An AI readiness assessment decides what to build before anyone builds it: which processes an AI system should run, which a plain script should run, and what has to be fixed first. We do it with 10 questions in four groups: the work, the data, the people and the proof.
Every question has an answer you can check, and a "not yet" becomes a task for the first weeks. Skipping the data questions is expensive: Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data.
The short version
- An AI readiness assessment decides, before any build, which processes AI should run, which a plain script should run, and what to fix first.
- Our 10 questions cover the work (hours, rules, what exists), the data (reach, accuracy), the people (owner, approver, where the output lands) and the proof (the number, the alarm).
- The answer that moves the budget most is script or model: steps that follow fixed rules belong in a script, judgment and language belong in a model.
- For one B2B data client, free rule checks settled ~75% of a data-quality step before any AI ran, and one list's error rate fell from 9.5% to 7.2%.
What an AI readiness assessment is for
An assessment comes before the build and an audit comes after: one decides what to build, the other checks what already runs.
The audit side is in what an AI operations audit finds. An assessment starts from the other end, with the list of ideas a company calls its AI opportunities, and asks of each one whether the company can support it today. It hands back that list ranked, with each idea marked script or model and the work it needs first.
You need one when the list of ideas is longer than the budget, or when a first AI project stalled and nobody can say why. Published prices for this kind of review run from $1,500 to $7,500 at the small end to $15,000 to $50,000+ for a regulated enterprise, and our cost post splits that range by scope.
The 10 questions we ask before building anything
Ten questions cover four areas, and the table shows what a ready answer looks like for each.
| Question | Not ready yet | Ready |
|---|---|---|
| 1. Where do the hours go? | A sense that the team is busy | Hours per week per process, with the people on it |
| 2. Which steps follow fixed rules? | Every step is an AI opportunity | Each step marked script or model |
| 3. What already exists? | A new build for every idea | Tools you already pay for, old scripts and half-built features checked first |
| 4. Can a system reach the data? | Someone exports a spreadsheet every week | Read access to the tools where the data lives |
| 5. Is the data right? | Nobody has checked | A sample checked by hand, with an error rate |
| 6. Who owns the process? | Everyone, so no one | One named person who can change it |
| 7. Who approves the output, and how fast? | Approvals wait on one busy person | A named approver and an agreed answer time |
| 8. Where will the output land? | A new dashboard to log into | The tool the team already opens every day |
| 9. What number says it worked? | Activity: runs, logins, messages sent | One outcome, measured before the build starts |
| 10. How will you know it stopped? | Someone notices, eventually | A check on the outcome that alerts a person |
The work. Questions 1 to 3 decide whether an idea is worth building and whether it is already built. In a replay of 4 past client tasks on 7 October, our own agents built something new twice where the code already had the answer, which is why we check what exists before designing anything.
The data. Questions 4 and 5 are where a sample checked by hand beats any interview. For question 5, pull a random sample of the rows the system would read, check each one against its source by hand, and write down the error rate. In Gartner's July 2024 survey of 1,203 data management leaders, 63% did not have, or were unsure they had, the right data practices for AI.
The people. Questions 6 to 8 fail without an error message. Between 3 and 7 October, agents in our own system asked 32 questions and 23 expired before anyone answered, and the missing input was one person's time: mine.
The proof. Questions 9 and 10 decide whether anyone will know that the build worked, or that it stopped. Our own small SaaS went 114 days without a working search before we noticed, because the alert we watched counted signups, and the checks we run now are in how to know if AI automation is working.
Script or model: the answer that moves the budget most
I think most items on a company's AI opportunities list are rules, and a rule belongs in a script that costs nothing per run.
My test: deterministic and repeatable means a script, judgment and language means a model. A script fails loudly when its input changes, while a model keeps going and fails confidently, which is harder to catch.
The split sets the price too. A script is written once and a model is paid per run, so an assessment that marks every step as AI quotes a bigger build and a bigger monthly bill than the work needs.
What this looked like for one client
For a B2B data client, the share of rows pointing at the wrong website tracked the error rate across 4 lead lists.
A wrong domain points the enrichment at the wrong company. The check we built caught a major bank's row pointing at a personal loan-modification blog, and a clearing firm's row pointing at a UK cricket club.
Question 2, which part of that check needed a model, decided the build: about a quarter of the rows did. The check runs cheapest first, with free tests for whether the domain exists or is parked, whether Wikipedia or Wikidata lists it as the official site, and whether the homepage shows the company's name.
Those free tests settle ~75% of rows, and an AI researcher handles the rest. The whole check costs ~$4 per 1,000 rows and takes ~15 minutes per 250-row list. On one list, fixing the domains took the error rate, the share of rows that failed the quality check, from 9.5% to 7.2%, and the rows pointing at the wrong company from 16 to 6.
Where a readiness assessment stops
An assessment reads the process as people describe it, and the real process shows itself only once something runs.
In the client example, fixing the domains pointed at the next error source: subsidiaries ran a 19.4% error rate, against 2.8% for independent companies. The next fix starts there.
An assessment also cannot tell you whether people will use the result. That shows up only in usage after launch, which is why question 8 asks where the output lands.
Because of both limits, I count the first small build as the last part of the assessment: it answers questions 9 and 10 with real usage.
Take your top three AI ideas and answer questions 2 and 5 for each: which steps follow rules, and is the data right?
If most steps follow rules, the project is mostly a script, with a smaller bill than an AI build. If nobody can say whether the data is right, checking a sample by hand is the first task.
When the list runs longer than one person can hold, the first weeks of an engagement are where we would run all ten.
Key takeaways
- Answer question 5 by checking a random sample of rows by hand and writing down the error rate.
- Question 3, what already exists, comes before any design: in a replay of 4 past client tasks, our own agents built something new twice where the code already had the answer.
- A script fails loudly when its input changes, while a model fails confidently, which is harder to catch.
- Approval paths fail quietly: between 3 and 7 October, 23 of 32 questions from our own agents expired unanswered.
- An assessment cannot show whether people will use the result, so the first small build finishes the assessment.
Common questions
What is an AI readiness assessment?
A review done before anything is built, which decides which processes an AI system should run, which a plain script should run, and what has to be fixed first. It checks the work, the data, the people and how success will be measured.
What questions should an AI readiness assessment ask?
Ours asks 10, in four groups. The work: where the hours go, which steps follow fixed rules, and what already exists. The data: whether a system can reach it and whether it is right. The people: who owns the process, who approves the output, and where the output lands. The proof: what number shows it worked, and how you will know when it stops.
Is a company with messy data ready for AI?
Not for AI that depends on that data. Check a sample by hand and fix what the check finds first: Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data.