The AI readiness checklist: 30 questions before your first agent.

How to use this checklist

Answer each question with a plain yes or no. "Partly" counts as no; the point is to find the walls before an agent runs into them. Involve two or three people who see different parts of the company: a business leader, someone from IT and the person who runs the process you have in mind. Where the three disagree, you have found a gap worth talking about.

The checklist takes about thirty minutes. Count your yes answers per dimension and in total, then read the scoring guide below. If you would rather have an outside view, our AI readiness assessment covers the same ground in two to three weeks, with evidence instead of self-assessment.

The 30 questions

1. Strategy and leadership

Agents succeed where leadership has decided what they are for. These questions test whether that decision exists.

1. Has leadership named two or three business outcomes AI should improve in the next twelve months, in numbers?
2. Is there a named executive who owns AI results, not just AI experiments?
3. Do you have a written list of candidate use cases, ranked by value and feasibility, rather than a collection of ideas?
4. Is there a budget for the first production use case that includes operations after go-live, not only the pilot?
5. Would leadership accept an agent that handles 80 percent of cases well and routes the rest to people?

2. Processes

An agent automates a process. If the process only exists in people's heads, the agent inherits the ambiguity.

1. Is the process you want to automate documented well enough that a new hire could follow it?
2. Do you know the volume: how many cases per week, and how long each one takes today?
3. Are the decision rules explicit, including the exceptions and who handles them?
4. Do you measure the quality of the process today, for example error rate, cycle time or rework?
5. Can the process be changed without a lengthy approval chain, so that the agent can be introduced in stages?

3. Data

Most agent use cases run on documents, emails and records you already have. The question is whether an agent can reach them and trust them.

1. Are the inputs the process needs available digitally, rather than on paper or in personal mailboxes?
2. Do you know where the master data for customers, products and suppliers lives, and which system is the source of truth?
3. Is the data quality good enough that a careful colleague would rely on it without re-checking?
4. Do you have historical examples of completed cases that could serve as a test set?
5. Is it clear which data contains personal information and on which legal basis it may be processed?

4. Systems and integration

The agent has to read from and write into your systems. Integration is where most pilots discover their real scope.

1. Do the core systems in the process (ERP, CRM, ticketing, document store) offer APIs or connectors an agent could use?
2. Can IT create a dedicated technical user with limited rights for an agent within days rather than months?
3. Do you have a place to run an agent that security has approved, whether a workflow platform, a cloud environment or a partner's infrastructure?
4. Is there a way to log what a system user did, so that an agent's actions can be traced?
5. Have you decided where AI models may run, for example EU-only, and which data may leave the company?

5. People and skills

Agents change jobs. Readiness means the people affected can shape that change instead of resisting it.

1. Is there an internal person who could act as product owner for the agent, at least two days a week during the pilot?
2. Have the people who run the process today been involved in the discussion, not just informed?
3. Does your team have basic AI literacy: what a language model can and cannot do, and how to check its output?
4. Do you have, or can you access, engineering capacity to build and maintain integrations?
5. Is there a plan for what the freed-up time will be used for, so that the change has a positive story?

6. Governance and compliance

Governance is not a later phase. Under GDPR and the EU AI Act it shapes what you may build at all.

1. Is your data protection officer, or an external adviser, involved before the design is fixed?
2. Have you assessed whether the use case could fall into a higher-risk category under the EU AI Act, and do you track the changing status of those obligations?
3. Do you have a policy on which AI tools employees may use, and with which data?
4. Is the works council, where you have one, informed early about planned AI use in the workplace?
5. Is there a defined way to switch an agent off and route its work back to people?

How to score your answers

Count your yes answers. The total tells you how far you are from a first production agent; the per-dimension counts tell you where to start.

One warning about the arithmetic. A dimension with zero or one yes answer is a blocker regardless of the total. A company can score 24 with a governance score of one, and that company is not ready to let an agent touch personal data.

The most common gaps, and how to close them in weeks

Across the companies we talk to, the same four gaps come up most often. None of them takes a year to close.

No owner

The most frequent gap and the cheapest to fix. Name a department head who wants the outcome, give them a share of their time and a small budget, and let them choose the first use case. An owner without a use case is better than a use case without an owner.

Undocumented processes

You do not need a process-mining programme. Have the people who run the process walk through ten real cases while someone writes down the steps, the decision rules and the exceptions. Two afternoons produce a document good enough to design an agent against.

Data that cannot be reached

The data usually exists; it is locked in a system without an API or scattered across mailboxes. A narrow integration layer for the two or three systems that matter, plus a rule that the process runs through a shared inbox or ticket queue, is typically enough to start. Our data foundations work focuses on exactly this minimum, not on a data platform.

Teams that have never worked with a model

Low AI literacy shows up either as fear or as blind trust. A half-day of hands-on AI training for the affected team, using their own documents, fixes most of it. The EU AI Act's AI literacy duty, which has applied since February 2025, is a further reason to do this early; check the current requirements with your advisers.

Close these four gaps and most organisations move from the middle band to ready within one quarter, and the first agent pilot can start while the last gaps are being closed. Our article on why AI pilots fail shows what happens when they are skipped.