Same tool surface
A custom agent reads and proposes through the same typed interfaces as the built-in ones. It cannot reach anything the AP agent could not, which is what bounds its blast radius.
AI agents · finance
Every business has one process that no vendor builds for — a reconciliation against a bespoke settlement file, an allocation nobody else needs, a compliance check specific to your industry. Custom agents run on the same tools, the same policy engine, and the same audit schema, which is what stops a bespoke agent from becoming an exception to your controls.
Post a coded vendor bill automatically when all of these hold:
What it does
Each of these is work a person does today. The agent does them in sequence and stops at the first thing it is not confident about.
A custom agent reads and proposes through the same typed interfaces as the built-in ones. It cannot reach anything the AP agent could not, which is what bounds its blast radius.
Its proposals are evaluated by the same deterministic engine against policies your controller authors. There is no bespoke enforcement path to review separately.
Levels 0 to 4, granted by an administrator, with the same four actions absent entirely. A custom agent cannot release funds or close a period either.
Model, context, sources, reasoning, confidence, policy, approver, resulting entry. Your auditor sees one schema, not one per agent.
Straight-through rate and confidently-wrong rate measured against your corrections, with the same regression gates blocking a change that makes it worse.
Built and run by us as a module rather than handed over as code you then own. Build fee plus monthly hosting, versioned like everything else.
Authority
Every custom agent starts here regardless of how mechanical the workflow looks. The first quarter is where the corpus and your confidence are built.
Only once accuracy is measured and a policy simulated against history. The same bar as a built-in agent, not a lower one because you commissioned it.
We will not build an agent that bypasses the policy engine, writes to the ledger directly, or logs to a separate schema — regardless of what it would cost you to ask for.
The workflows that work well as custom agents share a shape: high volume, repetitive, document or data driven, with a decision that can be checked against a rule or a history. The ones that do not work involve genuine judgement, low volume, or a decision nobody can articulate the basis for.
Good candidates we have built: reconciling a payment processor settlement file with a bespoke format, checking subcontractor compliance documents against expiry rules, allocating a shared cost across entities using a formula specific to a partnership agreement, and validating a royalty calculation against contract terms.
The obvious alternative is to build it and give you the code. We do not, and the reason is not commercial lock-in — it is that an agent handed over stops being governed. It falls outside the evaluation harness, the regression gates, and the audit schema, and a year later nobody knows whether it still behaves.
Hosting means it is versioned, monitored, and measured like the built-in agents, and when a model changes it is re-evaluated against your corpus rather than quietly drifting. You pay a monthly fee for that, and you can export everything it has produced at any point.
Build is quoted at the published rate — API cost times five with a $5,000 floor — after a scoping session. Hosting runs $250 to $1,000 a month depending on volume and complexity. Most custom agents land between $8,000 and $25,000 to build.
We will also tell you when the answer is not an agent. A workflow running twice a month is almost never worth automating, and a rule that never changes is a policy rather than an agent — cheaper, more predictable, and it does not need an evaluation harness.
Questions
We will tell you whether an agent fits, what it costs, and when a simple rule would serve you better.