Author an AI agent in plain English. It reasons over your data with real tools, writes you a report, and stages the changes for review — it proposes, you approve. It never touches your data on its own.
Plain English: the goal + your guardrails.
On demand, on a schedule, or on an event.
Reads the entities you granted, step by step.
A report + staged changes — never a direct write.
One click commits; reject and nothing happens.
This is the difference between a useful agent and a scary one. A BackOffice agent can read what you grant it and reason freely — but every change it wants to make is staged as a proposal, not written. You see exactly what it would create or update, with its rationale, and approve or reject.
Grants are enforced on the server at every step and re-checked at approval — the agent can only act inside the entities you gave it, and can never escalate its own scope.
Counts, aggregates and reads records to think it through.
A readable summary of what it found and why.
Changes wait for your approval — nothing auto-commits.
Only the entities you allow, enforced server-side.
Hit run whenever you want a fresh pass — a report on demand.
A daily or weekly cron. The reorder check that runs before you're awake.
Fires when a record is created or updated and matches your condition.
"Compare sales to stock and propose POs for anything running out in two weeks, within budget."
"Find deals with no activity in 14 days and draft a follow-up task for each owner."
"Flag overdue invoices over ₾500 and propose a reminder for each."
"Every Monday, summarise last week's revenue, top products and pipeline changes."
| What you'd otherwise do | Typical cost | BackOffice |
|---|---|---|
| Analyst doing the recurring check by hand | hours/week | ✓ |
| Separate AI-agent platform | $$$/seat/mo | ✓ |
| Wiring it to your data + approvals | Build it | ✓ |
| Agents that know your data, gated by approval | extra tool + time | included · AI metered |
Agents run on the same workspace as your records, so there's nothing to integrate — and every write goes through your approval. See pricing →
No. Agents never write directly — every change is staged as a proposal on the run, and a human approves before anything commits. Read-only agents can't write at all.
In plain English. Give it a mission like "check sales against stock and propose a purchase order within our budget", set guardrails (which entities it may read and write, max steps, a budget note), and pick a trigger.
On demand, on a schedule (a daily or weekly cron), or on an event — when a record is created or updated and matches your condition.
Only the entities you grant it. Read and write scopes are enforced on the server at every tool call and re-checked when you approve — it can only propose within its grant, never escalate.
Agents are part of the BackOffice workspace — free to try, with AI usage metered against your plan. There's no separate per-agent seat.
Write a mission, set the guardrails, and let it draft the work — you stay in control of every change. Free to start, no credit card.
Start free