“I never asked for that”: how to win disputes over what an AI agent did

The new dispute: "my assistant did it, not me"

AI assistants now act for customers: they track orders, cancel them, start returns, change bookings. Most of the time that's exactly what the customer wanted. Sometimes it isn't, or the customer says it isn't:

• "I asked about the return policy, I didn't ask to return it."
• "That booking wasn't me."
• "I approved 49, not 490."

When that happens, the business is usually the one that has to prove what was asked and what was approved. The large platforms say it plainly: the merchant stays the merchant of record, and handles returns and disputes. The card networks don't yet have a special category for "the agent went beyond what the customer allowed". So the question becomes simple: can you show what really happened?

What a winning file contains

People who handle disputes for a living give the same list. For each order an agent touched, keep:

1. Which agent asked, and something that proves it was that agent.
2. Who the customer was, and how they connected the agent to their account.
3. What the customer approved, in exact terms, not "a return" but "return order 1042 for 49".
4. When, for every step, with times that can't be edited later.
5. What your rules decided, and whether your team reviewed it.
6. What happened in your store afterwards.

Screenshots and log lines don't hold up well: anyone could have edited them. What holds up is a record that shows if it was changed.

How Allowfy builds that file for every operation

Allowfy sits between the agents and your store, so it sees every step, and records it as it happens:

• The agent proves who it is with its own signing key, and the file keeps that key's fingerprint.
• The customer signed in with you, with a one-time code to their store email, and the file shows when they connected the agent.
• The exact terms the customer approved are in the file, with the version they approved. If the terms change, the old approval no longer counts.
• Your rules' decision and the reason, in Arabic and English, and the time your team reviewed it, if they did.
• Every step from Allowfy's ledger, each linked to the one before it, so the order of events can't be rewritten.

Open any operation in your dashboard and press Open evidence. You get a clean page to print or save as PDF for the customer, the bank or the platform, and a signed file to keep. Allowfy signs the whole file; if anyone changes a single character, checking it shows it.

A short example

A customer disputes a return of a jacket. You open the operation and print the evidence:

• The agent "Helper" asked to return order 1042, signed with its key.
• The customer connected Helper to their account two days earlier, with a code sent to their email.
• The customer approved "Return order 1042 (49.00)" at 18:42.
• The store accepted the return request at 18:42.

That's a very different conversation from "our logs say so".

Before the dispute: fewer disputes

The best dispute is the one that never happens. The same rules that produce the evidence also prevent the risky cases:

• Customer approval of exact terms before anything changes.
• Limits over time, such as two returns a week per customer, counted across every assistant.
• A return window counted from the purchase date.
• Your team reviews large amounts, late returns, or cash-on-delivery orders.

Start free in test mode: everything runs end to end, nothing real changes, and you can open the evidence for every test operation.