AI Agent Readiness Checklist: 12 Checks

This AI agent readiness checklist has 12 checks. Each one has a pass test, so you can score yourself in an afternoon with no tool and no email gate.

Seven of the checks block a pilot. Five you can fix while the pilot runs. Knowing which is which saves weeks.

Readiness does not mean “we have a data lake and a strategy deck.” It means an agent could start on a real workflow next Monday and nothing on this list would stop it.

What readiness really means

Most readiness assessments measure maturity. They ask whether you have a strategy, a center of excellence, and a roadmap. That tells you little about whether one agent can run one workflow safely.

I’d define readiness narrowly. For one named workflow, can the agent get the data it needs, act inside safe limits, be tested before it touches production, and be stopped or corrected by a person? If yes to all four, you are ready for a pilot. Org-wide maturity can come later.

The stakes are real. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, citing rising costs, unclear business value, and weak risk controls. Each of those maps to a check below.

Pick the workflow first. If you have not, start with which workflows to automate first. This checklist assumes you have one in mind.

How to use the checklist

For each check, score:

2 if you meet the pass test today
1 if you partly meet it
0 if you do not

Checks marked BLOCKER must score at least 1 before you start. Checks marked FIX IN PILOT can score 0 as long as you assign someone to fix them in week one.

Total score is out of 24. Bands are my own rule of thumb, not an industry standard:

20 to 24: Ready. Start.
14 to 19: Close. Fix the weakest blockers, then start.
Under 14: Not yet. Spend two to four weeks on the gaps.

Data and access checks

Gartner’s survey of 1,203 data management leaders found that 63% of organizations either lack the right data management practices for AI or are unsure they have them. Gartner also predicts that through 2026, organizations will abandon 60% of AI projects that lack AI-ready data. Data is the most common gap, so it goes first.

Check 1: The agent can read the data it needs. (BLOCKER)

Pass test: Someone pulls 50 real cases from the workflow, with every field the agent would use, in under a day.
Who answers: Data or systems owner.

Check 2: There is one source of truth for each fact the agent uses. (BLOCKER)

Pass test: For each key field (price, policy, customer status), you can name the one system that is correct when two systems disagree.
Who answers: Process owner.

Check 3: There is a supported way in. (FIX IN PILOT)

Pass test: An API, database view, or export exists. You are not planning to scrape a screen or share a human login.
Who answers: IT or integration lead.
Why fix in pilot: A read-only export is enough for week one. A proper integration can follow.

Process and owner checks

Check 4: One named person owns the workflow. (BLOCKER)

Pass test: A named person agrees to spend a set number of hours each week reviewing agent output. Not a team. A person.
Who answers: The sponsor.

Check 5: The process is written down, or can be in a day. (FIX IN PILOT)

Pass test: A one-page description exists of the normal path, the inputs, and the outputs. If three people do it three ways, you agree on one.
Who answers: Process owner and a daily operator.

Check 6: Exceptions have a home. (FIX IN PILOT)

Pass test: You know roughly what share of cases are off-path, and each type of exception has a person or queue that handles it.
Who answers: Process owner.
Method: Sample 50 to 100 cases and sort them into clean, known exception, and novel exception. My post on which workflows to automate first walks through the count.

OpenAI’s agent guide names two triggers for handing work to a person: the agent exceeds a failure threshold, or the action is high-risk. Check 6 is where you decide what those look like for your workflow. The design details are in my post on AI agent escalation design.

Security and approval checks

Check 7: The agent gets its own identity with least privilege. (BLOCKER)

Pass test: A dedicated account or credential exists for the agent. It has only the permissions the workflow needs. It is not a shared admin login and not a named employee’s account.
Who answers: Security or IT.

OWASP’s entry on Excessive Agency lists three root causes: excessive functionality, excessive permissions, and excessive autonomy. Its prevention steps include minimizing tools, minimizing permissions, and requiring user approval for actions. Check 7 covers the permissions piece. For the full method see AI agent permissions and least privilege. OWASP also published a Top 10 for Agentic Applications in December 2025, built with more than 100 contributors. It is a good read for your security reviewer.

Check 8: Data handling is decided. (BLOCKER)

Pass test: You have written down what data classes the agent may touch, where the model runs, whether prompts are retained by any vendor, and where logs live. Your security reviewer has read it.
Who answers: Security and legal.

Check 9: The sign-off path is named. (BLOCKER)

Pass test: You can list who must approve (security, legal, the business owner), what each needs to see, and how long each usually takes. Nobody says “we’ll figure it out.”
Who answers: The sponsor.
Why this blocks: A missing approval step is the most common reason a finished pilot sits idle. If your problem is a stalled pilot, the pattern is in my post on moving from experimentation to production.

Test environment checks

Anthropic recommends extensive testing in sandboxed environments along with appropriate guardrails, because agents’ autonomy raises costs and the chance of compounding errors. The next three checks make that possible.

Check 10: There is a safe place to let the agent write. (BLOCKER)

Pass test: A sandbox, test tenant, or copy of the system exists where the agent can create, edit, or send without real effect. If the pilot is read-only plus drafts, mark this pass and say so.
Who answers: IT or integration lead.

Check 11: You have a test set with known answers. (FIX IN PILOT)

Pass test: At least 30 real cases exist where a person has already recorded the correct outcome. If not, the owner builds it in week one from the 50-case sample.
Who answers: Process owner.
Why it matters: OpenAI’s guide says to set up evals to establish a baseline before you optimize anything. A test set is the start of an eval.

Check 12: A person can stop it and see what it did. (FIX IN PILOT)

Pass test: The agent can be switched off in one step. Every action it takes is logged with time, input, and result. The pilot owner can read the log.
Who answers: Engineering lead.
What to log: My post on what to log for an AI agent audit trail has a field list.

Scoring and next steps

Add up your 12 scores. Then do these three things.

List your blockers that scored 0. There are seven blocker checks: 1, 2, 4, 7, 8, 9, and 10. Treat check 10 as a blocker only if the pilot writes to a system. Each blocker at 0 is a stop sign. Assign an owner and a date to clear it.
List your fix-in-pilot gaps. Assign each one to a named person with a day in week one. Put them on the pilot plan, not in a side document.
Decide the go date. If no blocker is at 0, set the start date now.

Here is a quick decision table.

Your result	What it means	Next step
All blockers at 1 or 2, total 20+	Ready	Write the brief and start
All blockers at 1 or 2, total 14 to 19	Workable	Start, with named owners for each gap
Any blocker at 0	Not ready	Clear it first, then rescore
Total under 14	Not ready	Two to four weeks of prep

Which gaps block a pilot?

Gaps in data access, source of truth, owner, agent identity, data handling, approval path, and (for write pilots) the test environment. Without these, the pilot either cannot run or cannot be trusted.

Which gaps can I fix during the pilot?

The integration method, the written process, the exception map, the test set, and the logging. They need effort, but they do not prevent a first, read-heavy, human-reviewed pilot from starting.

For how a short pilot is run, see my one-week AI pilot methodology.

Printable version

Copy this into a doc. Fill in the score and the name for each line.

AI AGENT READINESS CHECKLIST
Workflow: ____________   Date: ______   Scored by: ______

DATA AND ACCESS
[ ] 1  Agent can read the data (BLOCKER)         Score __  Owner ______
[ ] 2  One source of truth per fact (BLOCKER)    Score __  Owner ______
[ ] 3  Supported way in (FIX IN PILOT)           Score __  Owner ______

PROCESS AND OWNER
[ ] 4  Named workflow owner (BLOCKER)            Score __  Owner ______
[ ] 5  Process written down (FIX IN PILOT)       Score __  Owner ______
[ ] 6  Exceptions have a home (FIX IN PILOT)     Score __  Owner ______

SECURITY AND APPROVAL
[ ] 7  Own identity, least privilege (BLOCKER)   Score __  Owner ______
[ ] 8  Data handling decided (BLOCKER)           Score __  Owner ______
[ ] 9  Sign-off path named (BLOCKER)             Score __  Owner ______

TEST ENVIRONMENT
[ ] 10 Safe place to write (BLOCKER if writes)   Score __  Owner ______
[ ] 11 Test set with known answers (FIX IN PILOT) Score __ Owner ______
[ ] 12 Stop switch and action log (FIX IN PILOT) Score __  Owner ______

Total: __ / 24     Blockers at 0: ______
Go date: ______

Bottom Line

Readiness is about one workflow, not the whole company.

Score the 12 checks. Clear any blocker that sits at 0. Give each fix-in-pilot gap a name and a date.

Next step: print the sheet, get the owner and the security reviewer in one room for an hour, and score it together.

FAQ

How do I assess readiness for AI agents? Pick one workflow and score 12 checks across data, process, security, and testing. Give each a 0, 1, or 2 using its pass test. Total out of 24. Treat any blocker at 0 as a stop sign.

What do I need before a pilot? You need readable data, a named process owner, a dedicated agent identity with limited permissions, decided data handling, a named approval path, and a safe place to test. The rest can be built during the pilot.

Which gaps block a pilot? Missing data access, no source of truth, no owner, a shared or over-powered login, undecided data handling, an unnamed approval path, and no test area for write actions. Each can stop the pilot or make its results untrustworthy.

Which gaps can I fix during the pilot? You can fix the integration method, the written process, the exception map, the test set, and the logging and stop switch. Assign each to a named person in week one. Do not leave them as “later.”

Who answers each check? The process owner answers process and test-set checks. Security and legal answer identity, data handling, and approvals. IT or integration answers access and sandbox. The sponsor answers owner and sign-off. One person should hold the final sheet.

Do I need a readiness tool or a consultant to score this? No. The pass tests are written so a team can answer them in a meeting. A tool or outside review helps if you want a second opinion, but the sheet works without either.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect

AI Agent Readiness Checklist: 12 Checks Before You Start