What You Can't Take Back

A few decisions in this cycle are genuinely irreversible. Most of them don't look it at the time. · Juan Vegarra

I've written elsewhere about one-way doors as a general idea. This is the narrower version: which doors in this particular cycle only open once.

The list is shorter than people fear and different from what they expect. Almost nothing about model choice is irreversible. Almost everything about what you do with data is.

Data you gave away

The first genuine one-way door is data leaving your boundary under terms you didn't read carefully.

You cannot unsend it. You can terminate the contract, delete the account, get an attestation. What you can't do is restore the situation where that data had only ever been in your control. If it trained something, it's in the weights. If it was aggregated, it's in a benchmark. If it was a customer's, you've made a commitment on their behalf that you may not have had the standing to make.

This is worth real legal attention early and gets almost none, because it doesn't feel like a decision. It feels like a checkbox during onboarding.

Data you didn't capture

The second door is quieter and I think more expensive. It's the data you could have been collecting and weren't.

Every month a process runs without capturing its outcomes is a month of training signal you can never recover. You can start collecting tomorrow. You cannot go back and instrument last year.

In practice this is the difference between companies that have a real data position in five years and companies that have a licence to someone else's. And the cost of starting is usually small. The cost of not having started is enormous and arrives late, which is exactly the shape of decision organizations handle worst.

Trust with the people who have to use it

The third door is a group of people deciding the tool isn't for them.

I've watched this happen in a clinical setting and it's brutal to reverse. Ship something that makes a professional look bad in front of a colleague, or that quietly implies their judgment was the problem, and you don't get a second evaluation. You get polite non-adoption, which is much harder to diagnose than rejection because nobody complains.

The second version of the product can be genuinely better and it won't matter, because they're not looking anymore. First contact with a skeptical expert user is close to irreversible, and it's usually delegated to whoever has capacity that week.

What isn't a one-way door

Worth saying plainly, because fear of lock-in causes its own paralysis.

Model choice is reversible. Vendor choice is mostly reversible. Architecture is reversible at a cost you can estimate. Being on the wrong platform for eighteen months is recoverable and quite common.

People spend enormous energy on the reversible category because it's technical and legible, and very little on the three above, because those are legal, organizational and human. The genuinely irreversible decisions in this cycle mostly aren't made by engineers.

On Monday

Three questions. What data has left our boundary and under what terms. What are we running right now that we're not instrumenting. And who has already formed an opinion of our tool that we haven't asked about.

The first is a legal review. The second is a week of engineering. The third is a conversation you'll want to have before the next release, not after.

← Previous: Sequencing Beats Selection · Read the series → · Next: Scope It to a Number →

Juan Vegarra is the author of An Outsider’s Playbook. The views here are his own. More essays · Advisory · Write me