This is the actual list I run through before we agree to build anything for a client. Ten questions, answered honestly. Tick them and you are ready. Miss one and you know your first job.
This AI readiness checklist is the real one, the questions I go through before Orwell Lab agrees to take on a build, because taking money to build something that was never going to work helps nobody. It is deliberately short and deliberately blunt. I have watched enough projects fail, mine and other people's, to know the failures cluster around a handful of questions people would rather not answer. Answer them honestly here and you will save yourself a lot of money. The stakes are real: 95% of generative AI pilots deliver zero measurable return, and 56% of chief executives report no measurable return on AI so far. This list is how you stay out of those numbers.
1. Can you name the problem in one sentence? Not "we want to use AI". A specific, repeated, painful task. If you cannot say it in a sentence, stop here and start here.
2. Is there a number attached to it? Hours, error rate, cost, missed leads. If you cannot measure the problem, you will not be able to prove the fix, and unprovable wins get cut in the next budget round.
3. Does the data exist and can you reach it? This is where most businesses fall down. 52% of businesses name data quality and availability as the biggest barrier to AI. If the data is missing, locked in someone's inbox, or a mess, that is your first project, not the AI.
4. Is the data clean enough to trust? Different from question three. Data can exist and still be too dirty to build on. Be honest about the state of it, because the model will inherit every flaw.
5. Has anyone mapped the process? You cannot automate a job nobody has written down, because you cannot tell the system what "right" looks like. If it lives only in one person's head, map it first.
6. Who owns the output once it is live? Software is not finished at launch, it is adopted. A common gap: the AI capability gap between small and large firms sits at around 38 percentage points in the EU and is widening, largely because smaller firms have nobody assigned to own the thing. You do not need a data team, you need one named person.
7. Do you know where a human stays in the loop? For every output the system produces, someone should be able to say what happens when it is wrong and who catches it. If the answer is "nobody", that is a liability waiting to surface.
8. Have you decided build versus buy? Not everything should be custom. In the MIT research, externally sourced AI tools succeed about 67% of the time, more than double the rate of internally built ones. Default to buying unless the fit or the ownership genuinely justifies a build.
9. Can you start small? Is there a version of this you could ship in weeks, on one process, and learn from? If the only version is enormous, shrink it before you spend.
10. What happens if you do nothing? The honest control question. If the answer is "not much", this might not be the problem worth solving first. That is a finding, not a failure.
There is no pass mark, but there is a hierarchy. Questions one to five are the foundation: a real problem, a number, reachable data, clean data, a mapped process. Miss any of those and I would not build yet, however strong the rest looks. Six and seven are what keep a working build alive after launch. Eight to ten are what stop you overspending. If you tick the first five and have a plan for the rest, you are ready. If you do not, the gaps are your project plan, and they are almost always cheaper to close than people fear.
This checklist is the fast, self-serve version of the readiness assessment. If you want the opportunities ranked rather than a single use case checked, that is the AI audit. And if you tick everything and the honest next step is a custom build, the bespoke software buyer's guide covers how to buy one without the overrun. The thread through all of it is the same: be wrong cheaply and early, on paper, before it costs you a pilot.
A real problem with a number on it, data you can reach and trust, a mapped process, an owner for the output, a governance line for when it is wrong, and a build-versus-buy view. If you cannot tick those, you have your starting work list.
There is no pass mark, but the first three, a real problem, usable data and a mapped process, are non-negotiable. Miss any of those and a build will struggle no matter how strong the rest looks.
Data. It is the single biggest barrier for most businesses. The second is naming a specific problem rather than a general wish to use AI.
The checklist is enough to decide whether to proceed or pause on a single use case. A full audit is worth it when several opportunities compete for budget and you need them ranked with cost and payback.
Run the ten questions on your own toughest use case this week. If you get stuck on one, that stuck point is usually the whole answer. When you want a second pair of eyes on the result, send it to us at contact.
Send me the use case you are least sure about and I will run these ten questions against it with you, and tell you straight whether to build, fix something first, or leave it alone.
Book a 30-minute fit call