Who this checklist is for and how to use it
This checklist is for operations leads, product owners and managing directors who have a process in mind for an AI agent and want to know whether their organisation is ready to put one into daily use. It assumes you have already seen what large language models can do and are now asking a harder question: what has to be true on our side before an agent handles real work?
Most of the questions have little to do with models. They are about how well the process is understood, whether the data is reachable and trustworthy, who is allowed to approve what, and who owns the result once it runs. Those are the things that decide whether a first agent reaches production or stalls after a few weeks of promising tests.
How to use it:
- 01Pick one process. Answer the questions for a single, named workflow. A checklist filled in for AI in general tells you very little.
- 02Answer with the people who do the work. A team lead and someone who handles the cases every day will give you more honest answers than a steering group.
- 03Tick only what is true today. Plans and intentions count as a no. The point is to see the gaps clearly.
- 04Read the readout, then the gaps. The score gives you a direction. The unticked items tell you what to fix first.
The checklist in section four saves your progress in this browser, so you can come back to it after checking an answer with a colleague.
How to pick the right first process
Your first agent sets expectations for every one after it. Choose a process where success is easy to see and a mistake is cheap to catch.
Good first candidates usually share these traits:
- High volume, repetitive work. Dozens or hundreds of similar cases a week, such as classifying incoming requests, drafting replies from known information, checking documents against rules or preparing records for review.
- A clear definition of done. Someone on the team can look at an output and say within a minute or two whether it is right.
- Inputs that already exist digitally. Emails, forms, tickets, PDFs or records in a system you control. If the work starts with a phone call and a notebook, digitise that step first.
- A human step you can keep. The agent prepares, a person approves. This keeps risk low while you learn how the agent behaves on real cases.
- A measurable cost today. Hours spent, turnaround time or backlog size. Without a baseline you cannot show that the agent made a difference.
Processes to leave for later: anything where a single wrong action is expensive or irreversible (payments, contract commitments, medical or legal judgements sent straight to a customer), work that depends on knowledge only one person holds in their head, and processes that change every month because the business is still working out how it wants them to run.
If you have several candidates, score each on value (hours or delay removed) and on risk (the cost of a wrong output that slips through). Start with high value and low risk, even if another process looks more exciting. A small agent that runs every day builds more trust than an ambitious one that never leaves testing.
Start with the process: the first questions
Before data, systems or models, the process itself has to be clear enough to describe to someone new. If you cannot explain it to a new colleague, an agent will struggle to follow it too. These are the first questions from the checklist, the Process group:
- Is the process written down well enough that a new colleague could follow it?
- Is there one person who can decide how the process should work?
- Does the process run often enough to matter, weekly or more?
- Can someone check within a few minutes whether an output is correct?
- Do you know roughly how much time or delay the process costs today?
- Do you know the most common exceptions and how they are handled today?
If you answered no to two or more, that is useful information. It usually means the first piece of work is mapping the process and agreeing on the rules, which you can do without writing any code. Teams that skip this step often find that the agent exposes disagreements about how the work should be done, and the build pauses while those get settled.
The rest of this guide covers the full checklist (data, systems and access, risk and oversight, people and ownership), how to read your score and deal with gaps, why first agents stall, and what a first release should include.