Platform
You decide how much the agents may do without asking.
Every agent scores its own work against your master data and rules. Score above the line you set and it acts, posting to your ERP. Score below it and the item goes to a person with the reasoning attached.
The loop
What every agent runs, every time
The same four steps apply whether the item is an invoice, a receipt or a journal. The third one is where autonomy is actually decided.
Capture
Work is pulled from wherever it lives — mailbox, portal, EDI, bank file, ERP queue. Nobody uploads anything.
Understand
The document is read, not templated: header, lines, tax, intent — whatever layout it arrived in.
Validate
Scored against your master data, contracts, tolerances and history. That score is the agent’s own confidence in its work.
Act
Posted to your ERP, reconciled, and the outcome recorded with the reasoning behind it.
The line
One number, set three ways
The confidence threshold is not a global switch. It is set per agent, per entity and per amount band — because a supplier invoice under a thousand dollars and a revenue reclass above a materiality limit do not deserve the same treatment.
Agents clear more on their own. Faster cycle, fewer interruptions, and a larger share of decisions you are trusting the log for.
After a quarter of watching the agents work, this is the line the majority of controllers land on and leave alone.
Almost everything reaches a person. Fewer surprises, more review time, and a throughput ceiling set by your team rather than the agents.
The trade, in numbers
What 95% looks like on a real volume
Modelled on Meridian Supply Co. and its 11,200 invoices a month. Your own mix is what matters, and we model it in the first workshop rather than asking you to accept ours.
Below the line
What an escalation actually carries
An escalated item is not a ticket saying an agent failed. It arrives with the match the agent found, the score it gave itself, the reason the score fell short, and the resolution it proposes. The person decides — approve, edit or reject — and that decision is recorded against the item with their name on it. Nothing is posted quietly, and nothing waits without a reason attached.
The loop closes
Why the same exception should not come back twice
Whatever a person decides joins shared knowledge, so the next item of the same shape comes back already resolved. That is what moves the line over time: not a better model in the abstract, but your resolutions accumulating into the taxonomy. At a top global automaker, fifteen exception types ended up accounting for roughly ninety per cent of everything the agents met — which is a finite, teachable list rather than an open-ended problem.
Evidence
What that produced in twelve weeks
The honest part
Where the line binds
Autonomy starts moderate and climbs as agents absorb your resolutions. Teams that set the threshold too low at go-live spend the first weeks approving work the agents could have handled; teams that set it too high spend them unwinding postings. Neither failure is silent — both are visible in the log, and both are fixed by moving one number. The setting is not a go-live decision you are stuck with.
Questions
Asked before every deployment
What is governed autonomy?
A confidence threshold you set per agent, per entity and per amount band. Above the line an agent acts and posts to your ERP; below it, the item is held for a person with the agent's reasoning attached.
How does an agent decide its own confidence?
It scores its work against your master data, contracts, tolerances and history. That score is what is compared to your threshold, and it is stored with the decision.
What threshold do most teams use?
Around 95%. On a modelled volume of 11,200 invoices a month that leaves agents acting alone on 88% of AP volume, with roughly 1,344 invoices reaching a person each month.
Can we change the threshold after go-live?
Yes, at any time. Most teams start conservative and raise autonomy as the decision log earns it.
What happens to items below the threshold?
They go to a person with the match found, the confidence score, the reason it fell short and a proposed resolution. The decision is recorded against the item and joins shared knowledge, so the same shape returns already resolved.
How long does it take to reach high straight-through rates?
At a top global automaker it took twelve weeks to reach 85% straight-through, with fifteen exception types covering about ninety per cent of what the agents escalated.
See it running before you decide.
The tour is the product, not a video of it. Walk the floor yourself — no form in front of it.