Release gates for AI procurement agents
Your procurement agent creates requisitions and purchase orders in your customers' ERP and procurement suites. A reply that looks right can still leave the wrong order behind. Agentiqa is an AI Quality System for these agents: release gates run the agent's key tasks before each release and check the purchasing records it wrote.
Where it writes: ERP and procurement suites: requisitions, purchase orders, supplier records, approvals.
What goes wrong with procurement agents
These are the errors that pass a quick look at the reply and show up in the record:
| Error type | Example |
|---|---|
| Wrong unit | A request for 40 boxes saved as 40 pieces. |
| Wrong cost center | The cost center of the requester's previous team. |
| Skipped approval | An order split below the approval threshold, so no one signs off. |
| Duplicate writes | Two POs after a timeout and a retry. |
Behind most of them are a few named error types: wrong data, a false "done" (hallucinated success), duplicate writes, partial writes and regressions since the last release (AI agent regression testing).
What the release gate checks
Supplier, cost center, quantity, unit, price, approval path, and that the order was saved once.
The gate judges the result in the connected system. Different wording in the agent's reply doesn't fail a task. A wrong value, an extra record or a false success message does.
How it works
- Key tasks, in plain words. You and an Agentiqa engineer pick the tasks where a mistake costs the most, for example creating a purchase order from a request or routing an order above the approval limit. Each task is described in plain words.
- Run before each release. The gate runs each task in a real browser with a test account, then checks what changed in the connected system, not only the agent's reply. That is where wrong data, a false "done", duplicate writes and partial writes show up.
- Confirmed regressions only. When you rerun the failed tasks, AI Quality Manager reports a step as a regression only if it fails again. Unconfirmed failures are listed separately.
- Evidence for the decision. Each failure comes with a trace, the recording and steps to reproduce. Your team decides whether the version ships.
AI Quality Manager compares each release with the previous one, shows regression candidates, and proposes test updates and new tests. A person on your team approves them.
Rollout
First release gates are live within 7 minutes. Then comes the 3-day enterprise rollout: an Agentiqa engineer works with your team, on-site or remote, to cover the riskiest actions in your domain, and keeps the gates current as your prompts, models and product change.
Vertical Agents Safety Review
For teams whose agents write into business systems. Free, 20 minutes.
Security
- Credentials are encrypted with AES-256-GCM. The agent types them by name, so their real values stay out of the model's context.
- SOC 2 in progress (not yet achieved). Data processing agreement (DPA) on request.
- Agentiqa is independent of the team and the tools that built the agent.
More in Security.
Already running
Agentiqa already runs release gates every day for an AI procurement company whose agents write into their customers' procurement systems.
Read more
Vertical Agents Safety Review
For teams whose agents write into business systems. Free, 20 minutes.
FAQ
The purchasing records the agent wrote: supplier, cost center, quantity, unit, price, approval path, and that each order was saved once. It reads them in the connected system after each task.
Yes. A gate task can run orders just below and just above the threshold and check that orders above it go to an approver, including when a request is split to slip under the limit.
Yes, when the order is part of a gate task. After the task, the gate reads the purchasing records back and checks that the order was saved once. Two orders for one request, or a "done" with no order behind it, fail the task.
First release gates are live within 7 minutes. The 3-day enterprise rollout covers the riskiest actions in your domain with an Agentiqa engineer, on-site or remote.