What counts as an AI agent outcome?

An AI agent outcome changes what can happen next. A reply can be useful without doing that. A work product counts when it reaches its stated destination, passes its written checks, preserves required approvals, and lets the next person or system act.

That still may not be the final business result. A checked renewal packet can let an owner decide. It does not cause the vendor to accept new terms. A finished podcast episode can be ready to publish. It does not guarantee an audience.

Put the promise at the work-product level

The accepted work product is the part a provider can promise

Public-service designers use several levels of outcome because one provider rarely controls the whole journey. The GOV.UK definition of a service distinguishes the result of one process step, the wider institutional result, and the user’s overall goal. It also acknowledges that parts of the journey can sit outside the service’s control.

That distinction fits AI work. A provider can contract to prepare a traceable packet, apply stated checks, stop at an approval, and deliver inside a window. The provider can be held to that work. The buyer, customer, regulator, market, and outside systems still affect the later result.

The UK Department for Education’s product principle defines an output as the result of activity and an outcome as the benefit or change a user experiences. That is a useful warning against counting artifacts. For a commercial work order, however, the accepted artifact is often the strongest promise the provider can make without claiming control over the world beyond it.

FidelicAI calls that promised part the work product. The work-product promise covers the agreed delivery, written acceptance checks, approval boundary, and delivery window. The customer agreement shown before payment controls. The promise does not turn an outside result into a deliverable.

Placement alone does not make the work useful

A file in the right folder can still be incomplete. A draft in Slack can still lack its source. A monitor can stay silent because nothing changed or because its connection failed. A refusal can protect the business or reveal that the role misunderstood the request.

Five facts make a work product useful downstream:

1. Destination. The work reaches the system, channel, folder, queue, or person where the next action occurs.
2. Acceptance. A fresh reviewer can apply the written check without guessing what “good” means.
3. Authority. Every required approval is visible, and consequential action stops where the role’s authority ends.
4. Exceptions. Missing sources, conflicts, failed checks, and open decisions remain attached to the handoff.
5. Next use. The person or system that depends on the product can continue the job.

Done when: the destination contains the stated artifact or authorized action, each acceptance check has a result, required approvals are recorded, every open exception has an owner, and the downstream user can perform the stated next step.

One podcast episode makes the levels visible

Start with a recorded episode. The owner wants a release package another producer can inspect and continue.

SADIE, FidelicAI’s production AI podcast producer, carries the post-recording work for that job. Her current role covers transcript cleanup, timecoded paper edits, dialogue editing, audio repair, mix and mastering, requested show notes, quality-control records, and editable handoffs. The stated connections support the source recordings, transcript and document record, review, and delivery operations relevant to that workflow.

The checked files are work products. The operational outcome is that publication or another production handoff can proceed without rebuilding the episode. Listener response and revenue are outside business results.

For this job, SADIE can replace paid human work in transcript cleanup, story preparation, dialogue editing, repair, mixing, mastering, show-note drafting, and packaging. SADIE does not replace recording, video editing, rights decisions, publication, promotion, the show owner’s final editorial judgment, or the whole producer relationship.

The podcast editing work order gives the current product detail. The finishability question explains the delivery and remedy test.

Activity measures can reward the wrong behavior

Message count, files created, tickets touched, and hours active can describe activity. They do not establish acceptance or downstream use.

GOV.UK’s guidance on measuring service success recommends combining performance data with evidence such as task completion, user research, feedback, and financial information rather than relying only on digital analytics. The source concerns government services, not AI agents. The transferable principle is to measure the whole task the user needed to finish.

For an AI agent, keep five measures separate:

• Acceptance: Did the complete work product pass the stated checks?
• Correction: What had to change before acceptance, and who owned the cause?
• Downstream use: Did the next person or system actually use the handoff?
• Timeliness: Did the accepted result arrive inside the agreed window?
• Outside result: What happened later, and can the evidence support attributing any part of it to this work?

NIST’s AI Risk Management Framework asks whether an AI system achieves its intended purpose and stated objectives. NIST does not define FidelicAI’s work products or certify them. Its useful contribution is the insistence that the intended purpose, measurement method, and production behavior remain explicit.

Refusals and open exceptions can move work forward

A refusal is not automatically an outcome. “I cannot do that” may simply return the problem to the owner.

A useful refusal changes the downstream state. It records the request, the authority or source rule that blocked it, the affected work, the approval owner, and the permitted next move. The owner can then approve a different action, correct the rule, supply the missing evidence, or end the request.

The same test applies to an exception. A missing contract clause is not a completed vendor decision. A record that states which agreement is missing, who can supply it, which conclusion remains blocked, and when the decision must be revisited can still prevent a false renewal recommendation.

The agent-constitution guide explains the authority boundary. The production-survival question explains how exceptions, regressions, and repairs stay visible after launch.

Use four questions before accepting the result

Ask each in order:

1. What exists? Identify the artifact or authorized action.
2. Did it pass? Apply the complete acceptance check to a fresh copy or record.
3. What can happen next? Identify the downstream person, system, or decision that can now move.
4. What remains outside the work? Separate the customer, market, regulator, licensed professional, or owner action the provider cannot control.