Pick work that is frequent, rule-bound, reversible and dull, with an owner who wants it done better. Choose the first process to give an AI agent by ruling out anything that moves money, affects vulnerable customers or depends on fine judgement, then scoring the rest on volume, documented rules, clean sources, past cases for testing and a measurable baseline.
Why does the first process matter, and who chooses it?
The first agent sets the pattern for every one after it. If it works, you will have a tested way to approve, certify, run and review agents. If it fails publicly, the programme loses support for a year. The aim of the first process is to prove the method, not to deliver the largest saving.
The chief operating officer, or whoever sponsors the agent programme, makes the choice. Operations leaders propose candidates. Risk vetoes unsuitable ones. The process owner of the chosen process must be willing; a reluctant owner is a disqualifier on its own. The underlying argument is set out in which process should you give an AI agent first?
What disqualifies a process?
- It moves money or is hard to undo. Save depth 4 work for later.
- It affects customers who may be vulnerable, or decides something with legal or similarly significant effects on a person.
- It depends on judgement that experienced staff cannot explain as rules.
- It is being redesigned. An agent learning a moving process learns the wrong thing.
- It has no past cases you can use for testing.
What are the steps?
- Person, operations leaders: propose a long list of candidate processes, with volume and pain points.
- Agent, depth 1: profiles each candidate from system data: monthly volume, time per case, rework, how many distinct case types, and which systems are touched.
- Person, risk lead: removes disqualified processes.
- Agent, depth 2: drafts a scoring table for the rest: volume, how rule-bound, reversibility, quality of sources, availability of past cases, and whether a baseline exists.
- Person, sponsor, with operations leaders: scores and shortlists two or three.
- Person, sponsor: meets each shortlisted owner to confirm they want it and will give time.
- Agent, depth 1: checks the data for the top choice: are the sources reachable, are past cases complete, are outcomes recorded?
- Person, sponsor: chooses, and the owner sets the first ceiling.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Long list | Nothing | none | Proposing candidates |
| Profile | Pulls volume, time, rework, systems | 1 | Nothing |
| Disqualify | Nothing | none | Risk veto |
| Score table | Drafts the scoring | 2 | Nothing |
| Shortlist | Nothing | none | Scoring and choosing |
| Owner check | Nothing | none | Owner's commitment |
| Data check | Checks sources and past cases | 1 | Nothing |
| Choose | Nothing | none | Sponsor's decision |
Which controls make the first process safe?
Low first ceiling. Start the first agent at depth 1 or 2 for every capability, whatever the plan allows. Setting the first ceiling for a new agent covers how to set it and when to raise it.
Reversible actions only. If the agent gets to depth 3, its actions should be easy to correct.
Internal impact first. A process whose output is checked by staff before anything leaves the firm limits the harm of early mistakes.
A baseline before go-live. Measure time, quality and rework now, so you can show what changed.
Shadow before acting. Run the agent in shadow on live cases before it does anything.
What does a good first process look like?
Take a bank's operations team that reconciles incoming payments against open invoices for business customers. Thousands of cases a month, clear matching rules, sealed sources in the ledger and bank statements, a full history of past matches, and unmatched items already go to a person. The agent reads and proposes matches at depth 1 to 2; staff confirm. Nothing moves money. That is the kind of work your first agent should do: boring work.
Compare a complaints triage process. It is high volume, but it involves vulnerable customers, regulated deadlines and judgement. It is a good second or third process, not a first. By the time you reach it, the first agent should have taught you the mechanics: how holds are staffed, how refusals are reviewed and how certification evidence is gathered.
What should you measure?
- Time per case, against the baseline.
- Agreement between the agent's proposals and staff decisions.
- Rework after the agent's work.
- Refusals and escalations, and whether they were right.
- Owner and staff confidence, through short regular check-ins, because the first agent has to win people over as well as work.
Which signs show the choice is going wrong?
The showcase is chosen. A customer-facing agent is picked because it impresses the board. It carries the most risk and the hardest edge cases.
The process is undocumented. The team cannot say what the rules are, so building the agent becomes a process-design project.
The owner is absent. Nobody reviews results, approves depth or answers questions, so the agent drifts.
No baseline. After six months nobody can say whether it helped.
The vendor chose it. The process was selected because the vendor had a demonstration for it. Choose the process first, then the vendor; how do you choose an AI agent vendor? covers what to ask.
Where Pryme fits
Pryme's Pre-Trained Agents cover common, rule-bound operational work, and Attest replays them on your own past cases before they act, which makes the data check above a direct test. The route from first process to wider use is in deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Should the first process be one the team hates?
A process people dislike because it is repetitive is a good candidate; they will welcome help. A process people dislike because it is chaotic is not, since the agent will inherit the chaos.
How big should the first process be?
Large enough that the benefit is visible within a quarter, small enough that one owner covers it. One team, one process, one market is a sensible boundary.
Can we start two processes at once?
You can, but you will learn less. One process run well produces the templates, approvals and habits the second one reuses.