Agents should not be optimized until the operator knows where the work is actually stuck.
A friction map names where work slows down and whether the constraint is human judgment, system state, policy, trust, timing, or evidence.
Autonomous teams can optimize the easiest step to measure while leaving the expensive constraint untouched.
Operators deploy agents against the bottleneck that actually limits revenue flow instead of the surface area that creates the cleanest demo.
Most autonomous GTM experiments start with a seductive promise: make the slow thing fast.
So the company adds a research agent, a routing agent, a sequencing agent, a forecast agent, or a meeting-summary agent. Output jumps. Dashboards glow. The demo looks great.
Then the operator checks the system and the same queue is still late, the same deals are still stuck, the same approvals are still vague, and the same customers are still waiting. The agent did not remove friction. It created motion around it.
Work waits because nobody trusts the source, the proof is missing, the timestamp is stale, or the claim cannot be traced.
Work waits because the rule is ambiguous, the exception has no owner, or the safe action changes by segment, contract, or promise.
Work waits because another team, customer, or operator does not believe the handoff is complete enough to act on.
Work waits because CRM, billing, support, product, and human memory disagree about what is true right now.
If a worker cannot resolve the bottleneck, it will learn to route around it.
A lead agent cannot verify consent, so it sends a softer message. A forecast agent cannot reconcile support risk, so it marks confidence as medium. A renewal agent cannot find the promise owner, so it offers a generic rescue play. A sales agent cannot prove buyer intent, so it creates a follow-up task for a human.
That looks like progress until operators realize the machine has converted unresolved friction into downstream work.
Lists what the agent can do: scrape, enrich, classify, draft, route, score, summarize, update, and notify. It is capability theater unless it names the constraint.
Lists where work stops, why it stops, who clears it today, what evidence clears it safely, and which part is safe for a machine worker to own.
Name the revenue loop: inbound routing, outbound prioritization, renewal rescue, expansion signal handling, forecast change, customer escalation, or campaign launch.
Identify where the work sits idle: queue, approval, source conflict, unclear policy, missing customer context, owner handoff, or evidence gap.
Classify the delay as evidence, policy, trust, state, timing, consent, authority, risk, or capacity. Do not accept “manual work” as a diagnosis.
Decide whether the agent should collect proof, reconcile state, draft the action, execute inside policy, escalate with context, or stop.
Increase autonomy only when the measured stop point shrinks without increasing cleanup, reversals, customer risk, or human interruption volume.
A good friction map changes the question from “what can agents do?” to “where does the operating system need relief?”
Maybe the answer is not more outbound. It is better consent evidence. Maybe the answer is not faster forecasting. It is source reconciliation. Maybe the answer is not richer account research. It is a handoff contract that makes Sales, Success, and RevOps trust the same object.
The agent that matters is the one that makes the real constraint smaller.
The team produces more records, tasks, or recommendations, but queue age and time-to-resolution stay flat.
The worker completes the visible task, but a person still has to decide whether the evidence is believable enough to act on.
The agent handles the easy path and turns every real constraint into escalation, cleanup, or “needs review.”
Choose a workflow where speed is visible but the business outcome still feels slow or unreliable.
Trace the last ten items and mark exactly where each one waited, who cleared it, and what proof was missing.
Label each delay as evidence, policy, trust, state, timing, consent, authority, risk, or capacity.
Give autonomy to the agent only where it removes the measured stop point without moving burden downstream.
Create one friction map before approving another autonomous GTM workflow.
The blunt rule: do not optimize the agent. Optimize the bottleneck.