If agents quietly absorb broken process, the company does not get leverage. It gets automated duct tape.
Instrumentation that detects work agents had to invent because the official process lacked an owner, rule, field, source, or decision path.
Automation only compounds value when it exposes the constraint. If it hides the constraint, it turns organizational debt into machine-speed habit.
Heads of Growth can deploy agents without letting them normalize CRM chaos, broken handoffs, vague approvals, and missing accountability.
A good autonomous worker is helpful. It fills gaps. It finds the missing account owner. It rewrites the unclear reply. It reconciles fields that should have matched. It nudges the person who should have been tagged three days ago.
That is useful once. It is dangerous when repeated silently.
When agents keep compensating for weak process, leadership loses the signal that the process needs repair. The dashboard shows throughput. The customer gets a response. The queue moves. Underneath, the company has trained a machine worker to carry invisible operational debt.
The agent fills, translates, or guesses missing CRM, ticket, account, campaign, or lifecycle data before it can act.
The worker spends time finding who can approve, answer, unblock, or accept responsibility because the process has no explicit owner.
The agent chooses between conflicting runbooks, stale examples, and local team habits without a governed decision rule.
The worker digs through Slack, docs, tickets, email, notes, and old calls to reconstruct information that should have been structured.
The agent keeps moving edge cases forward without forcing the organization to decide whether the exception deserves a new policy.
Agent dashboards love completed work: replies sent, records updated, leads enriched, renewals prepared, tickets triaged, tasks closed.
But completion is not the same as health. If every completed item required hidden cleanup, the team is not becoming more efficient. It is outsourcing confusion to a worker that does not complain.
That is the trap: humans used to feel the friction. Agents remove the feeling before anyone fixes the source.
Classify the shadow work: missing data, unclear owner, contradictory policy, stale source, duplicated system, approval gap, or exception.
Track how many extra steps, lookups, retries, clarifications, and escalations the agent needed before completing the task.
Assign every repeated shadow-work pattern to the human owner who can change the workflow, not just the agent prompt.
Define when a workaround graduates into redesign: repeated three times, high-value account affected, revenue risk, compliance risk, or customer delay.
Convert the pattern into a changed field, updated rule, owner map, queue policy, runbook patch, or removed step.
A growth team uses an agent to prepare expansion plays. It reads product usage, CRM notes, support tickets, renewal dates, and campaign history. The agent works well, but every play requires hidden repair: account names differ between systems, owners are stale, usage thresholds conflict with the current packaging, and the approval rule for discount language is buried in an old Slack thread.
The team sees fast expansion briefs. The agent sees chaos. Without shadow-work sensors, the chaos never reaches the operating review.
With sensors, the agent logs each repair: stale owner, conflicting threshold, missing packaging rule, manual evidence lookup. After two weeks, the team can see that the bottleneck is not content generation. It is revenue data hygiene and approval ownership.
Average number of cleanup steps an agent performs before a workflow can be completed.
How often the same missing owner, field, rule, or source appears across autonomous work.
The percentage of repeated shadow-work patterns converted into actual process changes within a fixed window.
Make agents label every extra action they had to take: lookup, guess, reconcile, escalate, reroute, rewrite, or override.
Report completed work and hidden repair work as different metrics. Do not let one hide inside the other.
When the same patch repeats, create an owner-backed process fix instead of adding more prompt instructions.
Use the sensor output to remove fields, merge sources, clarify owners, and kill approval paths that only create waiting.
Before scaling autonomous work, make hidden cleanup visible.
The blunt rule: if agents keep saving a broken process, they are also preserving it.