How to Choose the First Process to Agentize

The first process you hand to agents matters more than any that comes after. If it works, the company learns how to do the next one. If it fails, the conclusion is usually "agents do not work here", and the second attempt never gets funded.

So the question is not which process is most exciting. It is which one is most likely to work and to be noticed when it does.

Start from a map

Before choosing, write down how the company works today: which processes exist, which systems they touch, who decides what, and where work waits. It does not need to be exhaustive. One page per area is enough to compare candidates side by side.

Five criteria

A good first process meets all five.

1. It takes real time, every week

Volume is what makes the result visible. A process that happens twice a year gives you nothing to measure and nobody who feels the difference.

2. Its rules can be written down

If the people who do the work can explain how they decide, an agent can be given those rules and evaluated against them. If every case is "it depends", start somewhere else.

3. It crosses more than one system

This is where agents differ from older automation. Work that means reading an email, checking the ERP, and updating the CRM is expensive for people and natural for an agent with the right tools.

4. Its mistakes can be undone

A wrong follow-up email can be corrected. A wrong bank transfer is much harder. Choose a process where errors are cheap to catch and cheap to reverse, and put approvals on the steps that are not.

5. Someone owns the outcome

One person must want the process to work, agree what "done well" means, and review the results. Without an owner, nobody decides when limits can widen, and the project stalls at the pilot.

Processes to avoid at the start

• The most critical one. The process the company cannot afford to get wrong is a bad place to learn.
• The one nobody understands. If the current process is undocumented and inconsistent, fix that first. An agent will copy the confusion.
• The demo. Something that impresses in a meeting but saves nobody any time will not survive the first budget review.
• Everything at once. A company-wide program before one process works spreads effort too thin to finish anything.

Good first candidates

These tend to meet the five criteria in most companies:

• Following up on overdue invoices.
• Reconciling payments against invoices.
• Answering and resolving routine customer requests.
• Qualifying inbound leads and booking the first meeting.
• Processing supplier invoices and purchase orders.

The right one depends on your map, not on this list.

What to measure from day one

Agree three numbers before any code is written:

• Resolution rate: the share of cases the agent finishes inside its limits, without a person.
• Quality: how often a reviewed result was correct.
• Cost per resolved case: what each finished case costs to run.

These numbers are the evidence that lets the process move from "a person approves every action" to "the agent works alone inside limits". They are also what you will show when someone asks whether to do a second process.

Then connect the next one

The second process should touch the first. If agents follow up on invoices, reconciling payments is a natural neighbor, because the two can hand work to each other. That hand-off is the first step from L2 to L3 on the five levels of company autonomy.

If you want help drawing the map and choosing, book a call.