If you have looked into AI for your business, you have probably been offered an "AI readiness assessment". In our experience these are mostly a slide deck and a maturity score, followed — surprisingly often — by a proposal to sell you something.

A small business does not need a maturity score. It needs to know whether one specific process is ready to hand over. That is a question you can answer yourself in an afternoon, and it is worth answering before you spend anything, because most of what determines whether AI works on a process has nothing to do with the AI.

This is the checklist we run before we let software — ours or anyone's — take over a piece of work. We run it on our own operations constantly, because every item on it was learned by skipping it and paying for the skip.

1. One process, not "the company"

Not "our operations". Not "marketing". One process, nameable in a single sentence: invoices get matched to deliveries, new enquiries get a first reply, the weekly report gets assembled.

If you cannot name it in a sentence, you are not looking at a process yet — you are looking at a department. Departments do not get automated; their pieces do. Pick the piece.

Green light: someone on your side can say the one sentence, and everyone else agrees that is the process.

2. Someone can define "correct"

Before you automate anything, somebody has to be able to write down what a good output looks like. Not describe it — write it down, specifically enough that a second person could check an output against it.

This is the check most often skipped, and the most expensive to skip. An AI system with no definition of correct will still produce output — confident, well-formatted, plausible output. You will then be paying for work nobody can grade, which is worse than the manual version, because at least the manual version came with someone's judgement attached.

If nothing else in this checklist lands, take this one: the hardest part of adopting AI is not the technology. It is that most businesses have never been forced to define "done". Writing that down is free, and it is the part you keep whatever you decide about the tools.

Green light: "correct" exists as a written definition a stranger could check against.

3. The inputs live somewhere reachable

The process runs on something — documents, spreadsheets, photos, messages, a system. For a tool to do the work, it has to be able to get at those inputs, and they have to be consistent enough that "the same thing" is recognisable from week to week.

This does not need to be perfect. Ours was nowhere near perfect when we started. It needs to be retrievable and predictably shaped. If every instance of the process arrives in a different format from a different channel, tidy that first — the tidying itself often pays back before any AI enters the picture.

Green light: you can point at where the inputs live, and a new hire could find them.

4. A wrong output is survivable — and survivable visibly

Everyone underestimates this one in the same direction. The question is not "how likely is an error" but "what happens when there is one".

A wrong first-draft reply an employee reviews before sending: survivable, cheap, useful even. A wrong payment instruction: not survivable, regardless of how clever the system is. Start with processes where an error surfaces before it costs, and keep the irreversible ones — money out, contracts, anything regulatory — behind a person until there is a track record.

We wrote about this at length from our own failures: the dangerous failure is the plausible one — the confident, well-formatted answer that nobody thinks to check. That post exists because we paid for the lesson more than once.

Green light: the worst credible error is caught by a checkpoint, not by your customer.

5. The baseline is measured — today, before anything changes

How long does the process take now? How often is it wrong now? Write the numbers down before the first tool is bought, not after.

Without a baseline, you will not be able to tell whether the AI helped, and the decision will be made by whoever is most enthusiastic. With one, the question answers itself within weeks. This is also the discipline that keeps vendors honest — a serious one will help you measure it; a vague one will change the subject.

Green light: two numbers on a page — time per instance, error rate — dated before day one.

6. A named owner answers for the result

"Operations will own it" means nobody owns it. One person, named, who answers for the output — including when the software produced it.

This is not bureaucracy. Accountability is the thing that cannot be delegated to software, and processes without an owner drift precisely when the automation makes them quieter. The quieter it runs, the more you need to know who is responsible for the day it doesn't.

Green light: if I asked you who answers for this process, you could say a name, and that person knows they are the answer.

7. A human checkpoint sits at the boundary

Wherever the output reaches the outside — a customer, a supplier, a government portal, a payment — a person looks at it. Every time, at first.

Not forever. The checkpoint is how the system earns trust: the day the human finds nothing to correct for weeks in a row, you can widen what passes through unreviewed. We still run this way on anything consequential, and the habit has caught more problems than any tool we have bought.

Green light: there is a named moment where a person checks output before it leaves — and it actually happens on busy days, which is the real test.

---

What readiness is not

None of the seven checks is about tools, budgets, or the maturity of your "AI strategy". That is the point. Readiness for AI looks a lot like readiness for any other form of delegation: a defined task, a definition of correct, a measurable baseline, and an owner who checks the work. Companies that have those get results from modest tools. Companies without them get nothing from impressive ones.

If you ran the checklist and your candidate process failed at check 2, that is not a bad result — it is the work list. "Write down what good looks like" costs a meeting, and it is the single highest-leverage thing on this page.

We run our own company on these checks — including the ones we learned by breaking them. The checklist above is the cheap version of that education.

If you are somewhere in this process and want to compare notes, the contact page is a real person, and the conversation costs nothing.

---

NEXT3LABS builds and runs high-stakes custom software from Singapore. We write up what actually happened on our own projects — including the parts that went wrong.

Working on something where being wrong is expensive? Message us on WhatsApp — no pitch, happy to compare notes.

More notes · Other ways to reach us