Automation fails in a particular, maddening way. Nothing crashes. No alarm goes off. The customer just doesn’t get the thing they were promised, and by the time you hear about it, you’re staring at a chain of systems — your app, the AI step, an integration, a payment provider, a calendar — and every one of them says “wasn’t me.”
That’s the real problem with an AI mistake. It’s rarely “the AI is broken.” It’s “something in a chain of five things went wrong, and you can’t tell which.” And until you can, you can’t fix it, you can’t bill the right party, and you can’t honestly answer the customer.
”The system says it’s fine” is not the same as “it’s fine”
Here’s a true-to-life one. An automation processes a refund. The log says
refund_processed: true. The customer says the money never arrived. The bank
says it never received a request. Three systems, three confident answers, and
they can’t all be right.
Where did it actually break? Did the AI step decide to refund and record it? Did it call the payment provider, or just think it did? Did the provider get the request and drop it? With ordinary logs you’re reduced to guessing — and to giving the refund again “to be safe,” because you can’t prove you already did.
This isn’t hypothetical. When a signed audit layer went into a live booking workflow, it immediately surfaced exactly this shape of bug: a run that returned an empty response to the customer while the internal record showed a complete, correct five-step decision. The gap between “what the customer got” and “what the system did” is invisible until something makes it visible. (Four such pre-existing bugs turned up in the first week — none of them caused by the audit layer; it just made them readable.)
You need a record of each step, not one log at the end
The fix isn’t a smarter AI. It’s a record that’s specific enough to point at the exact link that broke — and trustworthy enough that the other party can’t wave it away.
That’s what a per-step receipt gives you. Instead of one log line at the end (“refund processed”), each consequential step leaves its own signed record: what it received, what it decided, and when. So when something goes wrong you can see the chain — the AI decided to refund, issued the instruction at 14:02, here’s the exact request — and the gap jumps out. Either the receipt for the payment step exists (so the break is downstream, at the provider) or it doesn’t (so the break is upstream, in your workflow). Minutes, not guesswork.
And because each receipt is signed and tamper-evident, it’s not just useful to you. When the dispute is “your system vs. their system,” a record the other side can check independently — without trusting you — is what ends the argument. AXR is a small, open-source layer that produces exactly these receipts for automated workflows, including n8n.
What this does and doesn’t get you
Be clear-eyed about it. A receipt won’t stop the AI from making a bad call, and it won’t prove the call was right. What it does is make the truth recoverable: it shows precisely what happened at each step, and it makes a later edit impossible to hide. In a “whose fault is it” fight, that’s usually the whole game — you stop arguing about whose word to trust and start looking at a record nobody can quietly change.
It also scales down. You don’t need a compliance department. A solo operator who can pull up a signed receipt showing the refund instruction went out — with a timestamp — is in a completely different position from one who can only say “the system says it’s fine.”
Seeing is faster than reading
The quickest way to understand what a tamper-evident receipt actually is: look at one and try to break it. There’s a 30-second demo — a real signed receipt where you edit a field and watch the seal turn red. Once you’ve seen the seal break, the whole idea clicks, and “can you prove what your AI did?” stops being a scary question and becomes a button you can press.
The next time something in your automation goes sideways, you want to be the person who can say “here’s exactly what happened, and here’s the proof” — not the one giving a second refund and hoping.