AI Readiness Interview: Eight Questions to Ask Your Department Leads

An AI readiness interview is a structured conversation in which a business owner asks each department lead the same questions about how work really gets done. Its purpose is to find which processes, records and decisions are stable enough for an AI agent to take over part of the work, and which need a named owner, a written procedure or cleaner information first. The eight questions below are built for that purpose and take about one hour per lead.

Key takeaways

• Interview each department lead separately with the same eight questions so the answers can be compared side by side.
• Listen for a named trigger, a trusted source, a written approval boundary and a measure of success, not for enthusiasm about AI.
• Score each process as ready, ready with conditions or not yet before anyone chooses a tool or builds anything.
• Start with one process that scores as ready, in one department, with one approver who agrees to the boundaries in advance.
• Keep the first agent role narrow enough that you can observe its work and judge the result without guessing.

Why Interview Leads Before You Choose a Tool

Tools change quickly. The operating facts of a department change slowly, and the department lead holds most of them. An agent needs the same facts a capable new employee would need on day one: what starts the work, what inputs arrive, what a good output looks like, who approves what, and where the reference material lives. Most leads know these facts well but have never had to write them down in one place.

The interview gets those facts out of people's heads and into a document you can review. It also shows where your company's knowledge is consistent and where it depends on one person's memory. That second finding matters as much as the first. A process that works only because one senior employee always remembers an exception is a risk before it is an automation candidate.

Prepare the interview in three steps. Send each lead a one-page brief that says what you are trying to learn and why, without naming a product. Ask them to bring one recent example of the work, such as a completed request or a closed ticket. After the session, send back a summary in their words within two days and ask for corrections. A lead who corrects the summary is a lead who will later trust the agent's boundaries.

Once you have a candidate in mind, Draw Your AI Stack: A One-Page Architecture Map for Leaders shows how data, tools and people relate to each other.

Questions 1 to 3: How the Work Actually Runs

Question 1: Which recurring task takes the most of your team's time each week, and what starts it?

Listen for a named trigger: an incoming email from a particular sender, a form submission, a date in a calendar, or a request in a shared queue. If the lead describes the task only as 'handling customers' or 'keeping things moving', the work has no reliable starting point. An agent needs something concrete to respond to, and a vague trigger usually means the work is still decided case by case.

Question 2: Walk me through the last time this task was done from start to finish. Which steps were repeated, and which needed judgment?

Listen for the split. Mechanical steps such as copying data, checking a field, routing a request or preparing a standard reply are candidates for delegation. Steps that require a decision about a customer, a price or a risk stay with a person at first. Ask the lead to label each step as repeat or judgment, and note where they hesitated, because hesitation often marks where the unwritten rules live.

Question 3: How often do the rules for this work change, and who decides when they change?

Listen for a named decision owner and a rate of change the lead can describe in ordinary terms. If the rules shift in conversations and nobody records the new version, an agent will act on instructions that are already out of date. That is a process problem, and fixing it comes before automation. Leads who answer 'it depends on the week' should be asked for the three most common reasons it depends.

Questions 4 and 5: Where the Knowledge Lives

Question 4: Which documents, spreadsheets or systems does this work depend on, and when two sources disagree, which one does the team trust?

Listen for a single source of truth. A lead who names one system and one owner for it has given you a clean starting point. A lead who names four versions of a price list, or a shared folder where drafts and final files sit side by side, has shown that cleanup comes before any connection. An agent that reads conflicting sources will produce conflicting answers, and reviewers will lose trust quickly.

Question 5: If a capable new employee joined this team tomorrow, what would they get wrong in their first week because nobody had told them?

This is the most useful question in the set. The answers describe tacit knowledge: the exception for a long-standing client, the threshold above which a manager must look, the wording a regulator expects, the supplier who always delivers late. This knowledge is what an agent's memory and instructions must capture, and it is also what makes the work hard. A lead who lists five such points has shown where the process is more complex than the flowchart suggests. A lead who lists none may be underestimating the difficulty.

Questions 6 and 7: What the Agent Must Never Do Alone

Question 6: Which decisions in this work involve money, customer commitments, personal data or public statements, and who must approve each of them?

Listen for an approval boundary for each type of action. The useful answer is a short list in three groups: actions the agent may take on its own, actions it may draft but not send, and actions it must never take. The NIST AI Risk Management Framework asks organizations to make accountability for each AI use explicit, which is the same question in different words. A lead who answers 'everything goes through me' is describing a bottleneck. The agent should then prepare material for that person rather than act on outcomes until the approval load is redesigned. The AI agent permission review guide explains how to keep these boundaries current once they exist.

Question 7: Which systems would this work need to touch, and what is the least access that would still let it finish the job?

Listen for read-only versus write access, separate accounts for the agent, and an expiry for every grant. The OWASP authorization guidance is a useful reference: each connection should carry only the rights the task needs, checked against the role rather than the convenience of the setup. Access that a person would not keep for a year should not be given to an agent for a year. If the lead says the work needs five systems, the first role is probably too broad, so narrow it until the access list can be reviewed in one sitting.

Question 8: How Will You Know the Agent Is Working?

Question 8: What result would make you say, after a month of use, that this is worth continuing, and what result would make you stop?

Listen for a success criterion stated in the lead's own terms. Good answers are observable: the time between a request arriving and a first reply, the share of items that reach the approver without correction, or a backlog that clears by a set day. Weak answers such as 'better', 'faster' or 'more efficient' cannot be checked. If the lead cannot name a measure, the task is not ready, because nobody will be able to say whether the agent helped.

The measure also determines what the agent must record. Logs, traces and metrics give you a trail of what the agent did, which tool it called and what came back. The OpenTelemetry observability primer describes these signals in general terms, and they are what make the success criterion answerable after the fact rather than a matter of opinion.

Scoring the Answers and Choosing the First Task

After the interviews, score each candidate process against four conditions: a named trigger, a single trusted source, written approval boundaries and an observable success measure. A process that meets all four is ready. A process that meets three, with one gap that can be closed quickly, is ready with conditions. A process that lacks a trigger or a success measure, or whose rules change so often that no stable instruction set exists, is not yet ready, and the honest next step is documentation rather than automation.

Use this sequence once the scores are in:

• Write each lead's answers into a one-page brief per process, using their own words.
• Score each process against the four conditions and mark it ready, ready with conditions or not yet.
• Choose one ready process, inside one department, with one named approver.
• Draft the first version of the agent's role: memory needs, tools, approval boundaries and success measure.
• Agree the success measure and the stop condition in writing before anything is built or connected.

Illustrative scenario: an operations lead at a small company says quote follow-ups start when a quote is sent, that the customer record is the trusted source, and that success means a follow-up goes out within one working day. She cannot say who approves discounts. The process scores as ready with conditions. The first role can draft follow-ups from the customer record, but any discount question goes to the sales director until that boundary is written and agreed.

What a Managed Approach Does With the Answers

Be Human turns the interview into a first engagement that defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. These are written down before the agent runs, so the owner can see what it was given, what it may do and why. Once the role is live, managed operations can include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements. That continuity matters because the answers from the interview will change as the work changes, and the agent's instructions must change with them. Teams that will use the agent day to day also need a clear way in, which is covered in How to Get Your Team to Actually Use an AI Agent.

Take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Who should run the interview, the owner or an outside consultant? The owner should attend at least the first sessions, because the answers reveal priorities and approval limits that only the owner can settle. An outside consultant can prepare the brief, take notes and draft the summaries, which keeps the owner's time on decisions. Whoever runs it, the same eight questions should go to every lead so the answers can be compared.

How long should each interview take? Plan for about one hour per lead, with the brief sent in advance and a summary returned within two days. Shorter sessions tend to skip Questions 5 and 6, which contain the most useful material. If a lead needs a second session, schedule it rather than stretching the first one.

What if a lead answers most questions with 'we don't have a process for that'? That answer is itself a finding, and it should be recorded as not yet ready rather than treated as a failure of the lead. Ask them to walk through the last time the work happened, even if it was informal. The steps they describe usually form the first draft of a process that can be written down and reviewed.

Can these questions work in a company without a technical team? Yes, because they ask about work, sources, decisions and results rather than systems or code. A non-technical owner can run the interview and score the answers with the four conditions. Questions about connections and access come later, once a specific role has been chosen and someone is responsible for setting it up.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.