What should an AI audit actually give you?

A serious AI audit should leave you with four things you can act on without the vendor: a process map of how work actually moves through your company, a gap analysis of where automation fits and where it does not, an ROI estimate with its assumptions written down, and a build plan ranked by return. If it ends in a slide deck and a proposal for the next engagement, it was a sales document, not an audit. A useful test: a real audit tells you where AI is not worth it.

The failure mode: a slide deck that sells the next engagement

Most AI assessments are written by the people who profit from what comes next. That incentive shapes the deliverable: a maturity score, a market overview, a roadmap where every step happens to require the assessor. Nothing in it is usable on its own.

The tell is simple. Ask what you could do with the deliverable if the vendor disappeared the day after handover. If the honest answer is "schedule another meeting", you bought a sales document with an invoice attached.

What are the four things a real audit leaves behind?

An audit is a purchase of information. The information should arrive as documents that keep working after the auditor leaves. Each has a usability bar it must clear:

1. The process map. How work actually moves through your company: the systems, the handoffs, the volumes, the places where a person reads one screen and retypes into another. It must be specific enough that someone who has never visited your office could follow a request end to end.

2. The gap analysis. For each mapped process: whether automation fits, what kind, and why — or why not. It names your constraints: data that cannot leave your environment, systems without APIs, rules too unstable to encode. It must include the processes where AI is not worth it.

3. The ROI estimate. Hours and euros per process, with every assumption written down — volumes, hourly costs, error rates. You should be able to rerun the maths yourself when volumes change. In back-office automation, payback is typically 6–12 months; the estimate must show how that number was reached, not just state it.

4. The build plan. The build order, ranked by return, with scope, dependencies, and effort per item. The test is blunt: hand it to a different vendor, or to your own team. If they can build from it, the audit did its job.

If an audit never says no, it is a sales document

Some processes should not be automated. The volume is too low to repay the build. The rules change monthly. The cost of one wrong output exceeds a year of savings. Or the fix is a process change that needs no AI at all.

A real audit lists these processes, with reasons. That is not a hedge — it is the point. An assessment in which every process is a strong candidate for AI is not analysis; it is a price list. Expect a written "no" somewhere in the deliverables, or question the rest.

The same discipline applies to targets. In customer support, a realistic automated resolution target is 50–70%. If an assessment promises more without production evidence, the rest of its numbers deserve a second look.

Who should run an AI audit?

People who then build. An audit written by account managers estimates what sells. One written by engineers estimates what ships — and the difference shows up in the ROI estimate. Numbers stay honest when the person who wrote them has to deliver against them.

When we do this work it is founder-delivered. The two of us map the processes, write the four documents, and then build the automations on the same stack we recommend — N8N, LLMs, PostgreSQL, running in a cloud account you own. There is no handoff to a delivery team that never saw your operation.

Where to start, before you pay anyone anything

You do not need to buy an assessment to find out whether you are anywhere near ready. Our free AI Audit is a five-minute conversation with a chatbot that asks the questions we would ask in the first meeting: what your processes look like, what software you run, whether your data is trustworthy, who would own the change. It returns an AI-readiness score, how well you fit the kind of automation we build, and two or three candidates worth looking at first — and if the honest answer is "not yet", it says what to fix before automating anything.

It is not a process map and it does not pretend to be one. It is the cheapest possible way to find out whether the full exercise is worth anyone's time.

You come in with one problem. The map usually shows ten.

Companies call about the problem they can see — the inbox that eats Monday morning, the invoices that pile up at month-end. The process map almost always shows more. The pattern we see: you think you have one problem, the map shows ten candidates, and roughly seven are worth solving. The other three get a documented reason why not.

Two examples from our own work:

• An industrial components distributor asked about its support inbox. The map surfaced several document flows with the same manual retyping in the middle. The ticket flow ranked first and now auto-creates 85% of tickets from inbound email, running 24/7 — about 58 hours a month given back to the team.
• In a supply chain company, an invoice question became a system that has processed 10,000+ invoices at 100% automation — and found real fraud nobody was looking for.

Want to know where you stand?

Start with the free AI Audit: five minutes, no sales pitch, a straight answer about how AI-ready you are. If it says you are ready, book a call and we will talk about what to build first.