Human approval gates in agent workflows
Owner, threshold, evidence trail: keep consequential agent actions not activated until a named person approves — illustrated with the AiAx United purchase-proposal demo.
Agents can prepare a purchase proposal, gather prices and assemble the evidence. That does not mean the business has already decided. Risk- and compliance-minded buyers care about a simpler rule: no automation of the consequential step before a named person approves it.
This guide describes how human approval gates work when agent cases, Company Brain context and the United workspace sit together, with a clear owner, a threshold, an evidence trail, and a workflow that stays not activated until someone accepts it.
Why gates matter more than clever drafts
When agents run across systems, the dangerous moment is not the draft. It is the silent jump from “prepared” to “done”: a payment released, a purchase ordered, a customer message sent, a case closed. If that jump is implicit, process owners lose the audit story they need, and reviewers cannot see what they were asked to accept.
A useful gate is therefore not a decorative checkbox. It is a standing requirement that:
- Names an owner who may approve or reject.
- States a threshold (amount, risk class, or action type) that triggers the gate.
- Keeps an evidence trail of what the agent assembled and what the human decided.
- Leaves the workflow inactive for the consequential action until approval lands.
That pattern matches how AiAx talks about people staying involved on the marketing site: people set direction and review decisions where the workflow requires it; demonstrations show proposed next steps for people to review, not unsupervised completion.
Four beats of an approval-gated agent case
- Assemble: Agents retrieve authorised context (cases, prices, documents, prior decisions) inside workspace permissions and Company Brain access rules.
- Propose: The agent produces a concrete proposal with sources and a recommended next step.
- Gate: If the action crosses the configured threshold, the consequential step stays blocked; the named owner sees the proposal and evidence.
- Decide: Approve, reject or send back with conditions. Only then may automation continue that path.
Explore AiAxBrain · Explore AiAx Business
A concrete illustration (AiAx United · synthetic)
On the public AiAx site, an AiAx United demo workflow frames a purchase proposal: four HMI panels, two PLCs, prices and approval requirements. A close view of the recorded proposal shows an amount around EUR 6,400 together with a human approval requirement. The story is deliberate: agents coordinate the routine work; people approve at clear, visible points.
Treat that demonstration as synthetic AiAx United sample data, not a named customer deployment. In a live workspace the same shape would look like this:
Prepare a purchase proposal for the HMI/PLC package. Include prices and stop before any order is placed. The finance owner must approve above the threshold.
The agent gathers authorised catalogue and quote context, drafts the proposal and attaches the evidence. Because the amount crosses the gate, the order path stays not activated. The named finance (or process) owner reviews the trail and either approves, rejects or asks for a revision. Until that decision, there is no automated order.
Illustrative workflow using AiAx United sample information from the public site demo. Not a documented customer case. No claim that any particular customer has this workflow live.
The same gate pattern appears in other published AiAx stories: field observations become structured cases where actions that require approval return to the responsible person; CRM follow-ups that need approval stay for human review before send. The centre piece is consistent: prepare freely within scope, and activate consequential actions only after approval.
What “not activated until approved” protects
- Fraud and error risk: A wrong price or wrong vendor cannot quietly become a purchase.
- Mandate clarity: Agents work inside an agreed mandate; exceeding it is a review event, not a surprise.
- Auditability: Compliance can see who approved what, on which evidence, at which threshold.
- Trust with operators: People stay in charge of decisions that change money, commitments or external messages.
Company Brain helps the gate stay grounded: the owner, the process and the supporting documents are connected context the agent may retrieve within authorised access, not an open chat over every file. Discovery still comes first for new workflows: map the as-is, name the pilot owner and define human review before you treat a finding as ready to automate.
Decision checklist
Before you treat an agent workflow as ready for production traffic that can spend, commit or message externally, confirm:
- A named owner (role or person) is recorded for each gated action class
- A written threshold or trigger exists (amount, sensitivity, destination, or action type)
- Agents may prepare proposals and evidence inside authorised context only
- The consequential step stays not activated until approval is recorded
- The evidence trail (proposal, sources, decision, timestamp) is retained for review
- Reject and revise paths are defined; silence is not approval
- Synthetic demos and diagrams are labelled; customer sources and permissions are verified in implementation
- AiAx Cyber (if discussed) remains concept only and is never treated as a shipping control plane
