Start with the workflow, not the model

Ask a leadership team where AI agents should go first and you will usually hear a list of tools: a copilot for the sales team, a chatbot for HR, a pilot with whichever model was in the news that week. It is an understandable instinct. It is also the most common reason AI programs produce demos instead of results.

An agent is only as useful as the work it is given. The better question is not “which model should we use?” but “which piece of work should we redesign first?” Answer that well and the technology choices become much easier.

Map the work before you choose the tool

Start with one function, not the whole enterprise. Finance, procurement and revenue operations are good candidates because their work is documented, repetitive and tied to numbers someone already tracks.

Inside that function, list the workflows end to end: what triggers them, which systems they touch, who hands work to whom, and where things wait. You are looking for the places where people spend their time moving information between systems, chasing approvals or resolving the same kinds of exceptions again and again.

Two or three working sessions with the people who do the work will tell you more than any vendor survey. Ask them where the backlog builds up, which steps they would happily never do again, and which errors cost the most to fix later.

Score each candidate on five questions

Once you have a list, score each workflow from one to five on the questions below. The highest totals are your shortlist.

Notice what is missing from that list: novelty. The best first agent is rarely the most impressive one. It is the one that removes a real bottleneck, earns trust quickly and gives you a clean before-and-after number.

The best first agent is rarely the most impressive one. It is the one that removes a real bottleneck and earns trust quickly.

Give the agent a job description

Before anything is built, write down what the agent is responsible for in the same way you would for a new hire. A useful job description covers four things:

• Role. One narrow responsibility, such as triaging invoice exceptions, not “helping finance.”
• Permissions. Which systems it can read, which it can write to, and which actions always need a person’s approval.
• Hand-offs. Where its work comes from, where it goes next, and how it escalates what it cannot resolve.
• Audit trail. What it records about every action so a reviewer can see what happened and why.

This document does double duty. It keeps the build focused, and it gives risk, security and audit teams something concrete to review early, which is far easier than retrofitting controls after a pilot is already running.

Build on the systems you already run

A first agent should not depend on a platform migration. Agents can work across your existing ERP, CRM and finance tools through their interfaces, which keeps the scope small and the system of record unchanged. If a migration is genuinely needed later, it should be a separate, deliberate decision.

Deliver in stages. Start with the agent proposing actions for a person to approve, measure how often its proposals are accepted unchanged, and only widen its permissions once that number is consistently high.

Measure it like any other operating change

Agree the baseline before go-live: current cycle time, backlog size, exception rate or hours spent. Then report the same numbers on the same cadence afterwards. If the number does not move, the redesign needs work, whatever the demo looked like.

That discipline is what turns a single agent into a program. Once the first workflow shows a measured improvement, the next conversations get easier, because they are about where to repeat the result rather than whether AI works at all.