corporate decision-making and the company brain
TLDR: corporate decision-making today is highly inefficient due to human cognitive constraints, and in the future this function might be largely performed by a “company brain” (with human oversight)
limitations of the present
here’s a pretty typical example of how corporate decision-making works at a big company (based on true events):
- CEO of the company needs to make a decision, tasks it to the chief revenue officer (CRO)
- CRO staffs one of his VPs on it, who then has one of her Directors to pull together an analysis
- the Director gets their team together, where each team member analyzes some data, or does some market research, or interviews some people
- the team members put together a presentation for the Director, who provides some feedback based on their understanding of the world & of the preferences of the VP & CRO & CEO (repeat a few times)
- the Director & team then takes the presentation to the VP, who provides some feedback based on her understanding of the world & of the preferences of the CRO & CEO (repeat a few times)
- the VP & Director & team then take this to the CRO, who provides some feedback based on his understanding of the world & of the preferences of the CEO (repeat a few times)
- the CRO then talks to the CEO & CFO, using this presentation as a basis, maybe ask for some more follow-up analyses (go back to step 3, repeat a few times)
- CEO makes a decision, based on his updated understanding of the world
this is a particular solution to an optimization problem, which is just to pick the utility-maximizing decision given available info:
\[\displaystyle\arg\max_{\text{decision} \in \text{options}}\mathbb{E}[\text{utility}(\text{decision}) \mid \text{best model of the world, all available information}]\]on the basis of this formulation, we can discuss some inefficiencies with how corporate decision-making works today:
- “all available information”: everyone in this process has an imperfect subset of the relevant information, and communicating this is slow and lossy
- “best model of the world”: the models largely live in various peoples’ heads, and these heads are pretty constrained in size, and so the model may be quite wrong or lacking in nuance
- suboptimal computation of \(\arg\max\mathbb{E}\): humans are very cognitively limited, and thus can only contemplate a quite limited number of possible actions & potential outcomes
- suboptimal utility model: company executives are bottlenecked on their ability to deeply understand the preferences of their shareholders, and so their model of what utility function to optimize may be suboptimal
- speed: human communication is very low-bandwidth, and this process is highly bottlenecked on communications
the reason that things are the way they are is entirely due to human cognitive limitations:
- if the CEO could just understand everything relevant to the company, ingest all pieces of information, and make decisions that way, he would do so
- but he is cognitively constrained, and must allocate his mental faculties to the highest priority stuff
- this means delegating some tasks to a CRO
- who then delegates things to a number of VPs, and so on and so forth
corporate decision-making might look a lot more organismic
here’s how AIs stack up vs humans on these constituent sub-parts of decision making:
- accounting for “all relevant information”:
- LLMs are already able to ingest and process way more information way more quickly than humans
- they’re also getting increasingly good at integrating this information to derive relevant conclusions (e.g. see how GPT5.2 has been able to solve a number of Erdos problems by finding relevant lemma proofs in past literature & putting them together)
- computation of \(\arg\max\mathbb{E}\):
- machines are parallelizable in a way that humans are not, and so this sort of optimization procedure can be arbitrarily scaled, vs a human who has to do everything sequentially
- speed:
- machine communication can be much higher throughput than human communication
- also, there’s less need to communicate, because the models of the world & utility function aren’t distributed across people
- “best model of the world”:
- here’s a place where humans might still have an advantage: the implicit world models that AIs build may still be bad
- but there’s a significant probability that AI continues improving here
- “utility()”:
- AI may be able to integrate a lot more information about preferences of the company’s shareholders, and thus develop a more coherent notion of what the correct utility function to optimize should be
- but, alignment issue may loom larger here: humans might be better equipped to understand & optimize for the preferences of other humans
on the basis of this, here’s a view of how the future might look:
- the best model of the world + utility function is externalized into a “company brain” (e.g. implemented as an LLM), rather than implicitly living in the heads of a bunch of senior executives
- this company brain is hooked up to every possible data feed the company has access to
- decisions will be made by literally having this company brain think really hard a bunch of times in parallel & then aggregating up the results
- subject to final oversight by humans, at least in the near term
the role of human decision-makers in this world would be largely to guardrail the company brain:
- continually evaluating company brain’s quality & coherence
- checking the decisions made by the company brain against their internal understandings of the world
- (and doing a bunch of tasks not related to decision-making that we’re not going to discuss in more detail)
as in: if the corporation is the right entity for achieving certain outcomes in the world, then the optimal way to implement decision-making for this entity is to centralize its cognition, give it access to data, and have it optimize
the resulting entity might look a lot more like a single coherent organism vs the kind of organizations we have today that are very much agglomerations of humans weakly bound together
some counter-arguments to this view
the core contention above is just that there are various advantages to centralizing decision-making into a company brain vs how things are today, and so we might end up there
but:
- just because something is better doesn’t mean it’s going to happen:
- maybe the improvement in decision-making speed & quality are kind of minor, and so there’s not that much benefit to this company brain approach => limited market pressure to move in this direction
- maybe the market is inefficient, due to e.g. incumbents having large network effects or other moats, and the moat zeroes out any pressures to innovate on corporate decision-making
- maybe it’s not even better
- it’s entirely possible that all the advancements in AI capabilities we’ve seen only apply to strictly verifiable domains (e.g. math & coding), and corporate decision-making is too unverifiable & will remain a human-specific domain
- e.g. maybe AI will continue being hugely sample inefficient, and corporate decision-making is an area that requires making correct inferences from very few ambiguous pieces of information => humans continue being essential
- maybe alignment ends up being really hard, and it thus becomes infeasible to ever hand off decision-making to a company brain, even with extensive human guardrails
- it’s entirely possible that all the advancements in AI capabilities we’ve seen only apply to strictly verifiable domains (e.g. math & coding), and corporate decision-making is too unverifiable & will remain a human-specific domain