Agentic business automation
Not a chatbot. A worker that shows its work.
Aperator agents read what comes in, decide what needs to happen, act through the tools you allow, and stop for a human before anything consequential goes out. Every step is recorded.
- 01Input received500 units of Product A for Customer ABC
- 02find_customerCustomer ABC — found
- 03find_productProduct A — found
- 04check_inventory700 units available
- 05calculate_quote£7,500
- 06Approval requestedsend_email needs a human
- 07send_emailQuote sent
- 08Run completed8.4s · 375 tokens
The distinction that matters
A chatbot answers questions. This finishes the process.
Most “AI agent” products are a chat window with a system prompt. Aperator is built the other way round: business logic and permissions come first, and the model works inside them.
- Tools, not free text
- An agent can only call the tools you've explicitly granted it — find_customer, calculate_quote, send_email. It can't do anything you haven't wired up.
- Policy, not vibes
- Whether an action needs approval is decided by deterministic application code, not the model's judgment. The AI recommends; policy decides.
- A record, not a transcript
- Every input, decision, tool call, and outcome is persisted — inspectable per run, not just logged to a chat window that scrolls away.
How a run actually flows
The same fixed pipeline, every time.
Approval only appears in the loop when policy requires it — a £7,500 quote goes straight through; a £27,000 one waits for a person. Nothing skips the audit trail either way.
Tools & permissions
Each agent gets an explicit, per-tool allowlist. A tool call outside that list is refused before it ever reaches the model's request.
Policy engine
Deterministic application code decides ALLOW / REQUIRE_APPROVAL / DENY for every action — the model recommends, it never has the final say.
Human approval
Consequential actions pause and wait. Approve or reject from one place, with the exact proposed content in front of you, not a summary.
Audit trail
Every run, decision, tool call, and outcome is persisted from the start — not something you turn on after the first incident.
Give an agent a real process to run.
Set up an agent, hand it a test input, and watch the whole run — tool by tool — before it ever touches production.