To understand my view on this, we have to go back to the processes within a very successful business function fifteen years ago, the PokerStars support operation. It used to handle customer queries in multiple languages, from all around the world, 24/7 365 days a year. It processed thousands of emails per day, with staff in offices in three time zones (Costa Rica, London/IoM and Sydney). A good customer service agent would handle around 100 emails per day, picking up the oldest one from the queue every time they finished one. There was some variability in terms of capacity (agents would handle less if they were more senior or working on specific topics, more if the distribution of emails was more fortunate in terms of simplicity), and response times could vary from under five minutes to hours depending on the match between staff scheduling and email flow. One of the more impactful projects I did during my time there was to optimise that setup - I reduced the average response time from over an hour to under five minutes as a result. The support organisation had a fairly standard hierarchy of seniority. During a shift, if there was anything you could not answer or was concerning, senior members of the team were available via IM to escalate emails/issues to. At the end of each shift, you would send a brief email using a fairly standardised template, noting anything unusual that you came across. In that way, there was an important informational flow from a key interface point with the customer up towards the decision making entity at the top of the organisation, with each layer aggregating and filtering relevant data. Each individual aggregate and filter step was obviously inherently flawed - humans will miss the significance of individual items, or forget things, or over index on something, but a key aspect of that is that it was a known failure mode. No one at the top of the information hierarchy, receiving the highest level summary, assumed it has perfect reliability all the way down the chain. While imperfect, overall it was a good example of a competent organisational learning process and it was extremely effective. PokerStars support was easily the highest regarded service of its type in the gaming space at the time. Many companies have much less visibility and structure - early in my career I mapped a similar process in insurance for a startup, and we often encountered huge chains of cross function information flows where no one had visibility of what was happening at an aggregate level. Individual units were also often entirely incorrect in their understanding of their specific informational role/tasks and those of their process neighbours, yet somehow the systems still worked.

If you think about that support process today, you can envision all the points that can now be handled by AI agents. At the core task level, there are any number of entities offering customer support agents. You could also build your own, using RAG, evals and model routing to customise the responses to your exact requirements while managing inference API costs. The reporting and aggregation could be another agent. It could feed its summaries into an even higher level AI agent, one looking for trends, or mapping product features to automated bug reports and customer complaints. Start to link all that together and you have a nascent company brain. That automation is possible because, at an abstract level, an organisation is an information processing entity. It ingests and emits information flows, and there are also internal flows being created and handled. That’s not a particularly new perspective - it has been an angle that organisational theorists have been exploring for decades. Another way of thinking about this is that the organisation is a kind of state machine. It has a certain state of beliefs and assertions that are the essence of the organisation, and every action and information flow is related to those aspects (and could update them). To use a React/Redux/Flux pattern analogy, a company is a reducer, and the state is updated by actions/methods acting on specific informational events.

The thing to note here is that there are clear advantages to both the human and agentic approaches. Humans are pretty variable. Their memories, skill sets and competence can differ significantly. They can intuit patterns where none exist and can be highly inconsistent. AI agents never tire, and are (probabilistically at least) somewhat consistent. If anything, their issue is their tendency to be too literal, to give too much weight to a specific word or phrase. They also, as Armin Ronacher so eloquently describes it, add entropy to every interaction. For those of us who have been using them extensively as coding agents since the crossing of the competence Rubicon in April 2025, I think the most common take (at least amongst the people I most respect in the space), is that the speed up is real, but to let them loose without supervision or review on a codebase is to invite disaster. Small bits of entropy add up when you are building a wall, and every brick needs to support another.

And I think this is where the clash between the people who see a commercial opportunity to build company brains and practical reality lies. The biggest opportunities (in terms of impact) are by definition with the most complex organisations. Those are the organisations where some sort of company brain agent overseeing everything would massively increase both standards and consistent execution. But those are exactly the organisations where the volume of information causes the entropy from parsing it to become an invisible failure mode. If a company brain is 95% right in unknown dimensions, it is practically 100% unusable. What you see from people (mainly engineers) who have built something in this domain are the same issues as the quantity of data increases. They see drift in accuracy over time. If they have a rigid classification structure, it becomes unwieldy. If they build a flexible classification structure, it descends into chaos. Also, as the sheer volume of nondescript data increases, so the entire process of parsing and categorisation becomes slower (and more expensive). If you ever need to employ new staff to manage your company brain, or it becomes an unloved task for everyone, you end up with the worst of both worlds. Now the sellers of company brains will tell you they can subdivide the agentic tasks to overcome this, staying in the context/intelligence sweet spot. Or that you can throw more compute and inference at the issue by adding review agents or complicated consensus approaches. Or that the next generation of (expensive) frontier agents will fix it. But I don’t think this is solvable that way - you can’t create certainty from probability.

In a human organisational context we’ve been managing amorphous, incompletely defined, business states to more or less degrees of competency forever. It’s a diffuse state, and imperfect, being distributed amongst any number of humans and documents, but the unreliability of the company, and its staff, are effectively the working model. A certain amount of variability is baked in. In small startups there is obviously less issue (and one of their key advantages) - it is easy for one person to hold the entire company state in their head. Note that is also what eliminates the commercial opportunity amongst that class of company. That is not to say there is no value to a company brain for small startups - quite the opposite - an agent effectively becomes the human augmented if you like, able to multiply the capacity of some of the most important individuals.

There are two forces converging here. One is the rapid, possibly exponential, growth of LLM capability, albeit (as Demis Hassabis notes) with very jagged intelligence. The other is the reshaping, and in most cases downsizing, of the human organisational structures which can leverage the new tools for a given unit of work. Ultimately, in terms of having an effective AI company brain, there is basically an equilibrium between the capabilities of agents and the size of the company. Now the capabilities of agents is increasing rapidly, so this equilibrium is changing rapidly too. Jack is quite explicit about reorganising his organisation as a result of this, Satya is somewhat more measured (but it is still inherent to his vision). I think the missing piece here is that the optimal structure for a human information processing organisation is not the same as one in which an AI Agent is the structural lattice. I think an AI agent org is necessarily smaller and operates in different ways, but I think Satya is absolutely correct in that the value is not going to come from the AI side, but from the human side. The value is going to accrue to the organisation where the human intuition and effectiveness will optimise the agentic capabilities. Effectively, AI company brains are going exactly the same way as coding. There they have condensed the value into a very specific human role, which is now extremely leveraged. I think the best, most pragmatic, exponents of AI coding have realised that pure vibe coding is a catastrophic dead end for anything other than a product which can be eventually consistent (effectively only open source). It is the judicious use of (increasingly skilled) human intervention that captures both the speed and breadth values from coding agents, and I believe that with a company brain we will find that the most effective implementations are the ones with the human in the loop. Humans, at the right layers, executing the judgement and skillset an LLM can only approximate badly.

This tension between organisation size and company brain viability, and between the relationship between the human part of a company and an agentic one I think is where the missing half of Tom and Satya’s analysis is. The specific technical aspects of a company brain are interesting but not especially complex. I think they are basically going to be table stakes going forward, because ultimately most companies are already building the component parts in individual sections. The only aspect left is often to draw them together. That’s how we built ours - we started with making the main unpoker rivals codebase work well with coding agents. Then we created an accounting structure and a marketing creative and reporting system. At that point the central state becomes obvious. A feature you build into the product is relevant to the support system, the marketing system and maybe the accounting system. Once you link them up, once you realise that there is basically a master context for any organisation, then it becomes a matter of curation. Extracting that state, documenting it and making it useful is certainly a skill (and one smaller companies have a massive advantage on by definition), but the real value remains in the execution, not the code or the design of the brain per se.

And this is the conclusion the company brain sellers can’t reach. The companies large enough to pay for a company brain are precisely the ones it can’t work for. The companies it does work for are small enough to build it themselves. Indeed, as the agents improve, the size of company that can be run this way keeps dropping, so the organisation shrinks to meet the brain. There is no stable product sitting in the middle of all that. There is just a better way to run a small company, and I think all companies are going to at least resemble small companies. There has been a lot of doom mongering about the impact on AI on jobs. I remain somewhat undecided as to the aggregate effect (although it is abundantly clear now that there will be a significant transfer of capability and responsibility from human to AI in specific domains). What I think is clear, and what Satya probably couldn’t say, is that AI, and the AI company brain, is the beginning of the end of the organisation as we know it.

I started by noting that this trend has been driven from an engineering angle, and that is because AI is increasingly dominant there. Engineers are the vanguard of the AI march and they generally hate gruntwork and documenting. Everyone does! So naturally they have looked into the magic LLM void and seen a saviour - something that can remove all that work and be the nervous system and brain for an organisation, leaving them time to do all the enjoyable work. But this isn’t really a purely engineering problem. It’s a human challenge, and humans are quite a diverse bunch. That’s another thing about brains. My brain isn’t the same as yours, and the Unrational Games brain is not the same as the one you should build for your company. We’re going to open source the core of our implementation soon, but ultimately you need to build something to work with your company in your domain. And crucially, you have to find a way to keep the humans in the loop to get the value, not magically hope the next word guesser understands nuance. When everyone is building for autonomy, we think the solution is augmentation, indeed it is useful imprecision (and a bit of the unenjoyable work).

Philip Atkinson, CEO, June 2026