Sales and CRM agents

The right action. In the right record.

Improve the repeated decisions, API calls, and CRM updates inside a sales workflow.

Discuss your workload Read the evidence

Anatomy of the task

Illustrative workflow

Choose an action and update a CRM record.

Input

A conversation, the current account state, and the actions the agent is allowed to take.

Required output

The intended record changes, with valid arguments and no unrelated writes.

Evaluation

Select the appropriate tool
Use the correct record and field values
Reach the expected workflow state

Understand the workflow.

Sales systems repeat a relatively small set of operations across a large number of conversations. A good agent needs to understand the request, work within the available APIs, and leave the system in the right state.

Define the quality bar.

Use a bounded task and an inspectable target state. Evaluate the full action, including tool selection and arguments, rather than judging whether the response sounds plausible.

Improve the route.

Begin with the prompt, schema, and tool-call adapter. Compare model routes on the same task slices. Use specialist training when repeated failure modes and accepted corrections justify it.

Keep frontier capability.

Keep frontier reasoning for unfamiliar workflows, complex ambiguity, and cases where the strongest model changes the result.

Published evidence

See the comparison.
Keep the context.

The published CRM study examines prompt and adapter optimization on measured task slices. It is evidence about those routes, not a claim that model weights were trained.

Read the sales and crm agents study

Who this fits

A task your team
knows well.

A team with repeated agent steps, access to representative tasks, and a way to check the resulting CRM state.

Bring a workload

Another place to startOperations workflows

Own your intelligence

Talk to Orchestra