercial commitment. The pilot answers "does it work". The audit answers "is it worth it, and under what conditions".
How do you check at delivery that the audit was done well?
Three simple tests. Did the document make you give something up? Does it contain internal figures you can challenge? Do your teams recognise their work in the way it is described? An audit that drops no use case has arbitrated nothing. An audit your teams do not recognise was written from the outside.
What if the audit concludes that nothing should be launched?
That is a result, not a failure. An audit that concludes in favour of the status quo has spared you a budget and a project that is hard to stop. Best practice is to date the next review: data improves, costs fall, and a dropped use case can become profitable eighteen months later.
Conclusion
An AI audit is neither a market study nor a motivation document. It is an accounting artefact: it states what you commit, what you get, what you risk and what will make you stop. Five deliverables, a day rate, an announced duration, and a conclusion that may be negative. The rest is literature.
If you want to frame an audit before choosing a use case, our AI audit and advisory page describes the process and the deliverables. To discuss it with someone who knows your sector, get in touch.
Two useful reads before signing: what an AI agent quote should contain and comparing two commercial proposals on an AI project. If the audit leads to an agent in production, plan its maintenance and support from the start, and look closely at what task automation covers before handing over an entire process.
