Key takeaways
Track completion, escalation, reversal and cost together.
Compare against the humans doing the job today.
Raise autonomy in earned steps, and be ready to lower it.
Autonomous resolution is the number everyone quotes, and on its own it can mislead. An agent can close tickets quickly and still leave customers worse off. A useful scorecard looks at quality as well as volume.
Four numbers worth tracking
Start with resolution rate, then add reopen rate, escalation rate and cost per resolved task. Together they show whether work is finished, stays finished and is worth doing this way.
Compare against a human baseline
Metrics mean little without a reference point. Measure the same workflow as handled by your team before launch, then compare like for like. Agents do not need to be perfect. They need to beat the process they replace.
Raise autonomy in steps
Move from suggested actions to supervised actions, then to full autonomy within limits. Each step should be earned by a stable scorecard over several weeks, and it should be easy to step back.
Why deflection is the wrong number
Deflection counts the conversations that never reached a person. It says nothing about whether the problem was solved. A customer who gives up looks exactly like a customer who got an answer, which is how teams end up celebrating a number that hides a failing product.
Replace it with completion: the task reached a verified end state in the system of record. It is harder to measure and much harder to fake.
Segment before you celebrate
One overall autonomy score hides almost everything useful. An agent can finish ninety percent of easy requests and fail on the ones that matter most. Break the number down by task type, customer tier and value at risk.
Task type: which jobs are truly finished alone
Customer tier: where mistakes cost the most
Value at risk: how much money each decision moves
Time of day: whether quality holds when volume spikes
Watch what happens after the task
A task that closes today and reopens tomorrow was never finished. Track repeat contacts, reversals and refunds for at least a week after each completion, and count those as failures in your main number.
This lag is uncomfortable, because it makes early results look worse. It is also the only honest view. Teams that skip it tend to scale an agent quickly, then spend the next quarter cleaning up work that looked done on the dashboard.
Report it the same way every month
Trust comes from consistency. Pick one short report, keep the definitions fixed and send it on the same day each month. Show where the numbers fell as well as where they rose.
After a few cycles the conversation changes. Leaders stop asking whether the agent works and start asking which job it should take on next. That is what real autonomy looks like from the outside.
About the author
Mara Lind
Mara leads product at Falak. Before that she ran operations teams and spent years cleaning up after tools that stopped halfway.
