FAIRFLAI

Context can't be bought, only cared for.

Every time someone opens an AI assistant at work, they explain all over again who the customer is, who can grant the discount, which version of the contract is the valid one. They do it in three lines, from memory, and each time a little differently. Multiply that by a hundred licenses and two hundred working days: it is the biggest cost of enterprise AI, and it doesn't appear on any invoice.

The thesis of this article is simple: a company's context is a collective good, a commons in Elinor Ostrom's sense. It doesn't belong to whoever wrote it, nor to the model vendor. Someone has to look after it, with explicit rules, otherwise it degrades like a pasture left to itself (worse, in fact: with a pasture you can at least see when it's ruined). I owe the image of the pasture as a common good to Michele Kettmajer, and it echoes Ostrom's examples of rice paddies.

Superadditive is the name we at FAIRFLAI have given to our version of this idea: a queryable map of how an organization works, readable by people and by agents. The word was brought to us by Lorenzo: superadditivity, when the whole is worth more than the sum of its parts. Here I try to explain what it's for, who takes care of it, and why it's the only part of AI you can't buy.

The test we use with clients is a single question: who is accountable for this, under which agreement, and where can the data be read? The org chart knows who reports to whom, not who decides what. ERP and CRM know the numbers, not who vouches for them. The wiki knows what someone wrote two years ago, not what the organization considers true today. Each one holds a piece. No one holds the graph.

All the use cases stem from there. The first is the most prosaic: giving agents context. An agent drafting a response to a complaint needs to know who can grant an exception and up to what threshold; otherwise it optimizes the piece it can see and writes something that is locally consistent and wrong for the company. At an insurance company we worked with this autumn, delegations of authority, powers of attorney and job descriptions all existed. The practice of who really decides on an exception, however, was tacit, and an agent can't read what is tacit.

The second is rotation. When someone changes role, the knowledge of that role usually leaves with them, and whoever arrives spends the first months rebuilding it one meeting at a time. If context is attached to the role and not to the person, the newcomer inherits the map: the open agreements, the decisions taken and their reasons, the data they become the guarantor of.

The third is cross-cutting queries. Which practices have we used in industrial projects, with what results, and who led them? Which client had already asked us for something similar, and what did we tell them? Today the answer is a chain of Slack messages and the memory of whoever was there, and when that person leaves, the answer leaves with them. With a curated map, the question is asked once, to a person or to an agent, and the answer comes with its sources.

Elinor Ostrom won the Nobel Prize in Economics in 2009 for showing something many economists considered impossible. Communities of fishers, herders and irrigators manage common resources for centuries without privatizing them and without handing them over to the state, provided they have clear boundaries, rules suited to the place, and the ability to change the rules that affect them (Governing the Commons, 1990). In 2007, with Charlotte Hess, she applied the same framework to knowledge in Understanding Knowledge as a Commons.

Knowledge differs from a pasture in one important way: using it doesn't consume it. So the risk isn't exploitation, it's neglect. The document nobody updates, the procedure written by someone who has since left, the decision made on a call and never recorded. Corporate context almost always dies because nobody has the job of looking after it, almost never from overuse. That's why Superadditive is designed more as an institution than as software. Every object in the map has an owner, and the owner is always a role, never a person: people change, roles remain. Boundaries are drawn by the people who work there, across three overlapping spaces (personal, team or project, organization) that are genuinely separate from one another, because not everything I write for myself should become everyone's. Changes come in as proposals and are approved by whoever has authority over that domain. Even the vocabulary we use to describe the company, the ontology, changes the same way: whoever needs a new class proposes it within Superadditive and a role approves it. Those affected by the rules can propose changing them. Of Ostrom's principles, it's the one companies respect least, and without it the map becomes someone else's org chart. This is where the technology comes in. I'll keep it short because it's not the point, but without this part the rest doesn't hold up. A map that stays up to date only thanks to people's goodwill is dead within six months (we all know it: we all have a company wiki).

The maintenance work is done by agents. One gathers statements from meetings, documents and threads, attributes them to the right object and states how confident it is in the attribution. One consolidates the notes accumulated on an object and rewrites it, instead of appending lines forever. One, every night, runs the checks no human ever would: objects without an owner, data past its expiry date, two sources giving different thresholds for the same approval, a role declared in governance that never signs anything in the workflows. The last one, the most interesting, asks people questions in the channels where they already work. It starts from what the map knows it doesn't know: every object has a field for gaps, and the open gaps become the agenda of questions.

None of these agents writes to the map. They propose. A note doesn't become knowledge just because it has been indexed: if it confirms an object, it consolidates it; if it diverges, it opens a tension assigned to an owner, who resolves it. That's the difference from many enterprise document search projects, where a wrong set of meeting minutes from 2023 weighs as much as the policy approved yesterday. In our setup, active routines have a daily write cap (200, 150 and 30 depending on the routine). The cap is a deliberate choice. An agent that produces more proposals than an owner can read isn't helping, it's just moving the bottleneck.

The model can be bought, and in six months a better one will come out that your competitors will buy too. Licenses can be bought. Connectors can be bought. Even our method, if you like, can be bought. What can't be bought is the map of how your company works: who decides what, under which agreements, on the basis of which data, with which tolerated exceptions. That accumulates one recorded decision at a time, and it is worth something only to you.

Jack Dorsey, in the piece Block published on March 31, 2026 (From Hierarchy to Intelligence, written with Roelof Botha of Sequoia), describes hierarchy as a protocol for carrying information and the company world model as what makes it replaceable. He's half right. AI carries information better than three layers of middle management, but only if someone has decided what is true, who vouches for it and how it gets corrected. Removing the layers without building the commons gives you an organization that forgets faster.

That's why I think context will be what a company is worth in the years ahead, more than its technology stack. It's the only asset that agents multiply instead of turning into a commodity. It's a bet, and I'm declaring it as such.

Superadditive isn't software you install. It's a working framework and a technology framework at once. On one side, roles, proposal and approval rules, boundaries between personal, team and organizational space; on the other, the ontology, the connectors to source systems and the agents that do the maintenance. Rules without agents become yet another wiki nobody updates; agents without rules produce noise with great confidence. At FAIRFLAI we use it on ourselves first. Our governance is holacratic and lives on Nestr, projects and assignments live on monday, and the map reads from there. The ontological layer is written only by the service that syncs the source systems. People and agents write in the layer above, as proposals attached to nodes, which are either validated or superseded.

Now we're bringing it into client projects. We start with a diagnosis and a pilot domain with a real question, build the minimum registry for that domain, then connect the source systems and the first agents. At first we hold the role that looks after the map ourselves; then it passes to someone inside the client's organization, and the handover is written into the contract as a deliverable. A map that depends on the consultant goes back to being something you buy.

If you're buying licenses these months, before the next one try answering two questions. In which process does one of your agents get things wrong today for lack of context? And who in your company has the role of looking after that context? If the answer to the second is nobody, write to us.

Want to talk about it?

Tell us about your challenges and we'll propose a tailored solution for your organization and your business.