Automate the routine. Escalate the consequential.
Operations teams face a false choice: review everything, and the automation achieves nothing, or review nothing, and one bad action undoes a year of savings. The real work is deciding where the line goes.
Where is the line?
Autonomy should follow consequence.
- Low consequence
Read and prepare
Gather, compare, draft
Runs unattended
- Low to medium consequence
Reversible, low value
Update a ticket, send a reminder
Runs unattended, with limits
- Medium consequence
Reversible, higher value
Reassign work, change a schedule
Policy decides
- High consequence
Hard to reverse
Pay, delete, grant access
Requires a person
Example patterns
What this looks like, by function.
Reconcile vendor invoices and prepare payments.
Runs unattended, within limits
Retrieve records, match POs, flag anomalies, schedule low-value payments to verified vendors.
Payments above a limit, and any change to bank details.
Intended value: Finance staff review exceptions, with evidence attached, rather than every line.
Human in the loop
A good approval takes seconds.
Approvers do not want to read an essay from a model. They want the facts, the reason they were asked and a clear decision.
The agent stops, cleanly
The step parks and holds no compute. It could wait for days.
The right person is asked
Policy decides who, and the requester cannot approve their own work.
They see the facts
Exact parameters and evidence from records come first, and a model's summary is labeled as one.
The decision binds
Approval applies to this action only. If anything changes, it no longer applies.
Work resumes
The step continues, and the whole exchange is in the record.
Which process would you try first?
We would like to hear what you would automate, and what you never would.