AI Readiness Assessment: How to Run One Properly in a Fortnight

The pattern is remarkably consistent. An organisation picks a promising AI use case, runs a pilot, and the pilot works. Then it tries to move the pilot into production and discovers that the data it relied on was hand-cleaned for the demo, that nobody owns the process it was supposed to improve, and that legal has questions nobody can answer.

Six months gone, and the thing that stopped it was never the model.

A readiness assessment is the cheap way to find that out first. Done properly it takes about a fortnight, it costs a fraction of a failed pilot, and its most valuable output is often a clear no.

What readiness actually means

Readiness is not a single score, and it is not about how enthusiastic the leadership team is. It is about whether six specific things are in place well enough to support a change, and the honest answer is usually that four of them are and two of them are not.

Illustrative, but the shape is familiar: strong appetite, weak foundations. The lowest bar sets the realistic pace, not the average.

• IntentStrong
• TechnologyGood
• DataPatchy
• ProcessWeak
• SkillsMixed
• GovernanceAbsent

Intent. Is there a specific commercial outcome, owned by a named person, with a number attached? “We should be doing something with AI” is not intent. “We want to cut quote turnaround from four days to one, and Sarah owns it” is.

Data. Does the data the use case needs exist, is it accessible without a three-week request, and is it good enough? This is where most assessments find the real blocker.

Process. Is the process you are about to automate actually defined? A great many are not written down anywhere, vary by person, and have exceptions that live in one long-serving employee’s head.

Skills. Can the people who will use this thing evaluate its output? Not build models. Judge whether an answer is right.

Technology. Can systems be reached, integrated and monitored? Is there somewhere for this to run that IT will support?

Governance. Who approves a use case, who is accountable for an output, and what happens when it is wrong? For most mid-market organisations this scores lowest, and it is the one that stops things at the last moment.

The point of scoring all six is that the lowest one governs. An organisation with excellent data and no governance will get exactly as far as an organisation with excellent governance and no data.

The fortnight

Two weeks is enough if you are disciplined about scope. It is not enough to assess everything, which is fine, because assessing everything produces a document nobody acts on.

1. 01Days 1 to 3Interviews. Eight to twelve people across functions and levels. Ask what takes too long, not what they want AI for.
2. 02Days 4 to 6Evidence. Look at the actual data, the actual systems and the actual documents. Nothing here is taken on trust.
3. 03Days 7 to 9Score and shortlist. Six dimensions scored with evidence. Candidate use cases ranked on value against effort.
4. 04Days 10 to 12Test the top one. A day of hands-on work against real data to see whether the approach survives contact.
5. 05Days 13 to 14Report and decide. Findings, a sequenced plan, and a recommendation that may well be to stop.

Days 1 to 3: interviews

Eight to twelve people, thirty to forty-five minutes each, across functions and levels. Include at least two people who will actually use whatever gets built, and at least one sceptic.

The questions that produce useful answers are not about AI.

• Walk me through the last time this process took longer than it should have. What happened?
• Where do you wait on someone else?
• What do you copy from one system into another?
• What gets checked twice because nobody quite trusts it?
• What have you already tried, and why did it stop?

That last one matters more than it looks. Most organisations have already attempted something, and the reason it stalled is usually the same reason the next one will.

Days 4 to 6: evidence

This is what separates an assessment from a workshop. Nothing is accepted on assertion.

Ask to see the data, not a description of it. Open the system. Count how many records have the field you need populated. Time how long an access request takes by actually making one. Read the client contracts to see what they say about third-party processing. Find out what tools people are already using without approval, and be genuinely unbothered when you find them, because otherwise nobody will tell you.

Expect to find at least one thing that changes the picture entirely. There is nearly always one.

Days 7 to 9: score and shortlist

Score each of the six dimensions on a simple one to five scale, and write the evidence next to each score. A score without evidence is an opinion, and opinions do not survive a board meeting.

Then rank candidate use cases on two axes only: value if it works, and effort to find out. The best first project is rarely the most valuable one. It is the one that proves the organisation can actually take a change from idea to production, which builds the permission for the valuable one.

Days 10 to 12: test the top candidate

A single day of hands-on work against the real data. Not a polished prototype, and not something anyone should see outside the room. The question is narrow: does the approach survive contact with the actual inputs?

This is the step people skip, and it is the step that most often changes the recommendation.

Days 13 to 14: report and decide

The report is short. Findings with evidence, the six scores, a ranked shortlist, a sequenced plan for the next quarter, and an explicit recommendation.

What the output should look like

Scored profileSix dimensions, each with the evidence that produced the score. Re-runnable in six months to show movement.

Ranked shortlistThree to five use cases with value, effort and the specific blocker for each. Not a list of everything anyone mentioned.

Blocker registerWhat has to be true before each thing can work, who owns it, and by when. The most useful page in the document.

Sequenced planA quarter of work in order, with a decision point at the end of each stage rather than a single big commitment.

Measurement definitionWhat number moves if this works, where that number lives today, and what it reads right now. Agreed before anyone builds.

That last one is the one organisations most often leave out, and its absence is why so many AI projects end in a disagreement about whether they worked.

Five patterns that show up almost every time

The data is worse than anyone believes. Not corrupt, just incomplete in the specific field the use case depends on. Everyone knows their own part is fine. Nobody has looked end to end.

The process is not written down. Three people do it three ways and all three are defensible. You cannot automate a process you cannot describe, and the describing is often where most of the value turns out to be.

Enthusiasm is inversely proportional to proximity. The board is certain. The people who will use it are wary, usually for specific and well-founded reasons that nobody has asked them about.

Somebody has already started. There is a team using an unapproved tool and getting real results. Find them. They are your best evidence and your best advocates, and punishing them is the fastest way to lose both.

Governance is nobody’s job. Everyone assumes someone else approves AI use cases. Nobody does. This surfaces late, usually the week before launch.

Doing it yourself, and when not to

You can run this internally, and for a first pass you probably should. What you need is someone with enough seniority to get access, enough independence to write down an unwelcome answer, and about ten days of genuine capacity rather than ten days squeezed around a day job.

Where an outside view earns its cost is in three specific places. People tell an external assessor things they will not put in an internal document. An assessor who has seen twenty of these recognises a pattern in a morning that takes an insider a fortnight. And a recommendation to stop carries considerably more weight when the person making it is not on the payroll of the department that would run the project.

If you would rather not have that conversation internally, our AI readiness audit runs this process on a fixed price, and the pricing page sets out what each format costs before you speak to anyone. If your team simply needs to be able to judge AI output well, SEO and AI training is often the cheaper first move.

The part worth remembering

The most valuable outcome of a readiness assessment is frequently a decision not to proceed yet, with a clear list of what has to change first. That is not a failed assessment. That is a fortnight and a small fee spent instead of two quarters and a large one.

Readiness is not a gate you pass once. It is a set of gauges, and the honest reading of the lowest one tells you how fast you can actually go.