Before you can automate anything

Everyone’s shopping for the next model. Most companies haven’t sorted out the two boring things that actually gate whether any of it works. Neither one is a technology problem.

Your documentation has to be AI-ready

That starts lower than it sounds. First, it has to exist. A huge amount of “documentation” at most companies is really tribal knowledge — the thing Sarah knows, the thing that got explained once in a meeting nobody recorded. An agent can’t read what was never written down, and neither can the new hire who replaces Sarah.

Once it exists, it has to be consumable in a token-efficient way. A wiki written for a human skimming with their eyes, full of preamble and repeated context and screenshots standing in for the words that should be next to them, is expensive and lossy to hand to a model. And it has to be organised and searchable — not just present somewhere in a space with four hundred other pages, but findable by whatever is looking for it, whether that’s a person or an agent doing the search itself.

Existing, efficient, findable. Skip any one of the three and the docs function as decoration, not infrastructure.

Workflows have to be mapped — and agreed on

The second gap is that a lot of workflows were never written down either, because they never had to be. A process survived for years as “ask whoever’s done it before,” which works fine when the person who’s done it before is always around. It stops working the moment you want to hand the process to something that can’t ask around.

Worse, plenty of workflows don’t actually have one agreed version. Different people do the same job differently, and nobody’s had to reconcile that, because human variation absorbs it quietly. You can’t automate a workflow that has three unspoken versions — you have to pick one, or explicitly branch it, before there’s anything to automate at all.

Mapping the workflow comes first. Getting people to actually agree on it comes right after, and is usually the harder of the two — not because the mapping is hard, but because agreeing on “this is the correct way to do it” means someone has to actually decide, out loud, instead of everyone quietly doing their own thing. Once that’s done, the rest is comparatively easy: somebody just has to hit approve.

Neither of these is about the model

This is the same shape as the liability gap I wrote about separately — the scaffolding to automate a given piece of work usually already exists on the model side. What’s missing is company-side infrastructure: documentation good enough to hand to something that can’t ask a follow-up question in the hallway, and workflows clear enough that there’s actually one version to automate. Solve those two and the “we tried AI and it didn’t really help” story mostly goes away on its own.