The AI employee platform: pick a role, brief it, keep the last word.
A platform of role-based AI employees for small business — sales, support, accounting, HR and marketing. You choose the role and brief it in plain language. It works from your material, hands you drafts, and asks before it spends or sends.
From brief to reviewed work, in four steps
Brief
Describe the job in plain language: tasks, tone, rules and the outcome you want. That brief is the material every run reads from.
Configure
Pick a role, point it at the source of work, and set the rules it runs under — which actions need a human yes, and what the spend cap is.
Run
The role runs on a schedule. What it produces — a reply, an ad set, a proposal — lands as a draft with the reasoning attached.
Review
Every run is written to a journal with its inputs, its outcome and its error text. A run that failed stays visible as a run that failed.
How does a Unistaff digital employee work?
The reasoning layer
A general frontier model does the thinking. It never talks to your systems directly: it answers into a single JSON object, and prose that fails to parse does not leave the runner.
The role
A role is a prompt and a validator kept together in code. The validator throws out a result that breaks the role's rules — a number without a named source, an ad past Meta's length caps, a checklist item marked green with no reason written next to it.
Your material
Your brand pack — services, positioning, prices, tone — is read into the prompt on every run. When an input is missing, the role says so in its own output instead of filling the gap with an invention.
What is in the box
Every digital employee arrives with the parts a real hire needs: a defined job, your material to work from, limits it cannot talk its way past, and a paper trail.
Roles as code
Five roles ship inside the platform today: market analyst, creative writer, campaign proposer, lead responder, brief composer. Each is a prompt with a validator, and a renamed role breaks the build instead of going quietly missing.
Your material, on every run
Brief, service list and knowledge items are stored per brand and read into the prompt each time, so two runs a month apart answer from the same facts.
Connections
One live platform connector today: Meta ads, read access for reporting, with writes behind a separate token and the approval gates. Plus a lead endpoint any form or quiz can post to, and an email to you the moment a lead comes through your own site form.
Human approvals
An action in the approval class waits for a person. An action the rulebook does not recognise counts as needing approval, so a new capability starts life waiting for you.
The journal
Runs, sync attempts, proposed changes and decisions are written with timestamps and outcome, failures included. A decision carries who decided, when, and the note they left.
What is the difference between an AI employee and a plain AI agent?
| Feature | Plain AI agent | Unistaff |
|---|---|---|
| Role | Blank model | Prompt plus a validator, kept in code |
| Setup | Prompt + tool wiring | Pick a role, brief it in plain language, set its limits |
| Material | No context | Brand pack and knowledge items read on every run |
| Limits | None | Spend cap, kill switch and approval rules stored as records |
| Output | Sends and hopes | Drafts, with a person on the send |
| Trace | Chat scrollback | Journaled runs and decisions, failures included |
Oversight · six gates
What the human does
Reading, analysing, drafting and journaling run unattended. Money, outbound messages and anything the rulebook has never seen stop and wait for a person. Each line below is a gate in the code, labelled with the string it throws.
01 · Send
The first reply to a lead is a draft until a person sends it
The lead responder writes the reply ahead of time, so the answer is ready while the lead is still warm. It has no send path of its own. The phone number, the email address and the full name never enter the prompt either — the role works from an initial and the answers on the form.
02 · Approve
An action in the approval class waits for a human decision
Before any outbound write, the action is checked against the autonomy matrix. Launching a campaign, changing a budget and pausing something all sit in the approval class. An action the matrix has never seen counts as requiring approval, so the failure mode is waiting rather than acting.
03 · Cap
Spending is compared against your cap before the money moves
Daily and monthly caps are records in the database, checked inside the same gate. A mutation whose daily spend goes over the cap is refused at the gate, which is a different thing from noticing it on the invoice.
04 · Halt
The kill switch is checked before every outbound write
There are three switches — everything, ads, AI — and engaging one takes a written reason that is stored with it, because a halt nobody can explain later cannot be audited. A blocked action keeps its approval and waits instead of dying.
05 · Decide
A deadline never turns into a yes
A pending approval that runs out of time stays undecided. Expiry is never read as consent, and an item that has been decided cannot be quietly decided a second time. Each decision keeps who decided it, when, and the note they left.
06 · Fail shut
If the rulebook cannot be read, nothing goes out
The caps, the switches and the approval rules live in the database. When it cannot be reached, the outbound write is refused, because a write with unread checks is a write with no checks.
The same discipline applies to the way work is stored and who can read it.
What roles can I put to work?
Roles are scoped per deployment. Five of them exist as code in the platform today — market analyst, creative writer, campaign proposer, lead responder, brief composer. The ones below are the jobs we scope a pilot around, and we build them under the same draft-and-approve loop.
AI Sales Assistant
Reads every inbound lead and has the first reply written before a rep opens it.
Learn more →AI HR & Recruiting Agent
Scores applicants against your published criteria and drafts the scheduling thread.
Learn more →AI Customer-Support Agent
Drafts the answer to a repeat question from your own docs, and routes the rest to a person.
Learn more →AI Employee for E-commerce
Order status, returns and cart follow-up, drafted for a person to send.
Learn more →AI Employees for Agencies
Take the coordination layer off account managers and keep the client relationship human.
Learn more →HTTPS only
TLS 1.2+, HSTS on every response
Encrypted at rest
AES-256, Neon platform default
Approval gate
Unrecognised actions wait by default
Kill switch
Checked before every outbound write
Action journal
Runs and decisions, failures included
Each line above names a mechanism, and the mechanism is described on the security page together with the controls this platform has yet to build.
Frequently asked questions
Yes. You pick a role and write the instructions in plain language — the job, the tone, the rules, what a good outcome looks like. That brief is what every run reads from.
It reads data, analyses it, drafts, and writes every run to the journal. It does not send a message and does not spend money on its own: those sit behind the approval gate, and an action the rulebook does not recognise waits for you by default.
Meta ads for reporting, with writes behind a separate token and the approval gates. A lead endpoint any form or quiz can post to, with an email to you when a lead comes through your own site form. A daily capture of the public Meta Ad Library for competitor ads. Anything else is a connector we scope with you before we promise it.
A chatbot answers from a generic model and sends what it produced. A digital employee runs a role with a validator that can reject its own output, works from your brand material, and hands the result to a person before it reaches a customer or a budget.