Confluence for 7.5.26

Fable returns. The downsides of putting agents on org charts. How much time are you spending botsitting? A brief note on AI writing and logic.

Midjourney prompt: A Bauhaus interpretation of AI agents infiltrating an org chart

Welcome to Confluence. Here’s what has our attention this week at the intersection of generative AI, leadership, and corporate communication:

• Fable Returns

• The Downsides of Putting Agents on Org Charts

• How Much Time Are You Spending Botsitting?

• A Brief Note on AI Writing and Logic

Fable Returns

And it means the crossing of multiple thresholds.

The U.S. government reached an agreement this week that cleared Anthropic to once again make available its flagship general-population model, Fable. Others have covered the details of the export control, the security issues, the restoration, and what it means for U.S. AI policy, so we won’t get into that here. We recommend this post from Zvi Mowshowitz. Instead, we want to say a few words about what Fable represents for us, and what we think it might represent for you.

Overall, Fable is a crossing of thresholds. We’ve noted prior moments like this: the release of ChatGPT 3.5, the release of Opus 4.5, the release of ChatGPT 5 Pro. Each left us with the sense that we’d experienced a transition from one phase of generative AI to another. Fable does the same, but the thresholds we’ll note are perhaps a bit more macro than just the technology …

… but one is certainly the technology. Fable is very, very good. Shockingly so. When asked last week by a client to describe “what generative AI isn’t good at,” several of us were for the first time at a loss for words. Writing fiction? But between the general intelligence of the models and their ability to use tools, run code, and read and integrate data, there’s really nothing (at least in our work) where we wouldn’t try to find a way to use AI in the task. Here’s some perspective: we thought ChatGPT 4o could be as good as an intern in our firm. We thought Claude Opus 4.5 could be as good as a junior consultant. We think Fable is as good as someone who’s been in our firm for five or 10 years, perhaps longer. It’s that good.

Another threshold is the gap between what Fable (and the models that will soon rival it) can do and what the vast majority of people who have used generative AI think it can do. The generative AI community calls this gap “overhang.” There has always been overhang: a difference between what generative AI can do and what people feel is possible. But with Fable that overhang is massive. Most people use a free version of ChatGPT to do web-search-level work, and most people have given up on using Copilot at work because they’ve found it not helpful. People deep into using generative AI (not just software engineers but also professionals in many domains, including several in our firm) are using Fable to run teams of virtual agent employees, working on large-scale assignments over five- and 10-hour timelines. It’s a huge gap. And it matters because it means the vast majority of people are dramatically underestimating generative AI and what it will mean for them and their organizations. They have terribly poor assumptions and are likely making equally poor decisions about the future based on those assumptions.

A third threshold is security. Security is what got the U.S. government’s attention with Fable, and the model it is based on, Mythos, is still only available to a limited set of governments and organizations so they may use it to harden their security infrastructure. Mythos is, according to Anthropic’s testing, the safest model they’ve ever created, even if it, without guardrails and in the wrong hands, would clearly be the most dangerous model we’ve ever seen. To date, open-source models have lagged the capability of the frontier-lab models by about seven months, which means we may be five or six months away from the world having access to free models able to do what Mythos can do. Security concerns about generative AI are suddenly real.

A final threshold is cost. Fable is not cheap, and it’s the first model expensive enough that individuals and organizations are already rationing their use of it. Yes, there are cheaper models that can do most tasks perfectly well. But Fable is so impressive that it feels like one is making a meaningful decision about when and how to use it. When it was removed from public use, we felt a loss in what we could do, and we’d only been using it for three days. For the first time, there’s a generative intelligence available so good, and so costly, that it feels like a rationing of available intelligence. We’ve been writing for months that soon the cheap AI party was going to come to an end. Consider it over. And while Fable-level intelligence will become more affordable over time as the labs find system efficiencies, by then a more capable and more expensive model will have come along to replace Fable as the frontier model. Most people and organizations are not used to the idea of price discrimination on the level of intelligence available to them. That day has come.

So, we’re once again in a new phase. But it’s different from earlier phases. In this one, yes, the models have achieved surprising new capabilities. But this phase also brings increasingly incorrect assumptions about what AI can do and will mean, legitimate security risks, and new and likely limiting economics for AI use. This phase isn’t just about cool new things AI can do. It’s about the arrival of the first meaningful macro-level implications of incredibly capable generative AI.

The Downsides of Putting Agents on Org Charts

New research on how framing AI agents as employees changes human behavior.

The idea of putting a piece of software on an org chart, where we map out the roles and reporting hierarchies of people inside an organization, would have been absurd just a few years ago. Today, it’s an idea that many are taking more seriously. This week, The New York Times wrote about companies starting to use AI agents as employees, going so far as to put them on org charts. The piece drew on research from Emma Wiles from Boston University, along with a team from Boston Consulting Group, who ran an experiment to see what happens when managers think they’re reviewing work from an AI employee instead of an AI tool.

In the study, participants were given five documents to review, all with known errors, with a time limit of 20 minutes. Each participant was told one of three things about the documents: they were produced using an AI tool, they were produced by an AI employee named “ALEX-3” (no relation to our leadership AI, ALEX), or they were produced by a person named “Alex.”

Once they finished reviewing, participants had to make choices that mirror real management decisions. How closely do I check this before I sign off? Do I loop in someone else for a second look? If an error slips through, who’s actually responsible for it? And, going forward, how much should this producer be allowed to do without me checking in? Every participant reviewed the same five documents and answered the same questions afterward. The only thing that changed was what they were told about who produced the drafts: a tool, an AI employee, or a person.

The team found when AI agents are framed as employees, managers are more likely to cede responsibility for the outputs (even when charged with reviewing them), shifting accountability to the system. They are more likely to miss critical errors. And they are more likely to rely on additional reviewers before signing off themselves. In other words, people become worse managers just at a time when management skills are becoming more important to working with generative AI.

When we see leaders toy with the idea of adding AI agents to org charts the motivation is largely symbolic. If AI agents are on the org chart, it sends a message about their centrality to how a team operates. But that’s not the only effect such a decision has on the human colleagues of an AI agent. As Wiles notes, “it can change how work is evaluated and how responsibility is allocated.”

The changes this research shows are not ones we’d recommend our readers invite. The important decisions about AI agents have less to do with whether we should put them on org charts, and much more to do with how we handle questions around quality and accountability for their outputs. Spend your time there, and leave the org charts alone.

How Much Time Are You Spending Botsitting?

A new way to talk about a familiar issue.

A new study finds that workers who report saving hours each week by using generative AI are handing more than half of that time straight back. Glean’s Work AI Institute surveyed 6,000 full-time digital workers across the United States, the United Kingdom, and Australia, with researchers from Stanford, UC Berkeley, UC Santa Barbara, Emory, Notre Dame, and University College London, and published their Work AI Index 2026 on June 10. Adoption is reported as very high, at 87%, with 75% of users saying AI makes them more productive. Yet only 13% of organizations report performing significantly better as a result. We’ve written about this time and again, the chasm between what workers say is happening and what companies see at the bottom line.

So what gives? The Work AI Index offers a clever new name for this old-news problem. The missing hours go to what the authors call botsitting, the unbudgeted labor of making AI usable, which is the organizational overhang we describe above, but measured from the worker’s desk rather than the balance sheet. Reloading the same context into one tool after another, comparing outputs because the first answer was not good enough, verifying high-stakes work, and cleaning up the confident wrong answers together consume a reported 6.4 hours a week, more time than workers spend using AI to produce anything. We’ve talked about how a task now demands more judgment and less drafting than it once did. Here is that change in the data: workers spend more of their AI time supervising and repairing the tool than producing work with it.

There is also useful data here on workslop. As the burden grows, people stop checking and pass along output they cannot stand behind. The authors cheekily call this “botshitting” (their term), and 69% of AI users in the study admit to it, from delivering work they could not explain if asked, to knowingly passing along output they believe is wrong. If falling asleep at the wheel is a well-known danger, this one is more like glancing down at your phone though you know it’s not a good idea.

The practical lesson is that adoption metrics measure the wrong thing. A team can report excellent AI usage rates and a workday’s worth of time saved while quietly absorbing most of another workday in making AI output usable, none of which shows up on a bright green dashboard. Our advice remains the same. Leaders ought to reward the judgment behind good AI use rather than the output and ground tools in the organizational context that makes them useful, so that “saved time” is truly valuable.

A Brief Note on AI Writing and Logic

AI-generated writing can be deceptively repetitive. A good editor makes all the difference.

AI writing is often good, even very good, on the sentence level. It’s always grammatically correct, and it tends to use a short, direct sentence structure that’s easy to read. But it can also be deceptively repetitive, especially across longer materials. Read an AI-generated work closely and you’ll sometimes find the model returning to and rephrasing the same concepts in ways that obscure the main idea (or hide the fact that there isn’t a main idea at all). We’ve found that AI does this often in written materials that break up content into subsections (like, for example, strategic communication plans). The model will create sections that are largely repetitive of earlier parts but have enough new content to make them feel like they are adding something new. Simplicity at the sentence level masks bloat at the structural level.

This kind of repetition is an AI-ism that can blunt the impact of even the best ideas, but can be surprisingly hard to put your finger on. Tracking and improving how an argument progresses takes sustained time and attention, and the speed and polish of much AI-generated work make it tempting to simply accept it as is. But all writing is the result of a series of choices, and, as we noted above with botsitting, that is the supervising work now landing on all of us. Does this sentence actually build on the one before it? Is this information I would include if AI hadn’t already included it for me? Is it earning the space it takes up on the page? Often, the answer is no.

We’ve written about how working with AI will increasingly mean shifting one’s focus to reviewing rather than doing, and this is one example of that shift. So we may find that as more and more professionals start using AI to create their first drafts, editing will be at least as valuable a skill as writing. The kicker, of course, is that editing, like writing, takes time to get good at, and it resists the very things that make AI convenient. It rewards care over speed, skepticism over sycophancy, and concision over volume. But if you’re looking to take our advice above and prioritize quality AI use over quantity, ensuring you have a team full of good editors may be one place to start.

---

We’ll leave you with something cool: To keep the endangered dialect of Louisiana French from disappearing, linguists are feeding centuries of Cajun and Creole recordings into their own AI model, teaching a machine to hear a language that lives almost entirely in song and spoken memory.

---

AI Disclosure: We used generative AI in creating imagery for this post. We also used it selectively as a creator and summarizer of content and as an editor and proofreader.