How to choose the first function to automate

Key takeaways

• Score candidate functions on five tests: volume, rules, data access, cost of a mistake and measurability.
• Start where errors are cheap to reverse and a person can approve each action.
• One function carried end to end is worth more than five demos.
• Write a one-page brief before building: owner, systems, authority level, success number and stop conditions.

Almost every organization that decides to adopt autonomous AI faces the same question in the first week: where do we start? The answer matters more than it looks. The first function you hand over sets the level of trust people place in the system, the size of the budget for the next one and the story the organization tells about the whole effort.

Choose well and the second and third functions become easy to approve. Choose badly and the project joins the long list of pilots that never left the pilot stage, a pattern we describe in why agentic AI projects stall.

Five tests for a good first function

We score each candidate from one to five on the tests below. The numbers are only a way to force an honest conversation, but the conversation is the valuable part.

1. Volume. Is there enough repeated work that the time saved will be visible within a month? A task done twice a week is a poor first choice, however annoying it is.
2. Rules. Is the work mostly rule-based, with a written policy and a clear list of exceptions? If three senior people would handle the same case three different ways, the function is not ready.
3. Data access. Is the information in digital form, in systems that can be read and written, such as a CRM, an accounting package, an inbox or a shared drive? Work that lives in people's heads or on paper needs a preparation step first.
4. Cost of a mistake. If the system gets one case wrong, how bad is it and how easily is it undone? A wrong reminder email is cheap. A wrong payment is not. Early on, choose work where a person can approve each action before it happens.
5. Measurability. Is there a number the leader already tracks, such as days to close, response time, cost per invoice or share of enquiries answered within an hour? If you cannot measure the before, you cannot prove the after.

A function that scores high on volume, rules, data access and measurability, and acceptably on the cost of a mistake, is a strong first candidate.

Where this usually works first

• Customer enquiries: first replies, qualification and routing to the right person, with a hand-off when judgement is needed.
• Collections and reminders: matching payments to invoices and sending polite, compliant follow-ups.
• Lead follow-up: responding quickly, qualifying and keeping CRM records current.
• Appointments: booking, rescheduling and reminders for clinics, salons and service businesses.
• Document intake: reading forms, invoices or applications and preparing them for review.

Functions to leave for later are those where a single error carries financial, legal or personal consequences: final pricing decisions, hiring decisions, payments above a threshold and anything involving employee relations. They can still be automated, but only after the system has earned authority on easier work. We describe how that authority is earned in autonomy is a dial, not a switch.

An illustrative example

Imagine a mid-sized distributor choosing between three candidates: customer enquiries, collections reminders and supplier negotiations.

• Supplier negotiations score low on rules, because every deal differs, and low on cost of a mistake, because a bad concession is expensive. Leave it.
• Customer enquiries score high on volume and data access, but quality is harder to measure and tone matters. A good second function.
• Collections reminders score well on every test: high volume, a written policy, data in the accounting system, mistakes that are easy to undo and a number the finance lead already tracks, days outstanding.

The distributor starts with collections reminders, with a person approving each message at first. After a few weeks of evidence, the approvals narrow to unusual cases. That is the first function earning its place, and it makes the case for the next.

Small enough to finish, large enough to matter

There are two opposite traps. The first is choosing something so small that success proves nothing, such as an assistant that answers three questions. The second is choosing something so broad that it cannot be finished, such as automating all of customer service at once.

A good rule is to carry one function from start to finish, with real data, real approvals and a real number, rather than to build several impressive demos. A single function that runs reliably in production teaches more than five that never leave a slide deck.

Write a one-page brief first

Before anything is built, write one page that answers these questions. If you cannot fill in a line, that is the work to do first.

• Function and owner: what is being handed over, and who is accountable for the result.
• Systems: which tools hold the data and which the system must act in.
• Volume: roughly how many cases a week, and how long each takes today.
• Policy: the written rules and the list of known exceptions.
• Authority level: what the system may do alone and what needs a person's approval, starting conservatively.
• Success number: the one metric that will show whether it worked, and its current value.
• Stop conditions: what would make you pause or roll back, such as error rate or customer complaints.

Where to go from here

If you are weighing two or three candidates and are not sure which scores highest, we are happy to look at them with you. You can speak with Primues to get started, browse the use cases most teams begin with, or read how we structure each function on our platform page. For how engagements are scoped, see pricing. New to the terms in this article? Our glossary explains them in plain English.