What an AI employee is
A line of work, not a chat: it reads set areas of company data, prepares set kinds of drafts, has a daily budget and named people who may accept its proposals. A proposal changes nothing until a person decides; the code has no route for it to carry itself out.
Fifteen lines of work
The list is closed: no setting adds a sixteenth.
More detail
- Sales — qualifies requests, drafts replies and offers. A salesperson closes the deal and sends the e-mail.
- Support — drafts answers from the confirmed knowledge base; a staff member sends them. The website chat is a separate tool without a model.
- Marketing — collects campaign results from exports and flags deviations. It changes nothing in ad accounts.
- HR — drafts vacancy descriptions and hiring-stage summaries. It makes no decisions about people.
- Finance — compares plan with actuals before an invoice is approved. It moves no money.
- Legal assistant — extracts deadlines and obligations from contracts for a person to check. No legal opinions.
- Operations — runs requests by the procedure, watches deadlines, records deviations. People carry out the steps.
- Customer success — proposes an account status from a closed list of words. It computes no score.
- Quality — reviews conversations against a checklist, rating the answer, not its author.
- Procurement — puts requests and supplier terms into one table, unranked. The buyer chooses.
- Project management — collects tasks, sets deadlines, shows what is overdue; the database rejects cycles in the plan.
- Compliance — checks actions against written rules and shows missing confirmations. It does not certify legal compliance.
- IT helpdesk — takes and routes staff requests. It grants no access.
- Knowledge — finds outdated and conflicting documents and proposes fixes; a second person confirms publication.
- BI analyst — explains changes in metrics and names the source of every number. No forecasts.
How an agent is switched on
Every agent ships switched off with a zero budget.
More detail
It first runs in a test mode, where everything it produces stays in a sandbox; moving it into work takes a decision by two people. Budget and go-live steps require the password again and are written to the journal.
An agent whose module is off does not start. A run that would exceed the spend ceiling is refused in advance, with the reason recorded.
What the agents run on
By default, built-in rules without an external model: they select records by matching words, write no prose and refuse with a stated reason what they cannot do.
More detail
An external model is switched on by your decision: a key, the owner’s permission and its host on the outbound allow-list. Live runs on test data passed in September 2026; its quality on your data is measured in the pilot.
Basis: the AI employee registry; the registry of modules and industry templates. This page describes what exists in this build, not what is planned.