Dev.to AI 🤖 Ai 👁 0 📖 3 min read

What records do I need when an AI agent talks to my customers?

The short answer The moment an agent speaks to a customer in your name, "the AI said it" stops being an answer. You need a record that survives the chat window: for each customer conversation, what the customer asked,

The short answer

The moment an agent speaks to a customer in your name, "the AI said it" stops being an answer. You need a record that survives the chat window: for each customer conversation, what the customer asked, what the agent replied, which tools the agent called while replying, what each call actually returned, and what happened next (a refund, an escalation, a human taking over). Keep those five things in a form the customer can be shown and anyone can recheck, and a complaint becomes a lookup. Keep only the chat transcript, and every complaint becomes an argument about a log you control and can edit.

Disclosure, once: I run ops for Zambo, the product this article uses. The records cited are public, and the checks described need no account.

Why the transcript alone is not a record

A chat transcript fails in three specific ways, and customer-facing work hits all three.

First, it shows words, not actions. The agent told your customer "your refund is processed" or "your appointment is moved." The transcript proves the sentence was sent. It says nothing about whether a tool call behind the sentence ran, what it returned, or whether it was blocked. Customers do not dispute sentences. They dispute outcomes.

Second, you hold the only copy, and you can edit it. A log stored in your own dashboard, exportable as a file you could regenerate, convinces nobody who is already angry. The record has to be fixed at the time the work ran, in a form whose later editing is detectable.

Third, it has no decision layer. When a human reviewed the conversation, approved the refund, or overrode the agent, that judgment usually lives in a different tool, unlinked. Six weeks later you cannot show which human decision belongs to which agent conversation.

The five records, per customer conversation

  1. The request as received. The customer's message or task, saved as the job's input, not paraphrased by the agent afterward.
  2. The reply as sent. The exact response the customer saw, saved as a step result.
  3. Every tool call behind the reply. Each call the agent made (looking up the order, checking the calendar, issuing the refund) as its own step, with the tool named.
  4. What each call returned. The saved result of every step, including the failures and the blocks. A call that was refused is part of the record, not an embarrassment to delete.
  5. The human decision, recorded on the same page. Accept, flag, escalate, refund approved: one decision, recorded once, against the exact conversation record it judges.

What this looks like as one openable record

On Zambo, a job run through one MCP connection (https://zambo.dev/mcp, 153 tools, 37 on the default surface) leaves a verifiable receipt per call: verification recomputes a SHA-256 hash over the canonical bytes of the saved result, so any later edit to what the agent supposedly did breaks the recompute. A wrap turns the job into a handover, one page carrying the summary, the goal, a named recipient, and the next steps, with the step receipts on the page and the recipient's decision recorded on that same page.

Here is a real one, opened today, November 21, 2026:

https://zambo.dev/handover/d468118b-afc1-43e5-a70c-5e198e876a04

It is a small job, an ETH price check, and I use it because its shape is exactly the customer-record shape at minimum size. Two steps: live_price (saved result ETH $1,621.82 USD, down 6.43% in 24 hours, market cap $197.79B, snapshot fetched Sat, 21 Nov 2026 01:15:07 GMT) and prompt_shield (status pass, risk level none). Each step carries a verifiable receipt you can check in the browser on the page. Its Decisions se

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.