Human-in-the-loop approval belongs on irreversible and high-consequence actions, not on every step. An agent that asks before reading a file is useless; one that sends, pays or files without asking is dangerous. BiltIQ applies one rule in two products: Agent OS pauses consequential actions for a named reviewer, and Business OS freezes the 23 of its 66 tools that change data.
At a glance
| Point | Detail |
|---|---|
| The rule | Approval for irreversible and high-consequence actions; everything else runs |
| Agent OS | Findings land in an approval inbox; a named reviewer accepts or rejects; with no approver the action is blocked |
| Business OS | 23 of 66 copilot tools change data; each is frozen with a preview and expires after 72 hours |
| Permissions | Role permissions closed by default; bad inputs rejected before review |
| Trail | Business OS: append-only action log of every copilot call |
| Loops | Budgets stop runaway workflows |
Why not approve everything?
If a manager has to click fifty times for an agent to read, filter and summarise a folder, the agent costs more than the clerk. Reviewers stop reading what they approve, and the gate becomes theatre. The compensating-control trap, covered in a live BiltIQ post, is the same failure at the policy level: a control that exists on paper and not in practice.
So the question is not "should a human be in the loop" but "on which actions".
Which actions should pause?
| Runs on its own | Pauses for a person |
|---|---|
| Reading and parsing documents, including scans | Sending anything to a customer, supplier or regulator |
| Searching your own files and citing the page | Creating or changing a record that others act on |
| Scoring, deduping, assigning an owner | Posting to the ledger; preparing a statutory export someone will file |
| Drafting a reply, a finding, a summary | Accepting a finding that changes a claim, a tender score or a contract position |
| Calculating and comparing | Any call to an outside host |
The test is consequence and reversibility. A draft can be deleted. A sent email, a paid claim, a filed return and a changed ledger cannot.
How do Agent OS and Business OS apply the rule?
Agent OS. Agents read, check and cite. Each finding goes to the approval inbox, where a named reviewer accepts or rejects it. Irreversible and high-consequence actions pause, verified in code; if no approver is available, they are blocked rather than run. Budgets stop runaway loops, so a workflow cannot spend its way through a problem.
Business OS. The copilot has 66 tools; the 23 that change data (leads 9, RFQs 6, work items 4, revenue 2, platform 1, insights 1) follow four steps:
- Propose. The copilot calls a tool that changes data.
- Check. Role permissions apply, closed by default; bad inputs are rejected.
- Wait. The call is frozen with a preview in the approval inbox. It expires in 72 hours.
- Decide. Approve, and the frozen call runs. Decline, and the copilot is told.
Every copilot call goes to an append-only action log.
What makes an approval inbox usable?
Three rules keep reviewers reading:
- Show the trigger, not the file. "Invoice exceeds the PO by ₹14,500, page 3" beats a 20-page attachment.
- Show the change as a preview. The reviewer sees exactly what will run if they approve.
- Let the reviewer decline with a reason. The agent is told, and the decision is on the record.
A gate that expires (72 hours in Business OS) stops a queue from silently growing into a backlog nobody owns.
Why is a rule in the prompt not enough?
A prompt rule is a request to the model. The control has to sit under it. In Agent OS, outbound calls are blocked by default by the security gate; in on-prem-required mode no outbound host can be allowed, and other modes use an admin allowlist you are shown. Approval is verified in code. Patent application 202631094208 (fail-closed data boundary) was filed in Aug 2026.
How do you set the approval list for your own organisation?
A method that fits on one page:
- List the actions an agent could take in the process: read, search, score, draft, send, create record, change record, post, file, call outside.
- Score each on three questions. Can it be undone? Does it reach outside the team? Does it change money, a legal position or a person's record?
- Gate anything with a yes on the second or third question. Leave the rest running.
- Name the approver for each gated action, and a deputy.
- Set the expiry. In Business OS a frozen proposal expires after 72 hours; use the same discipline for any inbox.
- Review quarterly. Actions that were never declined may not need a gate; actions that were declined often may need a rule upstream.
| Action | Undo? | Outside? | Money, law, person? | Gate |
|---|---|---|---|---|
| Summarise a tender | Yes | No | No | No |
| Draft a reply | Yes | No | No | No |
| Send the reply | No | Yes | No | Yes |
| Change a lead's owner | Yes | No | No | Depends on your team |
| Post an invoice | No | No | Yes | Yes |
| Reject a claim | No | Yes | Yes | Yes |
What happens over a long weekend?
A proposal waits. In Business OS it expires after 72 hours if nobody approves it, and the copilot is told. In Agent OS a consequential action with no available approver is blocked rather than run. Neither product escalates by sending the action anyway. If a weekend backlog is a problem for your process, the answer is a deputy approver, not a looser gate.
When should you NOT gate an action?
- Read-only work. Searching, summarising and citing your own files should return instantly.
- Internal pipelines. Re-indexing documents on a schedule needs a threshold alert, not a signature.
- Drafts. A draft that nobody will send is not an action.
- A queue with no owner. If no reviewer is assigned, the gate blocks the work. Assign the owner first, then switch the gate on.
The one-line test for a new action
When a new action is added to a workflow, ask: if this ran once in error, what would we have to do on Monday? If the answer is "delete a draft", it runs. If the answer is "call the customer", "reverse a journal" or "write to the regulator", it pauses for a named person. Write the answer beside the action in the approval list, so the next reviewer does not have to re-derive it.
Frequently asked questions
Does every action an AI agent takes need approval?
No. Approval applies to irreversible and high-consequence actions, not to every action. Reading, searching, scoring and drafting run on their own.
What happens if the reviewer rejects a proposed action?
In Business OS the copilot is told and the frozen call never runs. In Agent OS the finding is rejected in the approval inbox and the workflow does not proceed on it.
Is there a record of approvals?
Business OS writes every copilot call to an append-only action log. In Agent OS, each finding and the reviewer's decision sit in the approval inbox.
Related: What is an enterprise AI agent? 10 tasks, each with a human gate · The 7 quality checks an AI agent passes before production · Securing an AI agent connected to ERP, mail and documents · The compensating-control trap
Next step: Agent OS and Business OS; a scoping call on contact.
Facts checked against the BiltIQ product pages, 5 November 2026.
