Before you start: assess one use case, not the company
Most readiness assessments fail because they try to score the whole organization on a maturity grid, produce a color-coded slide, and change nothing. Do the opposite. Run the assessment against one candidate use case you might actually put into production in the next quarter. A readiness score only means something when it is attached to a specific thing you are deciding whether to ship. This is the same discipline behind our Free AI Value & Readiness Assessment: pick the work, then score it.
Step 1: Name the use case and its owner
Write one sentence: this use case uses AI to do X so that Y improves, and Z owns it. For example, this use case drafts first-response replies to billing questions so that support handling time drops, and the head of support owns it. If you cannot name a single accountable owner, stop here. Unowned use cases do not ship; they become everyone's idea and no one's job. Watch for a common tell: if the sentence needs the word and twice to list what the AI does, you have three use cases pretending to be one, and you should split them and assess the most valuable one alone. The outcome of this step is a one-line definition and a named person, not a workshop.
Step 2: Write the value case as two numbers
Find the baseline: what does this task cost or produce today, measured. Then set the target: what should it be after AI does it. Cost per resolved ticket, hours per report, error rate, revenue per rep, pick one. Pick a metric someone already reports, so you are not inventing a measurement system alongside the AI. If the honest baseline is we have never measured this, that is your first finding, and the readiness answer for this use case is measure-it-first. A target you set with no baseline is a wish, and wishes do not survive a budget review. If you cannot state a baseline and a target number, the use case is not ready, because you will have no way to tell whether it worked. The outcome of this step is a before number and an after number you would be willing to be held to in front of a CFO.
Step 3: Score data readiness one to five
Rate three things from one to five: do you have the data this use case needs, do you trust it, and can the AI reach it without breaking a privacy or residency rule. Be specific about the third one. If the data holding your answer sits in a system that cannot leave a given region, or contains personal information that a model provider would then see, that is a control question you resolve before building, not after a customer complaint. Data is the bottleneck far more often than the model, and it is where honest teams find their lowest score. A three or below on trust means the assessment result is fix-the-data-first, not build-the-AI, and that is a useful answer, because it stops you spending on a model that would have been fed numbers no one believes. The outcome is three scores and a short note on the lowest one.
Step 4: Map the controls the use case needs
List what has to be true for this use case to run safely: which actions need a human to approve, what must be logged to answer an auditor, and which regulations apply. For regulated workloads the EU AI Act and ISO 42001 tell you what evidence you will need to keep, so decide now rather than after go-live. The outcome is a short control list that a delivery team can build against, described as what you will do, not a product you will buy.
Step 5: Estimate the path to production
Sketch the real work from here to a live system: integration with the systems it reads and writes, the human-in-the-loop steps, monitoring, and who runs it after launch. Be honest about the parts that are not the model. The prompt is usually a week of work; the integration, the approval flow, and the monitoring are the months. Most teams underestimate the after-launch job entirely, then discover no one owns the agent once the project team disbands. Name that owner now. The outcome is a rough effort estimate and the two or three things most likely to stall the project, written down where the sponsor can see them.
Step 6: Decide go, no-go, or fix-first
Put the four scores together: value, data, control, effort. High value with weak data is a fix-first. Strong data with no measurable value is a no-go, no matter how good the demo looked. Clear value, trustworthy data, and controls you can build is a go, and you can be shipping a first governed workflow in 30 days. The outcome of the whole assessment is one decision per use case and a short reason, so a leadership team can act on it in a single meeting instead of commissioning another study. Run this on your top three candidates and you will usually find one clear go, one fix-first, and one you are relieved to have killed before it cost anything. That is a good day of work, and it is the difference between an AI program that ships and one that keeps studying itself.
Frequently asked questions
How long does an AI readiness assessment take?
For one use case, a focused team can complete these six steps in a few days. Scoring twenty use cases at once is what turns it into a quarter-long exercise that decides nothing.
What is the difference between AI readiness and AI maturity?
Maturity models grade the organization in the abstract. Readiness asks a sharper question: is this specific use case ready to ship, and if not, what is the one thing blocking it.