Governance · 15 July 2026

What a position description for an agent actually contains

The document that turns an AI capability into an employable role, section by section, with the questions your risk team will ask about each one.

When we say every agent starts with a position description, people picture a marketing page. It is not one. It is a working document, agreed and signed before the agent is deployed, and it is compiled into the constraints the kernel enforces at runtime. This briefing walks through its sections.

The role, in one sentence

The first line states what the role is for, in language the whole firm can repeat. “Keeps the debtor ledger current and follows up what is owed” is a role. “Leverages AI to transform finance” is not. If the sentence cannot be falsified, it cannot be supervised.

Duties

Five to eight duties, each one observable. A duty must describe work you could point at in the audit trail: a reminder sent, a register updated, a summary delivered. Where a duty involves judgement, the duty says whose judgement finishes the job. “Drafts variance commentary for finance lead review” keeps the noun and the reviewer in the same sentence on purpose.

Reporting line

Every agent reports to a named human role, and the position description says which decisions travel up that line. This is the section risk committees read first, and it should answer their question directly: when this agent meets something outside its remit, who hears about it, and how quickly?

KPIs

Three or four, measurable from the decision log without extra instrumentation. Good agent KPIs look like good staff KPIs: debtor days, turnaround hours, error counts with a target of zero. If a KPI cannot be computed from the record, it is an aspiration, and it comes out.

The never list

The section that does the most work in a deal, and the one we refuse to soften. It lists what the agent will never do, and the kernel enforces each line as an absent capability, not a policy reminder. “Never alters an invoice amount” means the operation does not exist inside that agent’s remit. When a prospective client asks what happens if the agent is instructed to do it anyway, the answer is: it refuses, and the refusal is logged with the instruction that caused it.

Sign-off gates

Finally, the schedule of actions that stop and wait for a person: which actions, which approver roles, and what the approver sees at the gate. Gates are the difference between delegation and abdication, and the position description is where they are agreed rather than assumed.

A useful test for any agent vendor, ours included: ask for the position description before the demo. If there is no document to show you, the governance is a slide.

Book a demo

Tell us the role you would otherwise hire for. We will show you an agent at work on it in a sample firm: hand it a task and watch the sign-off gate hold its first draft. No deck, no obligation.

KERNEL v1.0 · NO DECK, NO OBLIGATION

THIRTY MINUTES · LIVE ON ONE OF YOUR WORKFLOWS

Book a demo

Tell us the role you would otherwise hire for. We put an agent to work on one of your own workflows, live, and you leave with a written scope.

KERNEL v1.0 · RESPONSE WITHIN ONE BUSINESS DAY

We use what you enter here to reply to you and for nothing else. Data is handled under the Australian Privacy Principles.