An AI readiness assessment studies one area of the business before anyone pays for a build. It maps the work step by step, checks the data with real samples, and ends with starting numbers, risks and a fixed-price build estimate. You can run one in 2 weeks, and sometimes the answer is not to build.

It answers the questions a proposal usually skips. A finance director gets asked to approve an AI project, and the paper doesn’t say how many hours the work takes today, which steps AI would take on, what the build costs, or what happens when it gets something wrong.

What an AI readiness assessment answers

The UK government’s 2026 survey found that 65% of businesses not using AI cited the lack of a clear business case or proven return as a barrier. A readiness assessment is how you build that business case for one area, on your own numbers.

It should answer 4 questions:

  1. Where does the time go? Each step of the process, who does it, in which system, and how long it takes.
  2. What can AI take on? Step by step: what AI can do, what it can prepare for a person, and what stays manual.
  3. What is it worth, and what will it cost? A saving range from the starting numbers, a build estimate and the running costs.
  4. What could go wrong? Data problems, regulatory limits, and the mistakes the system could make and how they’d be caught.

Maturity assessment or process assessment

The term covers 2 different products, and they’re useful for different decisions.

Maturity assessment Process assessment
Scope The whole organisation One area or process
Method Questionnaires and interviews Walkthroughs, system data and samples
Output A score per dimension and a roadmap Starting numbers, a step-by-step plan and a build estimate
Answers How ready are we in general? What should we build here, and is it worth it?
Useful for Board conversations, policy, training plans Approving or rejecting a build

If the decision in front of you is whether to build something, you need the process assessment. A maturity score won’t make that decision.

Choosing the scope

Pick one area with steady volume and an owner who can describe how the work should go. Good scopes are named by their output: supplier invoices, client onboarding, month-end close, service requests, weekly reporting.

Agree these before the first walkthrough:

  • The goal. What the owner wants to change, in their words, such as onboarding a client in 2 days instead of 9.
  • The boundary. Where the process starts and ends, and which teams and systems are in it.
  • The people. Who does the work, who approves it and who can pull data from the systems.
  • The decision. What the report will be used for, and who makes the call.

Keep the scope narrow, so every process in it gets watched and timed.

Who takes part, and how long it takes

Ours takes about 4 hours of your team’s time in total.

Role What they do Time
Process owner Agrees scope, goal and starting numbers, attends the report About 1.5 hours
2 or 3 people who do the work Show the work in a walkthrough About 30 minutes each
Someone with system access Pulls volumes, exports and samples About 1 hour
The assessment team Observes, measures, tests samples, writes the report 2 weeks
Day 1: kickoff, scopeand goalWeek 1: walkthroughsand volumesWeek 1 to 2: datasamples and tests onpast casesWeek 2: numbers,options and estimateReport and decision
2 weeks of an AI readiness assessment

Mapping the process

Watch the work being done. Asking people to describe their process produces the version in the procedure manual. Watching produces the version with the spreadsheet on the second screen and the supplier who always sends PDFs as photos.

For each step, record:

  • what triggers it and what it produces
  • who does it, and in which system
  • how long it takes, timed on real items
  • how often it goes wrong, and what happens then
  • whether it follows a written rule or needs judgement

Then tag each step. AI can do it, with a person checking. AI can prepare it for a person to decide. Or it stays manual. The tags come from what you saw and tested, and the volumes come from the systems.

This map comes from a client onboarding process at a professional services firm. Screening review stays with a person, because the decision carries regulatory weight. The check against the document checklist only halves, because AI prepares the comparison and a person still decides.

Checking the data

In the builds we’ve run, most surprises have come from the data, and they’re cheaper to find in an assessment than in month 3 of a build. Check with samples pulled from the systems. Documentation describes the data as someone meant it to be.

Check How What can change the plan
Where it lives List every system and spreadsheet each step uses A step depends on someone’s personal file
How to reach it Interface, scheduled export or screen only No interface means an export routine or a replacement
How complete it is Pull 50 to 100 real records and count empty or odd fields Key fields missing for a share of records
How consistent it is Compare the same item across 2 systems 2 systems disagree, and nobody knows which is right
How sensitive it is Personal data, financial data, special categories Masking, residency or a private model
How it changes New suppliers, new layouts, seasonal volume Peaks the build has to handle

In one engagement, a data check found the email field empty for every customer record in the data warehouse. Every rule that depended on that field had been treating every account as normal. No one had checked, because the reports built on it ran without errors.

Test a sample of past cases too. Take 30 to 50 real items with known outcomes and run the reading or sorting step on them with a general-purpose model. It won’t be the final build. But the share it gets right, and the kind of mistakes it makes, tell you whether the plan is realistic.

The numbers

The report needs 4 starting numbers for the area, signed off by the owner: volume per month, time per item, the rework rate, and the elapsed time from start to finish. They become the baseline every later result is measured against. Our guide to measuring AI ROI covers how to time them.

From the step map, estimate the time after the build, step by step. Then give the saving as a range. The high case uses the estimates as they are. The low case halves the saving, because assessments are made on samples and the first months of a build rarely match the plan.

A worked example, using the onboarding map:

Line Calculation Figure
Volume From the CRM 120 clients a month
Time today Sum of the mapped steps 80 minutes per client
Hours today 120 at 80 minutes 160 hours a month
Estimated time after Sum of step estimates 32 minutes per client
Hours saved, high case 120 at 48 minutes 96 hours a month
Hours saved, low case Half of the high case 48 hours a month
Value at a loaded $45 an hour High and low $4,320 and $2,160 a month
Running cost Hosting, model usage, maintenance $900 a month
Build estimate Fixed price $28,000
Payback Build divided by net monthly value 8.2 to 22 months

That range is the answer to take to the budget holder. If the owner can redeploy 96 hours a month, the build pays back in under a year. If only half the saving appears, it takes nearly 2 years, and the decision depends on what else the money could do.

Risk and regulation

The report should name what could go wrong and how it would be caught. For each step AI takes on, write down the mistake it could make, the check that catches it and who sees the exception.

Some uses carry legal weight before any build. The EU AI Act lists high-risk uses in Annex III, including AI used in recruitment and selection, in allocating work and monitoring staff, and in assessing a person’s creditworthiness. An assessment that touches those areas must say so and list the duties that follow before any build starts.

The other risks are practical:

  • Accuracy. In the same UK survey, 56% of businesses using AI had concerns about bias and accuracy. The sample test gives you a first measure of yours.
  • Skills. The survey found limited AI skills affected 67% of adopters. Say who will own the system after launch.
  • Data protection. Where personal data goes, how long it’s kept, and whether it may leave your region.
  • Dependency. What happens if a vendor changes its pricing or a model is withdrawn.

The report

The report should be short enough to read in one sitting and specific enough that another team could build from it.

  1. The process, mapped. Every step, who does it, which systems it touches and where the time goes.
  2. The goal and the starting numbers. Volume, time per item, errors and rework, and the target for each.
  3. What AI can take on. Step by step, with the sample test results.
  4. The options, ranked. Usually 2 or 3, with hours, cost and payback for each.
  5. A fixed-price build estimate for the recommended option: scope, price, timeline and the first change that would go live.
  6. Risks and data. What the build needs, where the data stays, and what could go wrong.
  7. A straight answer. Build, build later, or don’t build.
A page from an AI readiness assessment report with starting numbers and ranked optionsA report page for client onboarding. Starting numbers: 120 clients a month, 80 minutes per client, 9 days until an account is active. 3 options ranked by hours saved a month and payback. Client onboarding, 48 to 96 hours, 8 to 22 months, build first. Billing report, 15 to 30 hours, 11 to 32 months, later. Renewal reminders, 6 hours, over 3 years, don't build. Illustrative figures.Readiness report, client onboardingStarting numbers, signed off by the owner120clients a month80 minper client9 daysuntil activeOptionHours a monthPaybackCallClient onboarding48 to 96 h8 to 22 monthsBuild firstBilling report15 to 30 h11 to 32 monthsLaterRenewal reminders6 hOver 3 yearsDon't buildLow case halves the estimated saving
3 options ranked, and 1 marked do not build

Each option shows a low and a high case, so the payback range is visible before anyone approves a budget. The third option saves too little to cover its build and running costs, and the report says not to build it.

Red flags

An assessment that shows any of these won’t help you make a decision.

  • No volumes. Recommendations without counts from your systems are opinions.
  • A score with no process. “Data readiness 3 of 5” tells you nothing about what to build.
  • The tool chosen first. If the answer is a product the assessor sells, check whether the process was studied at all.
  • Percentages of work AI can do, untested. “AI can automate 70%” needs a sample of your past cases behind it.
  • No low case. A single saving figure with no range hasn’t accounted for anything going wrong.
  • No owner named. Every recommendation needs a person who will be responsible for the result.
  • No option to stop. If every assessment ends with a build, the assessment is a sales step.

Running one yourself

An internal team can run the same assessment, in this order.

  1. Pick one area and agree the goal, boundary, people and decision with its owner.
  2. Watch 2 or 3 people do the work, and time 30 to 50 real items.
  3. Pull volumes and samples from the systems, and run the data checks on them.
  4. Tag every step, and test the AI steps on a sample of past cases.
  5. Write the starting numbers, the saving range and the risks.
  6. Get a build estimate, from your own team or from outside, for the recommended option.
  7. Present the report to the owner and the person who approves the budget.

The hard parts are time and calibration. Someone has to spend most of 2 weeks on it, and the build estimate and the low case are only as good as the team’s experience of similar builds.

Where to start

Our AI readiness assessment covers one area in 2 weeks and needs about 4 hours of your team’s time. It ends with the starting numbers, what AI can take on and a fixed-price build estimate. The price is credited in full against the build if you go ahead, and the report is written so any team could build from it.

If you want a first view before you commit, the AI readiness scorecard takes a few minutes. For what the build itself involves, read the primer on AI workflow automation.

Next step: See the 2-week AI readiness assessment