Autonomous workers should treat evidence like inventory: useful when fresh, dangerous when expired.
Evidence TTLs define how long customer signals, approvals, policy snapshots, forecasts, segments, and task assumptions remain usable before an autonomous worker must refresh them.
Machine workers can move faster than the business updates context, so old facts quietly become operating instructions unless the control plane gives evidence an expiration date.
Heads of Growth get agent speed with fresher judgment, fewer silent context failures, and a visible path for when workers must pause, refresh, or escalate.
Human operators carry freshness in their heads. A Head of Growth remembers that the enterprise segment was rebuilt yesterday, that a pricing exception only applied to one quarter, that a board forecast was a temporary scenario, and that a customer signal came before the champion left. Autonomous workers do not inherit that ambient memory. They inherit records.
That is why stale evidence becomes one of the first invisible failure modes in agent ops. The worker is not hallucinating. It is obeying an old truth with new confidence.
A page visit, reply, meeting note, or product event should not carry the same weight forever. Intent needs a shelf life.
A human approval to change copy, route leads, or contact accounts should expire when scope, owner, risk, or campaign context changes.
Pipeline stages, close dates, next steps, and attribution become dangerous when workers optimize from last week's operating picture.
Playbooks, compliance language, escalation rules, and qualification thresholds need version-aware freshness before agents enforce them.
Review asks a human to catch every stale assumption. Expiration makes the stale assumption visible before the worker acts. Evidence TTLs turn freshness into an explicit control: this signal is valid for 48 hours, this approval for one campaign, this forecast for one board cycle, this policy until the next version, this owner assignment until territory sync.
That gives machine workers a simple operating rule: if the evidence is fresh, proceed within scope; if it is aging, reduce autonomy; if it is expired, refresh or escalate.
What fact, signal, approval, policy, customer state, or revenue record justified the agent's next action?
How long is that evidence allowed to influence autonomous work without refresh?
Who owns the evidence source when a worker needs a newer version or disputes the current one?
What should the agent do when the evidence expires: pause, downgrade, request refresh, route to a human, or use a safer default?
A GTM agent builds a target list from firmographic fit, recent funding, hiring signals, and product intent. Without TTLs, it keeps contacting accounts after the buying trigger has passed, the ICP changed, or the suppression list updated. With TTLs, funding data might last 30 days, intent seven days, suppression checks 24 hours, and messaging approval one campaign. The worker still moves fast, but it cannot pretend a stale trigger is current urgency.
A RevOps worker updates pipeline stages using CRM notes, call summaries, email replies, and historical conversion rules. Without TTLs, it can make the dashboard look cleaner while baking old assumptions into the forecast. With TTLs, opportunity stage evidence expires after a defined window, call summaries lose authority after a newer meeting, and historical conversion rules require version tags before changing forecast logic.
How often did a worker attempt to act from evidence past its usable life?
How long does it take the system to replace old context with current context?
How often does stale evidence correctly move a worker from write access to draft, ask, or pause?
Put TTLs on the inputs that create business consequences. Customer intent, compliance copy, segment membership, pricing exceptions, owner assignment, forecast assumptions, approval receipts, suppression lists, and routing rules all need freshness windows. Then teach the agent control plane to change behavior when the window closes.
The mature version is not a dashboard full of red warnings. It is a worker that naturally says: I can draft this from current context, but I cannot send it until intent and suppression are refreshed.
List the facts agents use before they change customer, pipeline, campaign, or task state.
Give each evidence class a default lifespan tied to risk and volatility.
Connect freshness state to permission: proceed, draft, ask, refresh, downgrade, or stop.
Review expired-action attempts weekly; they reveal where your operating context changes faster than your agents know.
Before giving an autonomous worker more reach, give its evidence an expiration date.