AI agent governance: what to require before agents operate
Resources · Governance
Governing AI agents means defining what they can do, who approves, how much they can spend and how every run is audited. In Qurable Autopilot nothing runs without approval from a person or from a policy your team published: every change arrives with its plan and risk, agents never approve themselves, and everything is logged.
- Agent
Proposes
- Plan, risk and impact
- Never approves itself
- Person
Approves
- Business rules
- Two approvers when needed
- Platform
Executes and logs
- Within the caps
- Audit trail and cost
In production
Autopilot agents work on the platform Qurable already runs in production.
- ~150 msEngine at paymentApproved rules only, no AI in the path
- ~3,000Requests per secondSustained in production
- 18MEnd usersIn brands that run on Qurable
- ISO 27001SecurityCurrent certification
Key takeaways
- An AI agent without approval rules, caps and an audit trail is a risk, not a tool.
- In Qurable Autopilot agents propose and a person decides; anything automatic runs only within published policies.
- Every run leaves a trail: who launched it, what it read, what it proposed, who decided, what was applied and what it cost.
The controls
Eight controls to run AI agents with confidence
What a risk committee asks before approving agents that touch budget and customers, and how Qurable Autopilot handles it.
Approval with business rules
Every change arrives with its plan, risk and expected impact. You set the rules: for example, anything over USD 10,000 needs two approvers.
No agent approves itself
A person approves, adjusts or rejects. Anything automatic runs only within a policy your team published.
Roles and permissions
Owner, admin, operator and viewer. Everyone sees and does only what their role allows, with full isolation between companies.
Test environments
Each agent is tried on an example before production and only connects to authorized destinations.
Spending caps
Caps on AI spend per company, per agent and per run; your team approves the rewards budget and its caps. When an agent hits its cap, it stops and explains why.
Pause anytime
Stop one agent or the whole operation whenever you decide. Work picks up where it left off.
Instructions only from your team
Documents, emails and external data are read, but they never give an agent orders. Credentials are encrypted per company.
An audit trail for every run
Who launched it, what it read, what it proposed, who decided, what was applied and what it cost. Nothing runs until it's confirmed.
Checklist
What your risk committee should ask.
Use it with any AI agent vendor. The second column says what answer to look for.
| Question | What answer to look for |
|---|---|
| Who approves what an agent proposes? | A person, with rules in business terms (amounts, segments, channels) and more than one approver for big changes. |
| Can an agent approve itself? | No. Approval has to come from a person or from a policy the team published. |
| Are there spending caps? | Caps on AI spend per company, per agent and per run; your team approves the rewards budget and its caps. |
| What happens at the cap? | The agent stops and explains why; it doesn't keep spending. |
| What data does the model see and where is it hosted? | Agreed with your security team, with isolation between companies and encrypted credentials. |
| Can an external document give it orders? | No: external content is read as data, never as an instruction. |
| Is it certified? | At least a current ISO 27001 certification. |
| Can it be tested before production? | Test environments and connections only to authorized destinations. |
| Can it be stopped? | Pause one agent or the whole operation without losing work. |
| Is the AI in the payment path? | It shouldn't be: the payment has to decide in milliseconds even if a model changes or fails. |
| What gets logged? | Who launched each run, what it read, what it proposed, who decided, what was applied and what it cost. |
| Can the result be proven? | Every action measured against a control group, with the money recorded in an auditable ledger. |
The difference
An ungoverned agent and one your company controls.
The AI is the same. What changes is who decides, how much it can spend and what gets logged.
| Ungoverned | With Qurable Autopilot | |
|---|---|---|
| Who decides | The agent applies what it proposes | A person approves, or a policy your team published |
| How much it can spend | No clear limit | Caps per company, agent and run |
| What happens at checkout | A model in the transaction path | Only approved rules, in ~150 ms; if an agent fails, the payment goes through |
| External content | A document can change its behavior | It's read, but never gives orders |
| Before production | Tested live | Test environment and authorized destinations |
| Trail | Scattered technical logs | A full audit trail per run, with its cost |
With Qurable
Agents that operate, people who decide.
Qurable Autopilot puts AI agents to work on your business metrics, with these controls from day one.
Qurable Autopilot
AI agents that propose the strategy, execute it with your approval and measure every action against a control group.
See Qurable AutopilotDouble-entry ledger
Every reward an agent applies is recorded, with its cost and contribution, so it can be audited.
See the ledgerInformation security
Current ISO 27001, any cloud and a shared or dedicated cell.
See the security policy
FAQ
What people ask about AI agent governance
Control, spend, data and audit.
It's the set of rules that defines what an agent can do, who approves what it proposes, how much it can spend and how every run is audited. Without it, an agent that touches budget or customers is a risk.
Only what a policy your team published allows, within its caps. Everything else arrives as a proposal with its plan, risk and expected impact, and a person approves, adjusts or rejects it.
With caps on AI spend per company, per agent and per run; your team approves the rewards budget and its caps. When an agent hits its cap, it stops and explains why.
Agents read documents, emails and external data as information, never as orders. Each agent also connects only to authorized destinations.
Who launched it, what it read, what it proposed, who decided, what was applied and what it cost. Nothing runs until it's confirmed.
No. Agents work outside the transaction; only approved rules run at checkout, in ~150 ms, and if an agent fails the payment goes through.
Last reviewed
Further reading
Keep reading about AI agents.
Where these controls show up in the product.
- ConceptQurable Autopilot
AI agents to retain customers and sell more.
- GuideAI in loyalty programs
The agents and the engine that decides at checkout.
- GuideRFP checklist
What to ask of an enterprise platform.
Review these controls with your security team.
In a short meeting we'll show you how each Qurable Autopilot agent is approved, capped and audited.