What do big-company CEOs want to know about AI? A group I sat with last month acknowledged AI’s transformative potential, but doubted they could send all their people on one long coffee break and replace them with agents. So they asked: what does the person do, and what falls to the machine? How do you move a fractious organization to adopt and get value from AI?
The US military defines a weapon system as not just the hardware, but “a combination of one or more weapons with all related equipment, materials, services, personnel, and means of delivery and deployment required for self-sufficiency.” US Army rifles like the M-4 and M-16 assume disciplined, well-trained forces; Mikhail Kalashnikov designed the AK-47 for an indifferently trained, poorly-educated conscript army. [9]
No less true in automating (or applying AI to) business domains than in equipping infantry platoons. Somebody has to use the bloody system in some way, and they may not want to use it in the way you want them to. Salespeople may not want to enter all their relationship information. Service techs may want to spend more time fixing customer problems and less time entering data on what they did. Doctors may disregard recommendations from a clinical support system.
Business or technology transformation initiatives are simple on Planet Milton Friedman. We all do rigorous strategic, operational and technology analysis. We formulate a business case. And if it plausibly exceeds the hurdle rate, senior leaders adopt it and implement the material incentives for everyone in the organization to prosecute it ruthlessly.
Frustratingly and wonderfully, we live on Planet Earth, not Planet Milton Friedman. Managers have other objectives and constraints in addition to maximizing their bonuses. They have different risk tolerances. Measuring contribution is hard, especially for more strategic and transformative initiatives. So client executives who want to achieve big things must be savvy political operators as well as astute business strategists. In the technology context, I framed this as CIOs needing to be demon game theorists to achieve the cooperation they need for success.
Getting any complicated bill through the US Congress requires careful coalition management. Add this provision and you pick up a few votes tied to one lobby or another. And that provision you pick up a few more votes tied to a second lobby, but you might lose a few too because you have disadvantaged some regional interest or another. And, take care, add too many expensive provisions in hopes of picking up votes and you might lose deficit hawks who want to keep the total ticket under a certain number.
We’ve all gone through this in constructing business cases. Exempt one local market from consolidation to avoid a political fight and the benefits decline more than the investments. Accept more conservative savings assumptions to get operations on board and the ROI attenuates further. The CFO insists on smaller investments in the first year. Figuring out a value-creating direction is hard. Assembling a ROI-positive path through political realities to get there is brutal.
Very little of the information required to determine how humans and machines should interface or how to assemble a politically-viable, ROI-positive business case appears in any database. Even when written down it’s ambiguous and stored in dozens of different spreadsheets, email threads and chat streams and, now, videoconference transcripts.
Ashwin Gopinath, Co-founder and CEO of Sentra, [2] wants to make the information substrate that busineses need to make complicated, multi-dimesional decisions legible. AI won’t solve the problem for you, but it make make the required information legible enough so that you can. [3]
Remember: ProsaicTimes never endorses companies or technologies. But we find insight in talking directly to builders about the future they are trying to create.
A Polymath’s Unplanned Path
In this section:
Ashwin describes his journey across physics, biology, and AI
How following interesting problems rather than a career plan connected disparate fields
Why the most interesting people rarely have a predetermined map
James Kaplan: Hi there, this is James Kaplan with another ProsaicTimes video podcast. With me I have Ashwin Gopinath — why don’t we start out with you introducing yourself a little bit, talking about your journey, and telling us how you came to Sentra?
Ashwin Gopinath: I started in one area and got pulled into working with biomolecules, because that was the only way I could get to the molecules I cared about. That opened two pathways at Caltech.
Then some personal moments led me to get interested in biology as a systems problem — something I could connect to a sense of social responsibility. That pulled me into a whole other set of work. And then I realized what was happening: I was chasing the unifying thread, not the discipline.
James: Well, the most interesting people never have a plan. Sometimes we create artificial divisions between knowledge domains. The physicists will say it’s all physics, and those of us who are history majors will say it’s all history.
Ashwin: You are only as good as your context. The way one person can connect things is remarkable. But the problem with teams is that people don’t know what they’re working on in aggregate. The collective has no privileged vantage point — each person knows only their slice. If you know 100 people, not all of them know what the others know.
James: Alan Kay [1] said that context is worth 20 IQ points, or context is worth 50 IQ points, depending on which version of the remark you’ve heard. History is context. And it is interesting to think about human context versus machine context, because all of us are operating in a professional environment, swimming in context and never quite able to manage it all. There are a zillion emails, a zillion chat messages, a zillion documents to read, and we can never quite assimilate it.
Ashwin: That’s precisely where the friction lives. The CEO, the C-suite — they believe they know what the organization knows, but they only know their slice of it. They may know their own knowledge, perhaps. But in practice, they don’t know what others know. That generates enormous friction.
James: Have you ever read anything by Peter Turchin? He overstates his point a little, but his theory is that asabiya — an Arabic term — is the discriminant between successful and unsuccessful societies. It means common purpose and willingness to sacrifice for that common purpose. But one definition I’ve seen is knowing what you’re supposed to do without being told — everyone has enough shared context that the next action becomes obvious, and everybody moves left or moves right without having to be directed.
Ashwin: Some of that gets called culture in any meaningful sense, or people describe it as a shared operating system. But the thing is, there are these ephemeral artifacts that some groups have — and the question I keep coming back to is: can we use AI to start building that? Can we get groups into that state faster?
James: You could argue that the humanistic aspects of AI are the most interesting ones. There’s an age-old question about how you generate what the military would call unit cohesion — without engendering groupthink. That is a tough issue. People have probably struggled with it from the beginning of time.
Ashwin: That is part of what led me to work on Sentra. The important problem — what Sentra is trying to do — is to be there and make sure everyone arrives at the same page. The question is how you sustain that as things move forward.
Context, Memory, and AI Alignment
In this section:
The challenge of aligning human and machine context in organizations
Human memory, utility functions, and lossy compression in AI context windows
Why structure in memory aids both programmatic analysis and human recall
James: It’s not only a hard problem — different motivations make it harder or easier depending on the environment. On an old-fashioned equity trading floor, decades ago before electronic trading, you could align people because there was a P&L at the end of the day. You had clear risk parameters and a clear profit-and-loss statement on the book. If you made money, that was good; if you didn’t, that was bad. But most of business isn’t like that.
Even where you have quantitative metrics, you may have multiple conflicting ones — this went up, but that went down. Or metrics that are overdetermined: this happened because you did this and I did that. How much was you versus me? Or you may have things that are hard to measure, ambiguous data. If you’re in the consumer world, measuring customer satisfaction is easy, but in a corporate world where you may have 50 different stakeholders in a relationship, how do you measure satisfaction?
Ashwin: When people think about utility functions in AI, they’re making the point that you define the behavior — and if you define the behavior poorly, it gets defined for you. The whole behavioral-economics school of thought is bound up with this definitional challenge: what does the model think it is optimizing, and what are we actually asking it to do?
The biological approach, if you think of the utility function as a kind of lossy compression, is that you compress away what you don’t believe is important. But the interesting thing is, if you need everything, you’re not running the compression correctly — you’re just storing.
James: When you say memory, do you mean memory in the context window, or memory stored to a file — which may be a graph?
Ashwin: I mean both, but I lean toward the latter. I think the ability to go to 100 million tokens is meaningful, but what matters is whether you can actually reach what you need at inference time.
James: Here’s why I’m skeptical of the context window. Thirty years of hanging around data centers tell me the constraint is never the compute — it’s always the memory. And as I read the charts, the compute is improving in capability more quickly than the bus to the memory is.
The context window is constructed using lossy compression [5] — which is obviously not the same as anything you write to a file. Anything involving lossy compression is, at some level, going to be unreliable — the question is how unreliable. And the beautiful thing about system memory as opposed to human memory is you do have infinite long-term memory if you mediate it with structure. [4]
Capturing Meetings and Extracting Structure
In this section:
How Sentra approaches capturing unstructured professional conversations
Identifying primitives — decisions, risks, issues — from fast-moving dialogue
The challenge of determining whether something is one decision or two
James: Now, I think you’re doing some really interesting stuff around capturing conversations. To me, this is one of those things that is incredibly difficult and deceptively important, because in a professional environment, we spend a lot of time in meetings — and then we walk out and say, my God, what just happened?
Ashwin: At the top level, you have the external artifacts — the Slack messages, the emails, the decks, the reports, the minutes. But underneath that, there’s a bunch of people having a discussion. And beneath that there’s something else: a set of primitives. Those primitives may be decisions, risks, issues, deliverables, questions, work streams — it will vary by business context.
What we’re trying to do is use the semantics available from the meeting itself, combined with the semantics you can get from the organizational graph — who are the visitors, the teams, the stakeholders — and pull out those primitives reliably. So the question we’re always asking is: what is the smallest meaningful unit of work or commitment that came out of this conversation?
James: That sounds like an object model. Do you have inheritance and polymorphism in there at all?
Ashwin: Yes — there is a hierarchical structure, and entities do carry properties inherited from their context. The graph structure lets you traverse those relationships.
James: The thing I always struggle with is: is one thing one thing, or is one thing two things? If you’re thinking about redesigning an IT organization — is the decision to consolidate across business units the same as the decision to optimize? You could argue they’re two discrete decisions. You could optimize in place without consolidating. But consolidating without optimizing seems pointless, and optimizing in place without consolidating gives you no scale effects. Theoretically, two decisions; practically, one. How do you think about that?
Ashwin: The way we handle it is probabilistic rather than deterministic. If the semantics of two threads are entangled — if resolving one consistently changes the frame of the other — we surface that dependency as a relationship in the graph rather than forcing a merge or a split. The human gets to see that these are linked and decide whether to treat them as one or two.
James: Ultimately a person or a group of people make a decision, sometimes very ambiguously. I tend to think of those as two separate and incredibly complicated problems. First, what is the set of decisions that have to be made? And second, what is the group of humans that have decision rights? Those rights can be very complicated — you might have negative control but not positive control. You can prevent an answer but not force one. Some people have influence only if three others also agree.
Ashwin: The act of making those dynamics legible changes them. The moment you surface who has influence over what, the political landscape shifts. That’s the observer effect applied to organizational behavior.
James: It’s Heisenberg’s uncertainty principle as applied to AI. [6]
Ashwin: And this is why I think the right model is the one you described — Team Cyborg. AI surfaces insights that inform human judgment; it doesn’t replace the judgment, because you can never capture information completely enough, and because the act of capturing it creates second-order effects.
James: Here is why it matters in a very practical sense. Every business case is a trade — at a big enough scale, it’s like getting something through Congress. I’ve been involved in multi-hundred-million-dollar business cases where there’s always a question: do we leave this country in or out of the consolidation? How aggressive do we estimate the savings here versus there? You’re always balancing the quality of the business case — what does the ROI look like? — against the quality of the political coalition. Smart senior executives do that calculation in their heads without thinking about it. Past a certain point, it gets hard to do in your head.
Ashwin: Political science has been mapping political dynamics for decades, if not centuries, and the models exist. [7] The challenge is applying them in real time to a specific organization with incomplete data. That’s where the second-order effects become dangerous — you don’t want to make the map so visible that people start playing to the map instead of doing the work.
James: This is why I’m on Team Cyborg, not Team Android — AI is a tool that provides insights to inform human decision-making, but we will never, or not in the foreseeable future, capture information perfectly enough to make instinct and judgment irrelevant. The act of capturing some information creates second-order effects.
AI Accuracy, Knowledge Work, and What Comes Next
In this section:
What “95% accuracy” actually means in different professional contexts
The distinction between being wrong at design time versus run time
How AI will reshape the toil and grunt work embedded in knowledge work
James: Where does this all go? Give us your few-sentence expression of where this issue — capturing and transacting with ambiguous data — will be in the next two or three years.
Ashwin: I think everyone believes AI is going to automate everything. It’s not going to automate everything — but I do think people’s work will change. What I find interesting to say is: if AI is 95% accurate on 100 tasks, you’ll know the 5 it got wrong. But if AI is working on 10,000 tasks at even a lower accuracy, the net volume of errors grows — and you have to think very differently about where you put humans in the loop.
James: I would put it slightly differently. First — if you’re running a hedge fund and you’re right 95% of the time, you’re a multi-billionaire. If you’re running an institutional brokerage processing trades and you’re right 95% of the time, you’re either bankrupt or in court. So 95% means very different things in very different contexts.
Second, it’s very different to be wrong 5% of the time at design time versus run time. If you’re wrong 5% of the time at design time, fine: we have quality assurance processes, we can test, we can adjust. But depending on the application, being wrong even 5% of the time at run time can be completely unacceptable.
Ashwin: And I think what’s important — and often under-appreciated — is how much toil and grunt work is embedded in most knowledge work: writing things down, recording things, organizing things. [8] AI is very good at absorbing that toil. Not eliminating the work, but absorbing the mechanical overhead so that humans can do the judgment-intensive part.
The quality of what the agent produces is going to be bound by the quality of the context it’s given. The response from a large language model is only as good as the prompt, and the context is what constructs the prompt. Getting that context right — at an organizational level, not just a session level — is the whole game.
James: All right, Ashwin — context is the whole game. Thank you.
Ashwin: Thank you, James.
Organizational context is a brutal problem of structuring information — decisions that were made but never documented, commitments that live in someone’s memory, dependencies that exist because of a conversation nobody recorded. AI may help us capture and understand: who decided what, when, under what assumptions, with whose sign-off. Done well, this will empower human decision-makers exercising judgement rather than replace them.
Footnotes
[1] Alan Kay massively influenced the development of the Mac via his work at Xerox PARC, and later was an Apple Fellow, but was not on the original Mac team. The “context is worth 80 IQ points” formulation (the number varies across tellings) is widely attributed to Kay from his OOPSLA keynotes and ACM lectures.
[2] Sentra describes itself as a “unified memory layer for your company” — capturing decisions, commitments, and organizational context into a queryable graph accessible to both humans and AI agents across 200+ business tools. (The 200+ figure is per Sentra’s own marketing; no independent verification available at time of publication.)
[3] Ashwin’s Company Brain post argues that most organizations confuse having data with having memory — data without temporal and semantic structure doesn’t tell you what the organization actually decided, or why, or when it changed.
[4] Ashwin’s analysis of how LLMs and biological memory fail the same way makes the case that both systems apply lossy compression, discarding what they judge unimportant at encoding time. The implication: if precision matters at retrieval, you need a different storage path — structured external memory, not the compressed context window.
[5] The arithmetic illustrating why memory bandwidth, not compute, is the binding constraint for LLM inference at scale — is worked through in Why LLMs Will Never Remember Your Enterprise.
[6] The intuition that measuring a system changes the system has a long pedigree in social science. The classic example is the Hawthorne effect: people often change their behavior simply because they know they are being observed. The analogy to Heisenberg’s uncertainty principle is metaphorical rather than literal — in physics, the principle describes fundamental limits on measuring physical quantities; in organizations, the comparable insight is that observation itself can alter the behavior being measured. Management scholars have recently formalized this analogy: Shelef, Wuebker & Barney (Academy of Management Review, 2024) argue that business experiments do more than reveal information — they can change the value of the underlying idea by influencing competitors, customers, and other stakeholders. A companion commentary in the same journal explores both the strengths and limits of the Heisenberg metaphor.
[7] The academic research connected to this point: Gailmard & Patty’s survey of formal models of bureaucracy — covering hierarchy, delegation, and information asymmetry across decades of political science literature — establishes the toolkit Ashwin is referencing. A foundational result (Crawford & Sobel 1982) demonstrates that complete information-sharing between agents with divergent interests is mathematically impossible; the degree of distortion is a function of how far apart their preferences are. Annual Review of Political Science 15 (2012): 353–77.
[8] Knowledge work contains substantial coordination and administrative overhead. McKinsey estimates that knowledge workers spend more than one-quarter of their time searching for information rather than applying their expertise. Cohen & Birkinshaw argue in Harvard Business Review that knowledge workers could reclaim roughly one day per week by eliminating, delegating, or outsourcing low-value activities.
[9] C.J. Chivers covers the history and design philosophy of the AK-47 in depth — the rifle was deliberately engineered for soldiers with minimal training and limited mechanical aptitude. Reliability under adverse conditions, not precision, was the design priority. Helpful comparison to the M-16 here.
