Human in the loop: designing AI approval workflows for regulated pricing

AI agents can now draft pricing models, apply rate changes and run test suites in minutes. That speed is useful, and it also changes where control needs to sit. When an agent can make a change faster than a person can read it, the approval workflow becomes the main safeguard.

Regulators in the UK, EU and US all expect firms to keep meaningful human oversight of automated decisions that affect customers. For pricing, that means designing who can do what, what evidence reviewers see and how every step is recorded.

Separate drafting, approving and publishing

A sound workflow splits model changes into three stages, with different permissions for each.

• Draft. A working copy of the model where changes are made. People and AI agents can both work here.
• Approved. A version that a reviewer has checked and signed off. Approval should only be possible for a person with the right role.
• Published. The version used to price live business. Publishing should also be restricted to named people.

The most important rule is that the person or agent who made a change should not be the one who approves it. This is the same separation of duties used for financial controls and software releases, and it applies to AI agents in the same way it applies to staff.

What AI agents should be allowed to do

Agents are most useful, and safest, when their permissions are narrow:

• Read models, tests and documentation.
• Edit draft versions only, never approved or published ones.
• Run tests and report the results.
• Explain what they changed and why.

Agents should not be able to approve or publish, change the expected result of a failing test without a person agreeing, or edit anything that is already live. A system should enforce these limits itself, rather than relying on the agent to follow instructions.

What reviewers need to see

Approval only works if the reviewer has the right information. For each change, a reviewer should be able to see:

1. The difference. Exactly which rates, rules or formulas changed, from what to what.
2. The reason. Why the change was made, with a reference to the rating guide, filing or decision behind it.
3. Test results. Which tests passed and failed, and for any test whose expected result changed, the explanation. Our article on testing pricing models covers how to set these up.
4. Example quotes. Traces for a few representative quotes before and after the change, showing how each step moved.
5. Who made it. Whether the change came from a person, an agent or both.

A reviewer who sees all five can approve with confidence. A reviewer shown only a summary written by the agent cannot.

Keep a complete audit trail

Every step in the workflow should leave a record:

• who created each draft and when;
• every edit, whether made by a person or an agent;
• the test results at approval;
• who approved and who published, with timestamps;
• every later change of status, such as unpublishing.

Combined with a trace for each quote, this lets a firm show which version priced any quote, who approved that version and what evidence they had.

Plan for urgent changes

Sometimes a pricing error needs fixing quickly. The workflow should allow this without removing the controls:

• Keep a small group of people who can approve and publish urgent fixes.
• Still require the change to go through a draft and a test run, even if review is faster.
• Allow a quick return to the previous published version while a fix is prepared.
• Review every urgent change afterwards and record what happened.

Rolling back to a known good version is usually safer than editing a live model under pressure.

Make oversight meaningful

Human oversight is only effective if reviewers have the time, information and authority to reject a change. Practical ways to support this include limiting the size of each change, so reviews stay manageable, training reviewers to read test results and traces, and tracking how often changes are rejected or sent back. A process where every change is approved without comment is a sign that oversight has become a formality.

How Ptally supports this

In Ptally, every change starts as a draft. AI agents can edit drafts over MCP but cannot approve or publish. Approving and publishing are done by people, and every status change is recorded with who made it and when. Every version carries its own tests, and every quote can return a debug trace. Our article on explainability in regulated industries covers the wider approach.

Getting started

If you want AI to speed up pricing changes while keeping people in control, send us your rating guide or current model. There is no implementation fee and nothing to pay upfront. Talk to our team.

This article is general information and is not legal advice.