Choose work that is frequent, dull, well documented and easy to check, where a mistake is cheap to undo and the facts come from systems of record. It needs a willing owner and a measurable baseline. Leave decisions that affect people's rights or move money for later agents, once your team has learned how to run one.
Why does the first process matter so much?
Because the first agent teaches the organisation how to run agents. It sets the habits for briefs, certification, review and ownership. If it goes well, the second agent starts from a working method. If it goes badly, the lesson most people take away is that agents do not work, when the real lesson was that the process was the wrong one to start with.
The first process is a training ground as much as a deployment. The argument that the first agent should do boring work follows from that: interesting work is usually interesting because it involves judgement, exceptions and consequences, which are the hardest things to govern well on a first attempt.
What should you look for?
| Criterion | Good first candidate | Poor first candidate |
|---|---|---|
| Volume | Hundreds of cases a week | A few complex cases a month |
| Rules | Written policy, consistent practice | Depends on who handles it |
| Sources | Systems of record | Emailed documents and memory |
| Checking | A reviewer can tell right from wrong quickly | Needs specialist investigation |
| Reversibility | Easy to correct | Hard or impossible to undo |
| Consequence | Internal, or small for the customer | Affects rights, credit, claims or money |
| Owner | Wants it and has time | Has it imposed |
| Baseline | Volumes and times already measured | Unknown |
Score each candidate against every row, not only the flattering ones. A process that scores well on volume but poorly on rules and checking will produce lots of fast, hard-to-verify output, which is the worst combination.
What does the agent do in a good first process, and what stays with people?
In most good first processes the agent gathers, compares, prepares and drafts, at depth 1 or 2. People decide anything with consequence. Typical examples:
- Preparing case files: pulling documents and data for an analyst's review.
- Chasing missing items: drafting requests for documents a file lacks.
- Matching and reconciliation: listing breaks and likely causes for a person to post.
- Internal queries: answering staff questions from a policy library, with sources cited.
- Triage and routing: classifying incoming work for the right team.
The person still owns the outcome in each. The agent removes the gathering and typing that surround it.
What does choosing look like in practice?
Say a bank's operations director has three candidates: answering customer complaints, assessing small business overdraft requests, and preparing files for periodic customer due diligence reviews.
Complaints score well on volume but poorly on consequence and checking, since tone and outcome matter and regulators watch them. Overdraft assessment involves credit decisions about people and businesses, which carry their own rules. Due diligence file preparation is frequent, follows written policy, draws on systems of record, and an analyst can check a prepared file in minutes. The agent's job is to gather and summarise; the analyst still decides.
The director chooses file preparation, with the financial crime operations manager as owner. Complaints become the third agent, after the team has a year's experience. Banks have particular considerations, set out in which AI agent a bank should deploy first, and the full selection method is in choosing the first process to give an AI agent.
Which controls should come with the first process?
- Redesign before deploying. Map the decisions first, as in redesigning a process for AI agents.
- A low starting depth, widened on evidence.
- Close attention in the early days, covered in running an agent's first week in production.
- A written lessons log, so the second agent starts from what the first one taught.
The full sequence is in our guide to deploying an AI agent.
Questions people ask
Should the first process be customer-facing?
Usually not. Internal work lets the team learn without customer impact. If it must face customers, start with drafts that a person sends.
What if leadership wants a visible first agent?
Visibility can come from results, such as time released and faster files, rather than from risk. A boring agent that works is a better story than an ambitious one that is withdrawn.
Can the first agent cover several processes?
Keep it to one. Several processes mean several owners, case sets and failure patterns, which is a lot to learn at once.