Autonomous agents: how it works and what to require
Proconomy runs the rules your administrators configure at the moment a record changes rather than when somebody reaches it. A clean invoice inside your amount limit is approved and dated with no person involved; one that fails any of eight checks is flagged with the check that failed. Model features annotate, draft and analyse — and none of them publishes, sends or awards anything.
Most of the queue was never a decision.
What runs when a record changes, and what waits for a person.
When an invoice arrives
- Check it against the amount limit your administrators set
- Match it to what is still open on the purchase order
- Confirm a goods receipt exists and quality is approved
- Check for contract variance, duplicates and tax mismatches
- Approve and date it where all eight checks pass
- Flag it with the failed check where one does not
When a request is submitted
- Suggest the category it looks like
- Compare it against the spend pattern
- Say whether a contract already covers it
- Give the supplier risk and budget position
- Point at possible duplicates already in flight
- Suggest an approver, and read an attached invoice
When somebody asks
- Draft a sourcing event below the value ceiling you set
- Propose the suppliers to invite to it
- Capture a request in conversation, slot by slot
- Mark each captured value as stated, inferred, defaulted or confirmed
- Raise the purchase order on an approved catalogue line
- Keep the requester’s own purchase order limit in force on it
When you open an analysis
- Commodity risk across what you buy
- Component lifecycle position
- Material planning
- Sourcing and rationalisation views
- Supply risk
- Extraction from a contract document
Automation a security team can sign. Because the boundary is a number.
Six properties, and the last two are what make the first four safe to switch on.
| Mechanism | What it has to do |
|---|---|
| It runs on arrival, not on attention | The checks run when an invoice is created, re-matched, arrives electronically or gets a goods receipt — rather than when somebody reaches it in a queue. |
| The boundary is a number you chose | An amount limit on automatic invoice approval, a value ceiling on drafted sourcing events, and automatic-approval ranges on your own approval rules. Anything outside them goes to a person. |
| One failed check is enough | Eight conditions have to hold for an invoice to approve itself. Any single failure stops it and flags the invoice with the check that failed. |
| Advice that changes nothing | The eight notes written onto a submitted request annotate it and alter no field on it. The decision stays with the person who has to make it. |
| Nothing model-written leaves on its own | Drafted sourcing events, captured requests and analysis views all wait for a person to publish, send or award. That is a design decision rather than a setting. |
| The reason is stored when the decision is made | An automatic approval records the limit that applied and how the invoice matched; a flagged invoice records the check that failed; an intake note records its reasoning. |
A description of what the practice requires, not a feature list.
Same boundary. Different places it pays.
Where automatic approval is worth the most, by the shape of the buying.
High invoice volume
The more of the ledger is routine, the more of it stops queueing behind the part that is not.
Heavy catalogue buying
Approved catalogue lines raise their own purchase orders, inside the requester’s existing limit.
Many small requests
Eight advisory notes on each one mean a buyer opens a request that already explains itself.
Tight audit expectations
Every automatic decision records why it went the way it did, at the time it was made.
Five things to ask for. The first one settles most of it.
Each takes minutes, and none of them can be staged in advance.
- Ask where the automation stops, and expect a number rather than a paragraph. If the answer is a list of principles, there is no boundary.
- Set a low amount limit in front of you, send a clean invoice through it, and watch it approve itself.
- Send one that breaks a check, and read what the flagged invoice says about why.
- Ask which parts involve a model and which are deterministic rules. A vendor who will not separate the two has not thought about what happens when the model is wrong.
- Ask what the platform can send to a supplier without a person. The answer you want is nothing.
Definitions. Asked and answered.
They are the parts of the platform that carry out a step without waiting for a person: approving a clean invoice inside the amount limit your administrators set, raising the purchase order on an approved catalogue line, and annotating a submitted request before a buyer reaches it. Each runs when a record changes, each stops at a limit you configure, and anything outside it goes to a named person.
Mostly rules, and we would rather say so. Invoice approval is eight deterministic checks — no model is involved. Models are used for the advisory notes on a request, for drafting a sourcing event or capturing a request in conversation, and for the analysis views. None of those changes a record on its own.
The limits your administrators set, and the fact that there are only a few automatic paths rather than a general permission to act. Invoice approval has an amount limit and seven further conditions, any one of which stops it. Sourcing drafts have a value ceiling. And nothing model-written publishes, sends or awards anything.
No. Nothing the platform drafts is published, sent or awarded without a person taking that action. It is a design decision rather than a setting.
Yes. Automatic invoice approval is the usual starting point because the boundary is a number you choose: set it low, watch what gets approved and what gets flagged, and raise it when you are satisfied. The drafting and analysis features are switched on separately.
In practice it means a piece of the system that carries out a step without waiting for a person — approving an invoice that passes every check and sits inside a configured limit, for example. The useful questions are narrow ones: which steps exactly, what stops each of them, and what the system is allowed to send on its own.
Often it is not, and the distinction matters less than vendors suggest. What matters is whether the step runs without a person, what bounds it, and whether a record is written explaining why it went the way it did. A deterministic rule that executes reliably inside a limit you set is a better thing to deploy than a model that exercises judgement you cannot inspect.
It depends entirely on where it stops and what it checks. An approval bounded by an amount limit you set, conditional on matching the order, a goods receipt, approved quality and no variance, and recorded with the limit that applied, is a narrow and inspectable thing. An approval that cannot tell you why it approved is not, whatever it is called.