Assistant Output Provenance Log (Delivery)
A sequence with 7 steps: Open a provenance record for every client deliverable before the first prompt is sent.
By InnovaAI ResearchPublished
What are the steps?
Assistant Output Provenance Log (Delivery)
- 01
Open a provenance record for every client deliverable before the first prompt is sent
Log the assistant used, model version and date, the operator's name, and the client account code. A single shared sheet per client beats scattered chat history when a client asks how a number was produced six weeks later.
- 02
Capture the prompt and the source inputs alongside the output
Paste the final prompt, list every uploaded file or linked document, and note which parts of the draft came from the model versus the operator. Client retainer work often mixes both, and the split matters during scope disputes.
- 03
Tag each output with a confidence tier and a named human reviewer
Three tiers are enough: verified against a primary source, plausible but unchecked, and exploratory. Anything client-facing below the top tier needs a second pair of eyes before it leaves the agency.
- 04
Re-test the same prompt after any model version change
Behavior shifts with model updates. Anthropic's testing on Claude Fable 5.2 found the model preserves existing code structure when prompts omit ownership framing, so a prompt that worked last month can quietly produce different output this month.
- 05
Record the cost per deliverable against the seat or API spend that produced it
Track tokens or credits consumed per client output so the agency can price AI-assisted work honestly. Forrester's November 2026 research flags that vendors set usage limits and arbitrate disputes over their own tools, which makes internal cost records the only neutral evidence you hold.
- 06
Run a monthly spot check on a random sample of logged deliverables
Pull five records per active client and confirm the reviewer named actually reviewed the work. Logs decay into fiction when nobody audits them.
- 07
Archive the record with the deliverable at handoff and retention close
Store provenance alongside the final file in the client folder, not in a personal drive. When a client leaves or a contract renews, the log is the artifact that shows how the work was made.