If an agent's claim can change action, the claim needs a visible mark showing where it came from.
Witness marks attach provenance to autonomous statements before they affect customers, forecasts, campaigns, tickets, handoffs, or management decisions.
Agent organizations reuse each other's outputs. One unsupported claim can travel through the operating system faster than any human can inspect it.
Heads of Growth get higher trust density: more autonomous work per operator because every important claim arrives with enough evidence to verify or dispute quickly.
Human teams are messy, but they carry social provenance. You remember who said the enterprise buyer was active, which CSM warned about procurement, and why a campaign promise was safe. Machine workers do not naturally preserve that social trail. They compress it.
Compression is useful until it becomes authority. A summary turns into a recommendation. A recommendation becomes a task. A task updates CRM, triggers outreach, shifts priority, or changes a forecast. If the original claim has no witness mark, the organization cannot tell whether it is current evidence, stale memory, model inference, or a confident guess.
Call transcript, CRM field, customer email, product event, approved playbook, human note, web page, or another agent's output.
The human, system, agent, workflow, or integration responsible for turning evidence into the claim.
The moment the evidence was captured and the moment the agent converted it into operational guidance.
The account, segment, campaign, team, risk class, or permission boundary where the claim is allowed to influence work.
A witness mark does not prove the agent is right. It proves the organization knows what the agent is standing on. That is the difference between useful autonomy and vibes with API keys.
The mark should be lightweight enough to travel with the work: source links, source class, timestamp, confidence, freshness state, responsible owner, and the action boundary. Operators should not need to reconstruct a forensic chain every time a machine worker says, this account is high priority or this deal moved to commit.
The exact statement the agent wants the organization to trust: risk, priority, status, recommendation, classification, or next action.
The evidence and actor behind the claim, including source links or record IDs where possible.
How strongly the worker believes the claim and whether that confidence came from direct evidence, corroboration, or inference.
What the claim may do: inform a draft, update a record, trigger outreach, change a forecast, escalate a risk, or merely ask for review.
A GTM agent says an account is warming up and should get executive attention. Without a witness mark, the operator sees a confident recommendation and a black-box score. With a witness mark, the claim shows three fresh product events, one champion reply, a funding trigger from 12 days ago, and no suppression conflicts. The Head of Growth can trust the recommendation quickly because the proof travels with it.
A RevOps worker moves an opportunity from best case to commit. Without a witness mark, finance and sales argue with a dashboard. With a witness mark, the update names the call transcript, buyer sentence, stage rule, rep confirmation, and timestamp. If the source is weak, the dispute is fast. If the source is strong, the forecast gets cleaner without another meeting.
How many agent outputs affected work without source, actor, time, confidence, or allowed-use metadata?
How quickly can an operator challenge, confirm, or roll back a machine claim because the witness is visible?
Which marked claims get reused by other agents, workflows, dashboards, or human decisions?
The real risk is not one agent making one questionable claim. The risk is that other workers inherit it. A research agent marks an account as expansion-ready. A campaign agent writes messaging from it. A routing agent creates tasks from it. A manager agent reads the task volume and reallocates labor. By the time a human notices, the original uncertainty has become a multi-system fact.
Witness marks make reuse safer. They let downstream agents downgrade autonomy when a source is weak, refresh when evidence is old, and escalate when a claim crosses into higher-risk territory.
List the claims agents make that influence revenue, customers, priority, ownership, or risk.
Make source, actor, time, confidence, and allowed use mandatory for material claims.
Let strong marks unlock action; weak or missing marks should produce drafts, questions, or escalations.
Review where marked claims travel and which downstream systems reuse them without fresh validation.
Do not ask operators to trust autonomous claims that arrive naked.