Skip to content
A-Eye Level
  • Articles
  • Explainers
  • Topics
  • About
  • Newsletter
  • Contact
← Back to newsletter
The 5-Minute AI Decision · Issue #33

When Two Corrections Disagree

October 29, 2026

Say an AI agent handles your customer renewals. One week a sales rep tells it a small, long-standing customer always gets a call before a price change. A month later a manager tells it calls are only for renewals above a set size. They came in separate chats, and nobody compared them. You may not even be able to inspect what the agent remembers. OpenAI’s help page for dots, the personal agents it launched in ChatGPT, says people “can provide updated guidance” but “currently cannot view, delete, or directly modify individual dot memories” (read October 28, 2026).

Why It Matters

Our article on agent memory argued for keeping the team’s accepted corrections in a playbook outside the agent. These two don’t add up to one rule, because that customer sits below the size line. A playbook can hold both lines, but it can’t decide between them. And a correction given to an agent isn’t automatically company policy: the rep knows the customer, the manager knows the pricing rule, and neither may have authority to change the rule.

The Decision

The question is where the general rule ends and the exceptions begin. Whoever sets the policy owns the rule, which treats customers alike. The manager accountable for a customer can approve documented exceptions within agreed limits, which keeps judgment close to the client. What you decide is how far that approval reaches: wide limits keep renewals moving, narrow ones keep them consistent.

What To Do This Week

  1. Ask the owner of one process an agent handles to review the conversations, instructions and corrections they can access for any that point different ways. If they can’t access those records, that’s a finding too.
  2. Have whoever sets that policy and the manager accountable for the customer agree which line is the rule and which the exception. If they disagree, you decide.
  3. Record the ruling, its owner and the exception limits in the playbook. Update the instructions or workflow the agent actually uses, then test the conflicting case. Until it passes, a named person approves any price change or commitment it affects.

What Not To Do

Don’t treat the latest correction, or the most senior person who gave one, as the ruling. Don’t read a review that finds no conflict as proof there is none. Some of what an agent learns may be unreadable. And don’t settle it only in a chat with the agent. Without the playbook entry, it’s one more correction.

Signal Boost

Where Does Your AI Agent Keep What It Learns?, A-Eye Level, October 11, 2026. It covers what the playbook should hold.

Get the next decision in your inbox

We use your email to manage and send the weekly newsletter through Buttondown, our email provider. Unsubscribe any time. Our Privacy Policy has the details.

← Previous issue #32 · The Token Bill Behind One Task Next issue → #34 · Whose Call Was It?
View all issues →
A-Eye Level

AI clarity for business leaders.

A-Eye Level
Privacy Policy Cookie Policy Terms & Conditions Accessibility
© 2026 A-Eye Level. All rights reserved.
The 5-Minute AI Decision Subscribe
Subscribe