Reading an AI Receipt in BearScope: A Walkthrough
Open any AI action and you'll see exactly what happened. Here's how to read the receipt.
When an AI agent does something on your behalf — issues a refund, sends a reply, tags a conversation — you should be able to open that action later and understand it completely, without being an engineer. In BearScope, every action leaves a receipt that does exactly that. This is a walkthrough of how to read one.
We'll go field by field, in the order a receipt is laid out, using a refund as the example.
An AI receipt walkthrough, top to bottom
A receipt reconstructs a single action from start to finish. It has six parts, and they're arranged in the order events actually happened, so reading top to bottom tells you the whole story.
Here's the shape of it, then we'll take each part in turn:
| Part | What it tells you |
|---|---|
| Trigger | What started this action |
| Data seen | What the agent looked at, and from where |
| Decision | What the agent proposed to do, in plain words |
| Check | Which guardrails ran, and whether they passed |
| Action | What actually executed, and exactly once |
| Result | The outcome and timestamp |
Open any action in the conversation timeline and the receipt expands. Let's read it.
Trigger and data seen: what set this off and what it knew
The top of the receipt is the trigger — the thing that prompted the agent to act. Usually it's a customer message ("My order arrived damaged, I'd like a refund"), sometimes a status change or a rule. You read this first because it frames everything below it.
Right under that is data seen: every piece of information the agent pulled before deciding, and where each came from. For a refund, that's typically the order record, the payment history, and your refund policy. Each item is tagged with its source — the store, the payments tool, your policy doc — so you can confirm the agent acted on real values.
The "data seen" section is how you prove the agent didn't make anything up. Every number it used has a source you can click back to.
This is the part people skip and shouldn't. If an action ever looks wrong, the cause is usually here — the agent saw stale or partial data — not in the reasoning further down.
Decision and check: what it wanted to do, and what stopped or allowed it
Next is the decision: the action the agent proposed, written in plain language. Not a code, not a payload — "issue a $42 refund on order 10482 to the original payment method." If the agent reasoned about alternatives, a short version of that reasoning sits here too.
Below the decision is the check — and this is the part that makes the receipt worth reading. Before any action runs, it passes through the guardrails you set. The check section shows:
- Which limits applied (for example, a per-refund cap).
- Whether the proposed action was inside those limits.
- Whether anything was blocked, adjusted, or sent to a person.
A receipt that shows the check ran is how you prove the guardrails were actually enforced, not just configured. If the check had blocked this refund — say it exceeded your cap — the receipt would show the block and the handoff to a person, and the action below would read "not executed."
Action and result: what actually happened
The bottom two parts are the outcome. The action is what truly executed — the committed refund, performed through the connector to your payments tool — with confirmation it ran exactly once. This matters because a retry or a double-click should never produce two refunds, and the receipt is where you confirm it didn't.
The result closes the loop: success or failure, the amount, the reference number from the payments tool, and the timestamp. If the action failed, that's recorded here too, with the reason — so a failure is just as legible as a success.
Read in order, those six parts answer the only questions that matter about an AI action: what started it, what it knew, what it wanted, whether the rules held, what it did, and how it turned out.
Where receipts pay off
Once you can read a receipt fluently, three situations get fast. A customer disputes a charge — you open the receipt, see the single refund and its timestamp, and the conversation ends. An action looks wrong — you trace it to the exact part (usually the data) and fix that gap instead of pulling the agent offline. And an auditor asks how the AI behaves — you hand them receipts, which is evidence, not assurance.
That's the point of building the receipt into the action itself: there's always a clear, readable account of what your AI did and why. See how the product works, read how we keep AI actions safe and auditable, or book a walkthrough to read real receipts on your own conversations.
See it on your own conversations.
Bring your busiest day. We'll score every conversation in it.
Book a walkthrough →