AI agents

Agents that own a workflow, not a text box

Each one has a scope, an authority level you set, a policy engine it cannot argue with, and a record of everything it did. They start work without being asked and hand you the exceptions instead of the busywork.

accounts-payable-agentrunning
Bill arrives [email protected]
Extracted vendor · date · lines
Coded 6420 · Cloud hosting
Matched to PO PO-2291 · within 2%
Policy checked level 2 · under $2,500
Approved M. Reyes · controller
Posted to the ledger JE-88104 · period open
Journal entry JE-8810419 Aug
6420 · Cloud hosting4,180.00
2010 · Accounts payable4,180.00
Balanced4,180.004,180.00
Work with your existing ledgerAuthority set per agentEvery action audit-logged

The agents

Eight roles, live today.

What makes these agents rather than features

The word has been stretched to cover almost anything with a model behind it, so it is worth being specific about what we mean. Three properties, and a thing has to have all three.

  • It initiates. The work starts without a person opening a screen. A bill arrives, a bank feed updates, a contract is signed, a period opens — the agent picks it up.
  • It owns an outcome, not a step. The AP agent is not a document parser. It is responsible for a bill reaching a correct posted entry, which includes deciding when it cannot do that alone.
  • It operates under explicit authority. It has a permission level, it is constrained by a policy engine it cannot address, and everything it does is recorded. Without this a thing that initiates work is not an agent, it is a liability.
A copilot waits to be asked. An agent notices. The difference is who is responsible for remembering the work exists.

They share one context

This is the part that is hard to see in a demo and matters most in production. Because every agent reads the same business graph, the AR agent knows a customer is mid-renewal before it sends a dunning notice, the Close agent knows the AP agent has forty bills still in the exception queue, and the CFO agent explains a margin movement using the project data the delivery team entered rather than a separate export.

In a stack of point solutions each of those is an integration project. Here it is the default, because there is nothing to integrate — and it is the single biggest reason a fleet of agents behaves sensibly rather than each one doing something locally reasonable and collectively wrong.

How customers actually roll these out

Almost nobody should turn on eight agents in a month. The pattern that works, and the one we recommend in the first call, runs roughly like this.

  • Weeks 1–6: accounts payable at Level 1. Everything drafted, everything reviewed. This is where the labelled corpus comes from and where your team calibrates its trust.
  • Weeks 6–10: AP to Level 2 for recurring vendors under a threshold. Typically once a vendor has thirty clean drafts. Dunning moves to Level 2 around the same time.
  • Quarter 2: reconciliation and close. Reconciliation at Level 2 daily, close at Level 1 driving the checklist. This is where close duration starts moving.
  • Quarter 3 onward: the analytical agents. CFO and Controller at Level 0 and 1, which need three or four closed periods of clean data behind them to be worth anything.

Procurement and revenue recognition tend to come later and are more industry-dependent — rev rec early for software companies, procurement early for distributors.

One agent at a time is not caution, it is method

Each agent changes a workflow your team already has habits around. Turning on several at once means that when something feels wrong, nobody can tell you which change caused it. Staging them is how you keep the ability to attribute a problem.

Questions

What people ask first.

Do the agents need us to replace our ledger?
No. They read and write through the business graph, so they work whether your books sit in QuickBooks, NetSuite, Sage Intacct, Acumatica, Dynamics, Odoo, or Xero. Most customers run agents for months before any ledger conversation.
How many can we run at once?
Technically all of them. Practically we recommend one workflow at a time so that when something behaves unexpectedly you can attribute it. Most customers are running two or three by the end of the first quarter.
Can we build our own?
Yes. Custom agents are built on the same tools, the same policy engine, and the same audit schema — so a bespoke agent is governed identically to a built-in one rather than being an exception to your controls.
What do they cost to run?
Each tier includes a monthly allowance of agent actions with metered overage. An agent reading and searching your data does not consume the allowance; completing a task does.
What happens when one is unsure?
It stops and routes the item to an exception queue with the reason stated. That queue is the agent telling you where it needs help, and its size is a metric we report rather than hide.

Pick one workflow and prove it.

Fifty documents and a week is enough to see what an agent would have done on your own data.