AI did the work. Who gets the credit?

Open almost any client timesheet and you will see a familiar record: a person, a project, a description, and a number of hours.

Now look at how the work was actually produced.

A strategist used an AI assistant to organize interview notes. A developer delegated a test suite to a coding agent, then reviewed and corrected it. A research agent ran overnight and prepared a source pack before the team arrived. The delivered work is a compound of human judgment and AI contribution, but the official record shows only the person who started the timer.

The alternative is not much better. AI observability tools can show model calls, tokens, traces, and cost. That is useful technical evidence, but it does not explain which client project moved forward, who was responsible for the result, or what was ultimately delivered.

One system sees the human half. The other sees the AI half. Neither tells the whole story.

The unit of work has changed

Time tracking assumes that labor and elapsed human time are close enough to treat as the same thing. That assumption was always imperfect. It becomes much less useful when a person can delegate part of a workstream to an agent that drafts, analyzes, codes, or reviews on its own.

This does not make human time irrelevant. Human effort still matters for staffing, capacity, cost, and many contracts. It does mean that human time alone is no longer a complete account of production.

The gap appears in ordinary questions:

• A client asks whether AI was materially involved in a deliverable.
• Finance wants to know whether AI tooling is improving margin or simply adding cost.
• A delivery lead wants to compare two workstreams that used different combinations of people and agents.
• Security needs to know which tools and models touched a sensitive project.
• Leadership wants evidence that AI adoption is improving delivery, not a count of purchased licenses.

A timesheet cannot answer those questions if the agent contribution disappears. A token dashboard cannot answer them if the human ownership and business outcome disappear.

The missing layer is work attribution.

Three modes of modern work

The simplest useful model begins with three modes.

Human work

A person performs the work without material AI involvement. Routine autocomplete, spell-checking, or background software does not need to become a special disclosure unless a policy requires it.

The familiar time entry remains sufficient: who worked, where the time went, and what changed.

Human-directed AI work

A person frames the problem, supplies context, delegates part of the work to an AI system, evaluates the result, and remains responsible for what ships.

The person should remain the primary owner of the delivery. The AI contribution should appear as detail beneath that ownership: what the system contributed, which tool or model was involved, whether the output was reviewed, and what evidence exists.

This is closer to a manager delegating work than to a piece of software being listed as the author. Delegation does not erase the manager’s responsibility. It adds context about how the result was produced.

Autonomous agent work

An agent performs a bounded activity without a person directing every individual run. Examples include a scheduled research monitor, an automated code reviewer, or a system that classifies incoming documents.

These runs deserve their own records because there may be no human time entry to anchor them. They should still connect to a responsible team, project, workstream, policy, and review process.

Autonomous does not mean unaccountable. It describes how the work was initiated, not who bears responsibility for deploying the system.

Attribution is not authorship

The question “who gets the credit?” can sound like a debate about whether a model is an author. That framing is too narrow for operational work.

Teams need to distinguish at least four things:

1. Ownership: Who was responsible for the workstream and the final result?
2. Contribution: What material part did the AI system perform?
3. Evidence: What facts can support the record, such as tool activity, a model identifier, cost, or a linked agent session?
4. Review: Who checked the output, and what level of review occurred before delivery?

These fields can point to different actors without creating a contradiction.

Consider a developer who asks a coding agent to prepare a test suite. The agent generates most of the first draft. The developer checks the assumptions, fixes brittle tests, runs the suite, and approves the pull request.

The honest record is not “the developer wrote every line.” It is also not “the AI delivered the test suite.” The developer owns the delivery. The agent made a material coding contribution. The final output was human-reviewed.

That is useful attribution because it preserves responsibility and describes the production process.

The human should remain the headline

Poor attribution can quickly become surveillance. A system that counts prompts, compares token usage between colleagues, or rewards people for selecting “AI-assisted” creates incentives to perform activity rather than improve work.

Use this design principle: the human is the headline, and the AI contribution is the tracked detail beneath it.

That principle changes the tone of the record. Attribution becomes evidence that a person can direct tools, exercise judgment, and deliver responsibly. It does not reduce the person to the percentage of a document they typed manually.

It also avoids a false choice between transparency and professional value. A team can say that AI contributed materially while still being clear that people framed the problem, supplied proprietary context, reviewed the result, and accepted responsibility for what the client received.

Record facts, not fictional precision

Once teams notice the gap, there is a temptation to invent a symmetrical measure. If the human worked for 47 minutes, perhaps the agent did “two and a half hours” of work. If a draft arrived quickly, perhaps the system “saved eight hours.”

Those numbers are rarely observed facts. They are estimates built on an imagined counterfactual: how long would this exact work have taken, at the same quality, without AI?

There are cases where a controlled baseline can support that comparison. Most daily work does not have one.

A stronger record starts with what is known:

• Actual human time
• Actual agent session or activity evidence
• Tool and model, where available
• Token use and compute cost, where available
• Type of contribution
• Human review status
• Workstream and delivered outcome

What a useful attribution record looks like

A good record should be understandable six weeks later by someone who did not participate in the work.

Project: Acme platform migration Workstream: Quality assurance Owner: Maya Chen Human time: 1h 20m AI involvement: Collaborative AI contribution: Generated initial migration test cases Tool: Claude Review: Maya verified coverage, corrected environment assumptions, and approved the final suite Outcome: 34 migration scenarios added, including six failure paths identified during review

The entry gives the person credit for the delivery without hiding how the first draft was produced. It also gives finance, delivery, and the client a shared set of facts.

Not every audience needs every field. Internal operations may need tool, model, cost, and evidence. A client may need only a concise disclosure that the work was AI-assisted and human-reviewed. The underlying record can support both views.

Start with a small model

Teams do not need to capture every prompt or instrument every tool before they can improve attribution.

Begin with four decisions:

1. Define what counts as material AI involvement.
2. Choose a small set of involvement levels and contribution types.
3. Name the person or role accountable for review.
4. Attach the record to a project, workstream, and outcome.

Then improve the evidence over time. Manual declarations can later be supported by provider data, IDE integrations, agent events, or privacy-preserving activity signals. The vocabulary should remain stable even as capture becomes more reliable.

This is the important shift: AI usage becomes part of the delivery record, not a separate technical exhaust stream.

One record for the work that actually happened

The goal of attribution is not to distribute credit with mathematical precision. It is to create a trustworthy account of production.

People frame problems, make choices, carry relationships, and accept responsibility. AI systems can research, draft, transform, code, and review at a scale that materially changes delivery. A credible work record needs to hold both truths at once.

Human hours still matter. AI agent contributions now matter too. Track them together, and the question “who worked on this?” becomes answerable again.