Individual AI made everyone faster. It hasn't made institutions smarter.
Spend any time with the people running AI inside a bank, the CDAOs and chiefs of AI, and the same tension surfaces. They have invested enormous sums. They have felt the technology work first-hand, as everyone now has: Claude and ChatGPT summarising a hundred-page document in seconds, drafting the memo, writing the article (though not, for the record, this one). The personal impact is obvious and immediate. And yet at the level of the institution, the returns have not appeared. Cost-to-income has not moved the way the business cases promised. The output of the firm looks much as it did before anyone had a model on their desk.
The puzzle is genuine. How can a technology this transformative for individuals leave the organisations they work for so little changed?
Individual AI is not institutional AI
A recent post from a16z named the gap precisely. There is individual AI, which makes a single person faster, and there is institutional AI, which changes what the organisation as a whole is capable of. The two are not the same thing, and one does not automatically produce the other. Institutional AI, the argument runs, needs things individual tools never have to supply: coordination toward shared goals, objectivity that resists simply telling each user what they want to hear, and an edge that is purpose-built rather than borrowed from a general model anyone can rent.
Their analogy is the electric motor, and it is worth dwelling on. Motors arrived in factories in the 1890s, but for three decades they changed almost nothing. Plants had been designed around a single central steam engine, machines crowded close to the driveshaft, and dropping a motor into that layout simply swapped one power source for another. The gains came only in the 1920s, once a generation of managers stopped retrofitting and rebuilt the floor around distributed power: machines arranged by the logic of the work, assembly lines, a factory organised around the task rather than the reach of a belt. The technology had been available for thirty years. The reorganisation was what released its value.
We think that is exactly right. Which leaves the question every bank is now facing, whether or not it has put it this way. How do you bridge the gap between individual and institutional AI? What is the modern equivalent of reorganising the factory floor?
It is not about model quality. It is about context.
The models are extraordinary, and improving by the month. AI is plainly a transformative technology. The constraint sits one level down. An institution cannot get institutional returns from a tool that knows nothing about the institution, and that is the situation most banks are in. A foundation model arrives at every task with its training data and whatever fits in the prompt. It knows a great deal about the world in general and nothing in particular about who relies on the thing that just changed, why a decision was taken two quarters ago, or which assumptions have since gone stale.
That missing material is context, and it is the stuff that makes JP Morgan JP Morgan, and Lloyds Lloyds. It is not held in any single head, and it certainly is not held in a model. It is distributed across people, encoded in past decisions, and expressed mostly in relationships: between systems, between commitments, between the person who made a choice and the reasons they made it. Much of it is tacit, and almost all of it has a history. Without it, even a brilliant model is working blind, and the promised returns do not arrive.
This is also where any durable advantage has to live. As the models commoditise, and they are commoditising quickly, the intelligence applied to a problem is increasingly available to everyone at similar cost. What is not for sale is the context it is applied to.
Closing the gap takes a System of Context
Bridging individual and institutional AI means giving the institution a memory that machines can use. Not a dashboard, which is a snapshot, and not a data lake, which is an archive, but a living model of the organisation’s entities and their relationships as they change over time, that both people and agents can query and reason over. We call it a System of Context, and it is what we have spent years building at Pometry.
It runs on a temporal graph engine, Raphtory, in which every entity and relationship carries its full history. The organisation can be queried as it stands today, or as it stood on the day a decision was made, so the “why” survives long after the people who knew it have moved on. Agents reach it through a standard interface, and because every answer traces back to the data beneath it, the result is auditable rather than merely plausible. It runs against the bank’s own data, in its own environment.
This is the substrate the a16z pillars quietly assume. Coordination, objectivity and a genuine edge are not properties of a cleverer model. They are properties of a shared context that every agent can reason over, the way every machine on the redesigned factory floor could draw on the same distributed power.
What this looks like in practice
None of this is hypothetical. It is already underway in the more progressive banks, and the easiest way to see the shift is through the work itself.
Transformation risk. A tier-one programme generates millions of signals across change boards, tickets, code and risk forums. A System of Context turns those into a live view of how the programme is actually behaving, where dependencies are accumulating and where fragility is forming, early enough to act rather than days before two critical-path dependencies fail at once.
Client coverage. Another tier-one bank is assembling a live view of its corporate and investment-banking clients, together with the employees, counterparties and suppliers around them and how those relationships move over time. Coverage stops being guided by a stale CRM extract and starts surfacing what matters: that a relationship manager already knows the newly appointed CFO of a target client, say, or that two parts of the bank are unknowingly competing for the same account.
Agents that brief each other. Someone commits a change to a codebase, or amends a clause in a counterparty contract, and then goes on leave. A week later the logic behind it matters and the person who holds it is unreachable. Rather than wait, guess, or reverse-engineer the reasoning, you ask their agent. Not the colleague, but the agent operating in the same context they did, which can account for what the change depended on, what it was responding to and what else it touched, with the evidence attached. Your agent asks theirs, and gets an answer.
The first wave of enterprise AI asked how much faster each person could go. The institutional question is larger: what an organisation could understand about itself, and decide, if its memory were continuous, shared, and legible to the machines now working inside it. Individual AI was the electric motor. The System of Context is how a bank rebuilds the floor around it.