Context in Atoms and Molecules

Preamble: Why AI keeps guessing what your company means, and how a small team fixes it without a data team.

Ask Claude or ChatGPT "which of our enterprise accounts are at risk?" and watch everything it has to do. First it has to work out what your company means by "enterprise" and by "at risk". It usually can't, so it either grills the person asking or quietly guesses. Then it builds its own filters, pulls companies from your CRM, fetches whatever it thinks signals risk, does the math, and hands you an answer. Every step is a chance to be subtly, confidently wrong. And two coworkers asking the same question can get two different answers, because they don't share the same definitions either.

That gap sits underneath all the talk about giving AI a "company brain". The phrase sounds great and gets repeated a lot. What it would actually take is rarely the part anyone discusses.

Big companies cover this with a data team. Pipelines keep the metrics updated, the definitions live in a semantic layer someone maintains by hand, and a team owns the whole thing. Most companies don't have that and won't. We don't think being fully enabled by AI should be reserved for big tech and hypergrowth startups. A smaller company growing at a steady pace deserves the same. So the question is how a thirty or seventy person team gets the same result without a data team, without a six month project, and without anyone becoming a data engineer or vibe-coding a security disaster.

This is how we build Midland.

The shift is to stop thinking of context as a pile of documents and start thinking of it as small pieces. Loosely inspired by atomic design, we split it into raw facts and the things built on top of them.

The facts are atoms. A metric, fetched and aggregated from your tools: a number, a date, a list of recent email subjects, whatever it is. A phrase: what your company means by a word like "trial" or "active user", written by a person or drafted from your existing docs and confirmed. Small, clear pieces anyone can read and edit.

Molecules are what you build from those atoms: rules, concepts, processes, the way you actually work. "Enterprise" is when company size is above some number. "Active" is last seen under four days and adoption above sixty percent. "How we onboard" is a short, verified description of the actual process. These are exactly the things AI gets wrong when left to guess, and the things a team actually holds opinions about.

With that in place, the question from the start becomes a few calls that return pre-calculated values the AI never had to interpret. "Enterprise" and "at risk" were defined once, by your team. The AI reads the answer instead of inventing it. You do give something up: an AI left to guess will sometimes surface things you didn't think to ask for. So this isn't about removing that. You can still point a live tool at a source and dig. It is about making the everyday answer consistent and trustworthy by default.

And the direct question is the smaller half of it. The larger half is the questions the AI asks itself. Drafting an email to a customer, it needs to know who that customer is, what was last agreed, and who owns the relationship. Pulling together a report, it needs your definition of "active user", not the one it would guess. Building a presentation, it needs the current number, not one from six weeks ago. Nobody types those questions. They get asked silently, many times, inside a single task. A context layer that only answers what people think to ask is serving the minority of its traffic.

That is also why the data is frozen into a daily snapshot rather than fetched live. Daily for the shared picture, live for the surgical task. For account reviews and planning, a stable picture everyone shares beats a number shifting under you mid conversation. And every value carries its source and a timestamp, so when the AI states a figure you can see where it came from and when. That removes a surprising amount of second-guessing.

The part that matters most to us is control. The big AI tools are adding their own account wide memory, which could solve part of this on its own. But memory the tool quietly manages for you is not the same as a canvas you can open, see, and tune. Teams want to set the small pieces themselves, change a definition, fix a threshold, see exactly what the AI looked at and what it still needs, rather than hoping it picked them up right. Context should be a glass box, not a black box.

The obvious worry is that this becomes one more thing to maintain. Another wiki that goes stale. That is the real test, and it is why the upkeep is mostly the AI's job: an agent that suggests metrics and definitions from your website, your Drive, Slack or Notion; an agent that turns a sentence you type into a new piece of context; the system noticing when the AI keeps reaching for a word nobody has defined and flagging it. A person still decides, but they approve suggestions instead of authoring a data model from scratch.

Edit: There is a follow-up to this, The Things They Just Know, about the half of this problem the defined pieces can't reach. The atoms and molecules here hold what you can write as a number or a rule. But the context that actually steers decisions, the things people just know and never wrote down, doesn't fit that shape. The follow-up explores how a chunk might also carry a small, living, conversational meaning alongside its computed value, captured in the flow of work rather than in a project nobody finishes.

This is what Midland does. We install a working context layer, adapted to how your company operates, and it is yours after that: you own it, you change it, you take it where your work needs it to go. The question we are testing in real deployments is a simple one: given a tool built for exactly this, can a smaller team give their AI the same grounded understanding that big companies get from a data team? We find the answer is shaping up to be yes, and we want to hear from both sides: the people who would build something like this, and the people who have to live with it.