Your first AI agent needs a job description.

An AI agent needs a clear job: what starts the work, which information it can trust, what it may do, and when a person takes over. Define those things before choosing the technology.

Imagine giving a new sales hire access to your CRM and saying, “Improve the pipeline.” No agreed qualification criteria. No clear territory. No definition of a useful meeting. You would expect inconsistent results.

“Build us an AI sales agent” is often the same brief with a different job title.

For a B2B team, a better starting point is one recurring task with a clear owner and a result someone can check.

Start with the handoff that keeps breaking

Consider the handoff from a signed deal to customer onboarding. The contract is in one place, sales notes in another, and the delivery team has to ask what was actually promised.

A useful first agent could prepare an onboarding brief from approved deal records. Its job would be to collect the evidence, identify missing information and create a draft for the account owner to review.

That is a specific piece of work. You can inspect its output and decide whether it made the handoff better.

Write the job description before the prompt

A one-page brief should answer six questions.

1. What triggers the work? For example, a deal moves to closed won and the signed agreement is available.
2. Which sources may it use? Specify the agreement, approved CRM fields and relevant meeting notes. Decide which source wins when they disagree.
3. What must it produce? Define the brief: customer goals, purchased scope, commitments, key contacts, open questions and the next owner.
4. What may it change? Preparing a draft is different from updating contractual commitments or sending a customer email. Make the boundary explicit.
5. When must it stop? Missing scope, conflicting dates or an unclear contact should create a question for a person, rather than an invented answer.
6. Who owns the result? Name the person who reviews the brief, handles exceptions and decides whether the agent is doing useful work.

If the team cannot agree on these answers, there is a process decision to make before there is an agent to build.

Keep the first assignment small enough to inspect

For this onboarding example, begin with draft preparation. Keep customer communication and approval of the handoff with the account owner.

A simple rule can start the workflow when a deal closes. AI can help turn the unstructured notes into a readable brief. Required-field checks can flag missing inputs. Each part has a defined role; the whole workflow does not need to make its own decisions.

The first version should make it easy to see where every important claim came from. A reviewer should be able to check a delivery commitment against the agreement without searching through five systems.

Test the awkward cases

Use a small set of historical deals that the team is allowed to use for testing. Include a clean handoff, incomplete notes, a last-minute scope change and two sources that contradict each other.

Write down what an acceptable brief would contain before looking at the agent’s answer. Check whether it captures the commitments, flags uncertainty and avoids introducing unsupported details.

A fluent summary is not enough. The delivery team needs an accurate account of what it has agreed to deliver.

Measure the whole job

Record how long the handoff takes today, including the questions and corrections that follow. Then compare that with the time needed to generate, review and correct an agent-prepared brief.

Track missing commitments, reviewer edits and follow-up questions alongside preparation time. Decide in advance which errors would stop the pilot.

If a brief takes less time to draft but more time to verify, you have moved the work. The goal is a better handoff with less total effort.

Build the next agent from what you learn

Once this assignment works reliably within its agreed limits, consider the next recurring task. You might prepare renewal briefs or assemble account research for a seller. Give each one the same clarity about inputs, authority and ownership.

This is how AI fits into commercial architecture: objectives determine what is worth improving, process defines the work, organization assigns responsibility, and tools support the execution.

Before you ask which agent to build, ask which piece of work your team can describe clearly enough to delegate.

Unsure where to start? A Commercial Architecture Stress Test examines the gaps in your commercial operation and helps establish what to fix first. Once the priorities are clear, Optimization is where we build the flows and agents that support them.