The AI Readiness Assessment: 12 Questions to Ask Before You Invest

Why an Assessment Beats an Ambition

"We need an AI strategy" is how most doomed projects begin — solution first, problem later. After deploying AI across dozens of enterprises, we've distilled the pre-flight check into twelve questions in four groups. Score each 0 (no), 1 (partially), or 2 (confidently yes). Prefer to work through it on paper? Download the printable scorecard. The total tells you what to do next — and it's the same instrument we run, in depth, during our AI Opportunity Sprint.

Group 1: The Use Case (Questions 1-3)

• 1. Can you state the problem without the word "AI"? "Invoice processing takes 11 minutes each and errors cost £180K a year" is fundable. "Explore AI in finance" is not.
• 2. Is the task high-volume and pattern-rich? AI compounds on repetition. A judgement made 10,000 times a month is a candidate; a bespoke decision made quarterly is not.
• 3. Would 90% accuracy be valuable? If only perfection will do, AI can draft and a human must decide — which is still valuable, but changes the design and the business case. Be honest about which you're buying.

Group 2: The Data (Questions 4-6)

• 4. Does the data the task needs actually exist, digitally? Not "could be collected" — exists now, in systems, with history.
• 5. Can you get to it? APIs or database access, with the permissions and privacy basis to use it. A six-month integration project hiding inside an "AI project" should be named and budgeted as such (our API strategy guide is the companion read).
• 6. Would you trust it in front of a customer? If the CRM is 40% duplicates, the AI will be confidently wrong at scale. Data cleanup isn't a blocker — it's often phase one — but it must be in the plan.

Group 3: The Workflow (Questions 7-9)

• 7. Is there a named owner whose process this is? AI adopted into a workflow succeeds; AI installed next to one dies. No owner, no project.
• 8. Do you know where the human sits? Reviewing everything, reviewing exceptions, or spot-checking — the human-in-the-loop position determines cost, risk and throughput, and should be a decision, not an accident.
• 9. Can success be measured against today's baseline? If you don't know the current cost-per-case, error rate or cycle time, measure that first. You cannot prove a return on an unmeasured process.

Group 4: The Risk (Questions 10-12)

• 10. What happens when it's wrong? A mis-sorted email is cheap; a mis-priced quote is not. Map error cost, and design guardrails proportionate to it.
• 11. Is the use regulator- and customer-explainable? Decisions about people (credit, hiring, claims) carry duties — explainability, appeal routes, and increasingly EU AI Act obligations. Know your category before you build.
• 12. Can you run it where your data must live? Sovereignty and confidentiality constraints decide architecture (your cloud, on-prem, or a managed provider) — cheaper decided now than retrofitted.

Scoring: 18-24 — pick your first use case and move; the risk is now in delivery, not readiness. 10-17 — invest 4-8 weeks fixing your two weakest answers first; it will halve the project risk. Under 10 — you don't have an AI project yet, and the cheapest possible outcome is discovering that today.

The Point of the Exercise

Notice what the questions never asked: which model, which vendor, which framework. Those are week-two decisions that pretend to be strategy. Readiness lives in problem clarity, data reality, workflow ownership and risk honesty — and every one of those is fixable, cheaply, before you spend serious money. That's the entire philosophy behind starting small, fixed-price, and measurable.