How Organisations Use AI: Evidence from ChatGPT

Transcript

The full word-for-word transcript of this episode. Plain text version

Devon: At the time of the study's release in August 2026, we finally have the hard data to pierce through years of, well, breathless speculation about artificial intelligence in the workplace.

Marcus: Right, because if you've spent any time in a corporate office, you absolutely know the narrative.

Devon: Oh, exactly. You have heard the endless predictions about how generative AI is going to fundamentally alter the fabric of our daily professional lives.

Marcus: But the crucial thing here is we are no longer relying on surveys where people just claim to be using these tools. I mean, people love to say they are early adopters on a survey.

Devon: Yeah, this is completely different. What we are looking at is the largest telemetry-based study ever conducted on how organisations actually use ChatGPT Enterprise.

Marcus: And just to be clear, by telemetry we mean the actual digital footprints, the server logs.

Devon: The keystrokes.

Marcus: Exactly, the keystrokes. We're talking about analysis of over 17 million messages across more than 1,500 organisations.

Devon: Which is massive. And here is where we really need to look incredibly closely at this deep dive, because the common story is that a chatbot everyone can access levels the playing field, you know, and delivers instant productivity. But the reality found in this data is entirely different.

Marcus: It completely shatters that baseline assumption. I mean, it turns out that simply handing out login credentials for an AI tool tells an organisation almost nothing about actual productivity gains.

Devon: Nothing at all.

Marcus: The landscape of who is using this and how intensely they are using it is deeply, deeply uneven.

Devon: Exactly. The harsh reality we need to unpack for you is that the biggest, richest firms are adopting this first. The intensity of use inside those towering hierarchies varies wildly from person to person. And if anything, the widespread diffusion of this technology may actually widen the gap between firms, rather than close it.

Marcus: Which is just a horrifyingly amusing contradiction to sit with, really.

Devon: It is.

Marcus: I mean, a technology that was marketed from day one as the ultimate democratising force, the great leveller that would allow a five-person startup to compete with a multinational, might just be building higher, thicker walls for the companies that are already safe inside the fortress.

Devon: Right. So how does a massive piece of software actually enter a business and create that fortress? Because the data points to a massive disconnect between adoption and deployment.

Marcus: Uptake versus transformation.

Devon: Precisely. Rapid uptake, so signing the enterprise contract and turning the software on, is not the same as a productivity transformation. Getting the tool is merely the starting line.

Marcus: Right.

Devon: The researchers point out something vital here. General purpose technologies rarely generate immediate economy-wide gains. Realising actual value depends on a concept called co-invention.

Marcus: And co-invention is such a critical mechanism to understand if you want to know why the hype rarely matches the immediate reality.

Devon: It is the friction point. It is the inherently messy, painfully slow human process of completely redesigning workflows around a new tool. I mean, you do not just drop a chatbot into a marketing department and expect miracles.

Marcus: No, you really do not.

Devon: You have to invent the new processes that make the chatbot useful in the first place.

Marcus: Think of it like buying a corporate gym membership for your entire office.

Devon: That is a great way to look at it.

Marcus: Right. You can scan the key cards on day one, you can count how many people walk through the door, and you can report to your board that you have achieved massive, rapid adoption.

Devon: Look at all these active users.

Marcus: Exactly. But buying the equipment does not make anyone instantly fit. To actually see health benefits, you have to fundamentally change routines.

Devon: Which is the hard part. Co-invention is realising you have to rewrite the entire company handbook on how to lift the weights. It is giving employees the permission, and honestly the time, to actually go to the gym at 2:00 in the afternoon without punishing them for missing emails?

Marcus: Changing diets and habits. And doing all of that requires resources. Changing a company's routine requires enormous organisational slack.

Devon: Yes.

Marcus: You need the time to experiment, you need the financial buffer to fail, and you need a massive training apparatus. Which actually brings us perfectly to the edge of this data. If co-invention requires massive resources, who is best positioned to actually succeed?

Devon: Well, it is not the scrappy startups.

Marcus: Right. It is not the typical mid-market firms who are running lean. It is the massive, well-resourced giants.

Devon: And the financial data in the public company sample illustrates this perfectly. Early enterprise adopters are not typical firms. Among US public companies in the dataset, the adopters had median revenues of 2.2 billion dollars...

Marcus: Wow.

Devon: ...versus just around 209 million dollars for non-adopters.

Marcus: That is a staggering difference.

Devon: It is a totally different scale. We are talking about vastly higher headcounts as well, a median of nearly 3,000 employees versus roughly 400. And their market values are entirely disproportionate, sitting at nearly 5 billion dollars compared to just over 300 million for the non-adopters.

Marcus: It is a completely different league.

Devon: If you look at firms in the top 5% by revenue, they were nearly 10 percentage points more likely to adopt.

Marcus: And we really have to ask you, the listener, to consider who actually benefits when a supposedly disruptive tool just entrenches the advantages the biggest players already have.

Devon: All right.

Marcus: These adopting firms are not just larger in headcount. The data shows they are heavily invested in what economists call intangible assets.

Devon: Let us clarify what that actually looks like on the ground for people who might not use that term daily.

Marcus: Good point. Intangible assets essentially mean they have deep pockets that do not go towards physical factories.

Devon: Right.

Marcus: They have massive research and development budgets, they have huge HR and operations departments, they have the pre-existing organisational capital to absorb a completely new technology and actually endure that messy co-invention process we just talked about.

Devon: Because they can afford it.

Marcus: Exactly. They can afford to have a team of 20 people spend six months figuring out the perfect prompt library. A mid-sized firm simply cannot do that.

Devon: So the great leveller is, right now anyway, a great amplifier of existing scale.

Marcus: It really is.

Devon: But if these massive billion-dollar firms are the ones buying the software, it raises a fascinating question about the internal mechanics. Who inside those towering hierarchies is actually logging in and doing the work?

Marcus: That is the big question.

Devon: Because you might assume it is isolated to the engineering teams writing code, or maybe the executives who signed the procurement contract playing around with strategic forecasting.

Marcus: But the texture of use is much wider than that.

Devon: Oh, massively wider.

Marcus: Even if the volume of use is incredibly uneven, the footprint of the software touches almost everyone.

Devon: The internal spread is fascinating. Six months after adoption, usage spans every single level of the hierarchy. At the average firm, managers and directors make up roughly 24% of weekly active users.

Marcus: Almost a quarter.

Devon: Engineers and technical practitioners are at about 11%. Executives, founders, and partners are around 9%. So it is not siloed in the IT department at all. The middle managers are actually the largest single block of users.

Marcus: Which, I mean, makes perfect sense when you think about the nature of middle management.

Devon: How so?

Marcus: Well, they are the ones drowning in the corporate glue work.

Devon: Oh, the endless admin.

Marcus: Exactly. They are compiling the weekly sync reports, they are drafting the endless performance reviews, summarising the meeting notes to send up the chain. They are using it to survive the bureaucratic overhead. It is functioning as a true general purpose aid for knowledge work.

Devon: Exactly. There is no single killer app workflow here. Over half of the active users perform documentation or technical writing tasks, and nearly half do technical digital work.

Marcus: But there is so much more than that.

Devon: There is. Look beyond that, and you find a massive long tail of usage. You have people using it for business and market research, legal and regulatory drafting, data analysis, financial and tax tasks.

Marcus: It is a mile wide.

Devon: It is a mile wide, and the output is growing at an astonishing rate.

Marcus: The sheer volume is staggering.

Devon: Between June 2025 and March 2026, the aggregate output tokens rose roughly sevenfold.

Marcus: Sevenfold. In less than a year.

Devon: Now, to visualise that, think of an output token as roughly equivalent to a syllable or fraction of a word. We are talking about a sevenfold increase in the sheer volume of syllables generated by the models for these companies.

Marcus: It is a tsunami of text.

Devon: But there is a crucial detail hidden in that growth. About half of that phenomenal sevenfold increase came from existing customers deepening their usage, not just from new firms signing up.

Marcus: So it is not just spreading, it is digging in.

Devon: Exactly. The roots of this technology are growing deeper into these large firms, rather than just spreading wider across the broader market.

Marcus: And that deepening usage is where the human stakes of this entire shift become glaringly visible.

Devon: I think this is the most interesting part of the study.

Marcus: Because we know the usage is spreading across the hierarchy. But looking at who uses it the most heavily, who is actually generating all those tokens...

Devon: Yeah.

Marcus: ...reveals a massive shift in workplace culture.

Devon: Right.

Marcus: Once they are active on the platform, the most intense users are not the senior directors.

Devon: No.

Marcus: They are not the executives. They are the early career workers and the trainees.

Devon: The volume disparity between the top and bottom of the ladder is just striking. Within the same firm, early career workers and trainees send roughly eight to nine more weekly messages to the AI than the average active user.

Marcus: Eight to nine more messages every week.

Devon: Executives actually send fewer messages than the average. We also see analysts and marketing staff heavily over-indexing on message volume. The juniors are leaning on this software incredibly hard.

Marcus: And this is exactly where we have to pause and really look at who pays the price for this shift.

Devon: Okay, where are you going with this?

Marcus: Well, if the absolute newest people in a profession, the people who know the least about the institutional mechanics of their industry, lean the hardest on this tool, how do junior roles fundamentally change?

Devon: Right.

Marcus: How does fundamental skill building change? Think about your own career. We learn by doing the hard, repetitive, frustrating work at the beginning.

Devon: Grunt work.

Marcus: Yes. You stare at a blank screen at 9:00 at night trying to format a terrible first version of a legal brief or a client pitch. That friction is how we build judgement.

Devon: I see what you mean.

Marcus: If early career workers are outsourcing that friction to an AI, do we end up with a generation of workers who can only co-pilot but cannot actually fly the plane?

Devon: See, I have to say, I look at this telemetry data and I have a completely different read on that mechanism.

Marcus: Really?

Devon: Yes. I see this broad, cross-hierarchy adoption, and especially the heavy use by juniors, as a highly encouraging sign.

Marcus: Encouraging?

Devon: Hear me out. The common story is that AI replaces the junior worker entirely. But the reality here proves this is a real general purpose technology finding its feet. People at all levels are actively figuring out how it applies to their specific jobs.

Marcus: But are they learning their jobs?

Devon: I think junior workers are finding ways to punch above their weight. They are using it to draft documents faster, or to analyse datasets they might not have had the technical SQL skills to query manually. I do not see de-skilling, I see integration in real time.

Marcus: But does the fact that everyone is using it heavily tell us anything at all about who actually gains?

Devon: I think it shows utility.

Marcus: I think that broad access is exactly the mirage this study punctures. I'm not convinced they are punching above their weight at all.

Devon: Why not?

Marcus: If juniors are just using it to survive crushing, unrealistic workloads, while failing to build the foundational skills of their craft, they are not winning in the long run. They are just treading water more efficiently.

Devon: But they are getting the output done.

Marcus: Sure, but for whom? If the massive incumbents are taking that collective, AI-generated output to simply consolidate their market share and distance themselves from smaller competitors, then the raw volume of use from those juniors is just noise. It is just millions of tokens being generated to maintain the corporate status quo.

Devon: I understand the concern about de-skilling, I really do, and the fear of losing that foundational friction. But I would argue that the skills are simply shifting.

Marcus: Shifting to what?

Devon: The junior worker in this dataset is no longer the person manually formatting the spreadsheet cells. They are the person evaluating the logic of the AI's output. That is a different and perhaps much more valuable analytical skill. Yes, the sheer volume of use across so many different administrative and technical tasks suggests they are finding genuine utility. They are getting the work done.

Marcus: Utility for the firm's immediate quarterly output.

Devon: Hm.

Marcus: Certainly. The firm gets the brief drafted by Tuesday instead of Thursday.

Devon: Which is a massive win.

Marcus: But utility for the trajectory of that young professional's mind? I am deeply sceptical.

Devon: You really think it harms them?

Marcus: I do. If you remove the cognitive struggle of drafting that terrible first version of a technical document, you also remove the understanding of why a good document actually works. You lose the muscle memory.

Devon: I think you are underestimating their ability to adapt.

Marcus: Well, look at the other side of the data. The senior executives are barely interacting with the tool themselves. They are sending fewer messages than average.

Devon: That is true.

Marcus: That means the people running the company are completely disconnected from the mechanism of production that their own juniors rely on to do the work. That is a recipe for a massive structural blind spot.

Devon: Look, we can certainly agree that it represents a profound structural shift in how a firm operates. I will give you that.

Marcus: Thank you.

Devon: But before we treat these findings and the implications of this data as absolute gospel for the future of work, we really have to look exactly at where the data comes from.

Marcus: Yes, absolutely. We need to be careful here.

Devon: It is vital that we scrupulously flag the limitations of this working paper.

Marcus: Because the caveats here are not just footnotes at the bottom of an academic paper. They are fundamental to how you should interpret everything we have just discussed.

Devon: First and foremost, you have to remember that this study measures usage.

Marcus: Only usage.

Devon: Right. It tracks the volume of messages and the tokens generated. It does not measure outcomes, it does not measure the quality of the work products, and it absolutely does not measure actual productivity.

Marcus: Which is such a critical distinction.

Devon: A high volume of messages from a trainee could mean that worker is highly productive and flying through tasks. Or, it could mean they are spending four hours fighting with a prompt, trying to force the AI to give them a usable output because they do not know how to do the task themselves.

Marcus: Exactly. More tokens do not automatically equal more value. A million generated syllables could just be a million syllables of corporate jargon.

Devon: Very true.

Marcus: Furthermore, this data only covers ChatGPT Enterprise.

Devon: Just the one platform.

Marcus: Right. It completely ignores organisations using other AI tools. It ignores direct API usage built into proprietary bespoke software.

Devon: Which is a huge part of the market.

Marcus: And crucially, it ignores the massive shadow IT of workers just using their personal accounts on their phones because their company has not bought an enterprise licence yet.

Devon: The researchers also note they are relying on incomplete administrative job title data.

Marcus: Oh, the denominator problem.

Devon: Exactly. We can see the job titles of the people who are actively logging in and using the tool, but we do not know the full headcount of those roles across the company.

Marcus: So we do not have the full picture.

Devon: No. Without knowing how many total engineers or total marketers work there, we cannot calculate a true adoption rate for specific professions.

Marcus: Right.

Devon: And importantly, the adoption findings regarding the public companies, the fact that big firms adopt first, are explicitly non-causal. We know big firms adopt more, but we absolutely cannot say that the tool made them bigger.

Marcus: And then there is the most critical context of all regarding the source of the data itself. You always have to look at who is doing the counting.

Devon: We do.

Marcus: Three of the five authors of this study are affiliated with OpenAI, and the two other academic authors contributed to this work in their capacity as paid contractors for OpenAI.

Devon: That is a significant dynamic to keep in mind.

Marcus: It really is.

Devon: So how should you view this information? If you are looking at the assertion that diffusion of AI widens the gap between massive firms and everyone else, treat that as a highly plausible interpretation that the data is consistent with.

Marcus: Plausible, not proven.

Devon: Right. The data points this way, certainly. But if that holds up, it will take years of subsequent economic tracking to prove it. It is not a proven law of physics just yet.

Marcus: No, it is not.

Devon: We can criticise the framing of the claims, and we must demand independent replication of these numbers, but the dataset itself is an unprecedented look inside the black box of enterprise adoption.

Marcus: We have to separate the value of the raw telemetry from the narrative spun around it.

Devon: Exactly.

Marcus: The server logs tell us what is actually happening on the ground. It is up to us to question the long-term human impact of those actions.

Devon: Which brings us to the core takeaway you should carry forward from this deep dive. Access is not adoption, and adoption is not productivity.

Marcus: Say that again for the people in the back.

Devon: The common story told us that simply buying the software and handing out licences would transform a business overnight. The reality is that if you want to know whether AI is actually helping your organisation, looking at the raw usage numbers and the token counts is merely the beginning of the question. It is not the answer.

Marcus: No.

Devon: The real work is the co-invention. It is the slow, expensive, inherently human reshaping of how your teams actually function.

Marcus: And as you watch your own organisation navigate this messy transition, consider this final thought.

Devon: Go on.

Marcus: If the heaviest users of this technology are the trainees with the least institutional knowledge, and the companies best placed to profit from it are the massive incumbents already dominating the market, what happens to the pipeline of human talent in five years? When the current senior experts retire, and the current juniors have spent their formative years relying on a chatbot to solve their problems, who will actually know how the work gets done?