Some decisions should cost a fingerprint

Software has spent two decades making everything one click away. Mostly that's good. But somewhere along the way we decided that every action should be equally cheap — and a few of them really shouldn't be.

I ran into this building the compliance system I've written about before. It holds a graph of a company's regulatory world: the laws, the obligations those laws create, the capabilities and controls that satisfy them. Most of what you do to that graph is safe and reversible. But one operation isn't, and it taught me something about where AI and automation actually need to slow down.

The operation that can't be casual

As the graph grows, you get near-duplicates. Two entries that look like the same obligation, created by different people reading different regulations. The obvious move is to merge them.

And merging is where it gets dangerous. Because sometimes two things that look like duplicates are actually two distinct legal obligations that happen to use similar words — and if you merge them, you have just quietly told your entire compliance posture that one duty exists where there were two. You've erased a requirement. In a regulated business, that's not a data-hygiene mistake; it's the kind of thing that surfaces in an audit two years later with your name on it.

So a merge is high-impact, hard to reverse, and easy to get wrong under time pressure. By every instinct the industry has trained into us, that means… a confirmation dialog. "Are you sure? [Cancel] [OK]." Which everyone clicks through on autopilot.

That's not good enough.

Make the weight physical

In my system, merging two entities doesn't take a click. It takes a fingerprint — a passkey signature, the same WebAuthn mechanism your phone uses, except bound to the exact pair of things being merged and the exact state of the graph at that moment. It's separate from logging in. Being signed in got you the right to propose the merge. Actually committing it requires you to physically, deliberately sign this specific irreversible act.

Three things change the moment you do that:

The decision gets a body behind it. Not a session token that could belong to anyone with a borrowed laptop — a specific human, authenticating a specific act.
It cannot be automated away. An agent, a script, a half-asleep bulk operation cannot quietly perform it at 3am. A person has to be present and intend it.
The record is real. The approval is cryptographically bound to what was approved. "Who collapsed these two obligations, and did they mean to?" has an answer that holds up.

Everything is logged anyway. But a log tells you what happened. A signature tells you someone meant it.

This is the whole thesis, in miniature

I keep coming back to one principle in everything I build: AI proposes, humans decide. The fingerprint is just that principle made physical for the decisions that can't be undone.

And it points at a design idea I think most teams adopting AI are getting wrong. The race right now is to make everything frictionless — let the agent do more, click less, ship faster. But frictionless is a property you want for the reversible 95%, not the irreversible 5%. The craft isn't removing all friction. It's putting the friction exactly where the cost of being wrong is highest, and nowhere else.

A system where deleting a customer and renaming a label feel identical has a design bug, not a convenience feature.

The takeaway

As you hand more of your work to AI and automation, draw the line deliberately: which actions are cheap and reversible, and which ones should cost a fingerprint? Make the first set effortless. Make the second set demand a present, deliberate human.

Not because people can't be trusted. Because the ones you trust most are exactly the ones who'll be moving fast when it matters — and they deserve a system that makes the weight of an irreversible decision impossible to click through by accident.

I'm Tete Mensa-Annan. I help engineering and compliance leaders adopt AI without losing control of the decisions that matter. If your teams are automating faster than you can tell the reversible from the irreversible, that's worth a conversation.