Which white-label agent design can your agency support?

Pilot a bounded workflow and a client-specific runtime against the same synthetic queue. Sell the option that meets the client’s acceptance target, requires support your team can staff, and passes the agreed persistence checks.

Price the minutes spent on exceptions

Start with a single deliverable such as drafting a reply to an inbound support request. Compare a bounded workflow that routes uncertain cases to a reviewer with a client-specific runtime that can consult an additional approved knowledge source. Give both a batch containing common cases, unfamiliar cases, and incomplete requests. Count an accepted draft only after a reviewer checks its factual basis and whether the customer can act on it. Keep first-pass acceptance and repair minutes in the same worksheet. In an illustrative pilot of 100 requests, twelve four-minute repairs consume 48 minutes, while five fifteen-minute repairs consume 75. A design with a better first-pass acceptance rate can still require more support. Use the measured difference to decide what review volume belongs in the client’s service agreement.

Building effective AI agents \ Anthropic

Decide which pilot outputs should survive

Map each byproduct to a retention decision before picking the deployment. A draft awaiting approval may need a durable queue, while a scratch copy of the synthetic support attachment may only need to live for the session. Run one staged task that creates both. Record the approved draft identifier, close the session, and check that the approved draft is still available through the normal operator view. Check separately whether the scratch copy is absent when you reopen the workspace. Configure the pilot to exercise the distinction described by the deployment guide between temporary storage and a dedicated volume for durable output. Compare the two designs on how many records an operator must reconcile after interruption. A draft that reappears with an extra queue entry adds review work every time the service resumes.

Securely deploying AI agents - Claude Code Docs

Sources