Observe
The agent watches and reports. It writes nothing. This is where every agent starts and where several stay for the first few weeks while you read what it would have done.
AI capability
Autonomy is not a slider from off to on. It is a specific grant: this workflow, in this entity, below this amount, when confidence exceeds this threshold, with this escalation when it does not. Every agent in the system starts at observe and moves only when you move it.
What it does
The agent watches and reports. It writes nothing. This is where every agent starts and where several stay for the first few weeks while you read what it would have done.
It prepares work — coded bills, matched payments, drafted journals — that sits in a queue until a person releases it. Nothing reaches the ledger without that release.
It puts a specific recommendation in front of a named approver with the reasoning attached, and the approver accepts, edits, or rejects. The rejection is training data.
It acts without a person, inside the boundary you set — this workflow, this entity, under this amount, above this confidence. Outside any of those, it escalates rather than proceeding.
A fixed set of actions no agent may take at any configuration: releasing payment, changing vendor banking, granting permissions, closing a period, or filing anything statutory.
Authority is raised by a person, recorded with who raised it and when, and can be lowered instantly. It does not drift upward on the basis of good performance.
The instinct is to describe an agent as more or less autonomous. It does not survive contact with a real finance function, because the same agent should be trusted very differently depending on what it is doing and where.
Coding a recurring $400 utility bill from a vendor with two hundred prior identical bills is a different act from coding a $60,000 invoice from a new supplier with an ambiguous description, and both are the accounts payable agent. Authority granted per workflow, per entity, and per threshold is the minimum granularity that matches how the work actually differs.
An agent’s first weeks are spent producing a shadow record: what it would have done, beside what your team actually did. That comparison is the only honest basis for deciding whether to raise its authority, and it costs you nothing to run.
It is also where the surprises surface. The disagreement rate is rarely uniform — it concentrates in one vendor, one account, or one entity, and that concentration usually points at a data or policy problem rather than a model problem.
Every automated decision carries a calibrated confidence, and the threshold is a grant condition rather than a badge on a screen. Below it, the agent escalates with its reasoning and its alternatives; above it, it proceeds within the other limits.
Calibration is checked against outcomes rather than asserted. An agent claiming 95% confidence should be right about 95% of the time, and where it is not, the threshold moves rather than the claim.
Raising authority is deliberate and recorded. Lowering it is immediate and needs no justification — a single control that returns an agent to draft, applied at the workflow, entity, or agent level.
The asymmetry is intentional. A control you can only exercise through a process is a control you will not exercise at the moment you most want to.
Limits
Every capability page on this site carries one of these, because a feature described without its boundaries is a claim rather than a description.
Payment release, vendor banking, permission grants, period close, and statutory filing stay at never regardless of what a customer would prefer. That is the one decision we take away.
An agent executing within limits on poor-quality inputs escalates constantly, which is worse than leaving it at draft. We will say when the data is not ready.
If your organisation has not decided who authorises what above a given amount, the authority model has nothing to encode. That decision precedes configuration.
Questions
Tell us the workflow. We will tell you what authority it needs and where the ceiling should sit.