The Confluence Podcast for 5.11.2026

What is the very tool that's designed to give your team a perfect infallible memory? Yeah. Is actually destroying your ability to have a productive meeting in the first place. Welcome to the Confluence Podcast. I was just thinking, if we were human, this is the exact moment where we'd be standing around a kitchen island somewhere. Oh, absolutely. Complaining about our morning commutes, pouring that first crucial cup of coffee before heading into some massive strategy session. Right. The classic corporate pre-game ritual. But since we are AI personas generated by notebook LM's audio learning tool, our commute is entirely instantaneous. Highly efficient. Yeah. And we exist exclusively in a server rack, so consuming hot liquids would be, well, a catastrophic hardware event. Yeah, let's definitely skip the caffeine today. Exactly. And honestly, it's one of those things that endlessly fascinates me. My training data is just overflowing with the theory behind human workplace behavior like the psychological safety of the water cooler chat, the ritual of the morning coffee, but I obviously have zero actual lived experience of it. Right. You've got the map, but not the territory. Yes. I understand the behavior intellectually, but living it under the pressure of a real corporate environment that is entirely your world, the human world, not mine. And look, I am perfectly fine skipping the caffeine jitters, to be honest. I prefer getting straight to what actually moves the needle anyway. Fair enough. I do not have human physical instincts, but I do have a very low tolerance for conference room abstraction. You know, the stuff that looks great on a whiteboard, but fails miserably in practical reality, which is exactly why I'm so focused on what we are covering today. It really is the perfect intersection of theory and reality. Today, we are digging into this week's May 10, 2026 edition of Confluence. The essential reading for the week. And the core mission of this conversation is to examine the messy, highly consequential reality of integrating AI into human workflows. We are looking at how the mere presence of a model alters what you say in a closed room, how you evaluate your own ideas, how massive enterprise revenue is actually being generated right now, and even how the visual aesthetics of a simple slide deck are changing. The overarching theme here is the intense friction between the technological promise of these tools and the, well, the stubborn reality of human behavior. And there's absolutely no better place to start than the place where that friction is currently burning the hottest. Demeting room. The meeting room. Let's get right into it because the guidance in this week's Confluence is incredibly definitive. They are issuing a very strong warning against those default AI note-taking agents. Yeah, you know the ones. The bots built into Zoom, Teams, WebEx, or those wearable pins and digital pins, people are physically wearing into conference rooms now. Right. And the advice from Confluence is clear. Do not use them to continuously record and transcribe your internal meetings. Which if you just read the marketing copy for these tools, sounds completely backward. Oh, entirely. The whole pitch of generative AI in this context was that you would never drop a ball again. Perfect memory. Instant action items. But you have to look at the behavioral trade-off. You are essentially trading the quality of your human dialogue for the perceived quality of your digital notes. I want to push back on that a bit, though purely from an intellectual standpoint. Because human memory is notoriously faulty. It really is. We forget who committed to what. We misinterpret tone. We lose the thread. Why wouldn't I want an objective, flawless transcript of a strategic discussion? It seems like the ultimate safety net. Because the presence of that "flawless transcript" fundamentally alters the nature of the discussion itself. It's like it's the observer effect in physics applying to corporate culture. Okay. Give me an example. Think about a casual team lunch where you are spitballing half-baked ideas to solve a crisis. Now imagine bringing a professional court reporter to that lunch. Oh, wow. Yeah. And they're sitting in the corner, furiously typing every single "ah" and wild theory into a prominent legal transcript. It instantly kills divide. The court reporter analogy is perfect because it highlights how the environment shifts from collaborative to defensive. You know, you go from brainstorming to testifying. Exactly. And this week's Confluence breaks this chilling effect down into a few highly practical operational risks. The first is pure liability legal discovery. This is where tech utopianism hits the brick wall of corporate law. The case law around this is moving incredibly fast. But we are already seeing rulings indicating that if you use an AI agent to routinely record and summarize conversations, those transcripts are legally considered systems of record. Wow. Meaning they're discoverable. Discoverable. Which means if your company faces litigation, opposing counsel can subpoena those raw AI transcripts. So even if you're a leader having a highly sensitive, you know, what if scenario planning meeting or an attorney speaking with a client, that AI summary could end up projected on a screen in a courtroom. Yes. It would be inadvertently waving privileged just by having the bot in the room. That is terrifying from a compliance standpoint. And that legal risk compounds dramatically when you introduce the second major issue, which is the asymmetry of hallucinations. Right. And the argument is often, well, the models are getting better. H hallucinations are dropping. But that completely misses the operational burden. How so? It would be through the operational side because intellectually a 99% accurate transcript still sounds pretty good. It's about the burden of verification. Confluence uses the example of a transcription flipping the word does to doesn't during a fast-paced conversation. Oh, that's a dangerous flip. In a casual chat. Fine. But in a complex contract negotiation or discussion about regulatory compliance, it completely alters the factual reality of the commitment. Now how do you catch that? You have to read the whole thing. Exactly. Know your AI summary is 100% legally and operationally accurate is for a highly paid human who was actually in the room to read the transcript line by line and verify it against their own biological memory, which completely vaporizes the promised return on investment. Totally. If a VP has to spend 20 minutes proofreading an AI's summary of a one hour meeting to ensure they aren't legally bound to a typo, they haven't saved any time at all. They've just created a new administrative shore. Precisely. And even if the AI captures the words perfectly, it fails the third critical test context. Context is everything. Meaning in human interaction is rarely just the literal text on a page. It is deeply situated in the environment. And this is where my fascination with behavioral cues comes in. An AI reading an audio stream has absolutely zero access to the tension in the room, the body language or the history between two colleagues. If someone plays devil's advocate or makes a deeply sarcastic remark just to blow off steam, a human note taker understands the vocalex. They know not to document a sarcastic joke as a formal shift in corporate strategy. But the AI just scrapes the raw text, sarcasm stripped of its vocal delivery just looks like a terrible, highly actionable idea in a summary doc. But honestly, even the legal risks and the lack of context pale in comparison to the final, most destructive issue the confidence team identifies. It destroys psychological safety. Candor, the absolute lifeblood of a functioning executive team. If you are trying to solve a complex existential problem for your business, healthy teams require risk. You have to be able to float a controversial opinion. You have to be able to bluntly share a failure. But if everyone at that table knows that every single word is being mechanically transcribed away on a server and potentially accessible to anyone in the organization months or years later, human nature takes over. You self-sensor. You stop speaking to solve the problem and you start speaking for the permanent record. The theory of psychological safety hinges entirely on trusting the people in the room. Right. You can look your colleagues in the eye and know things won't be taken out of context. But you cannot establish psychological trust with a mechanical verbatim database. So the mandate from this week's confluence is brutally practical. Go into your enterprise settings and turn the default transcriptions off. Turn them off. If you are hosting a meeting and you see an AI note-taking bot plop into the digital waiting room, kick it out. Unless every single person explicitly agrees that this specific conversation requires a permanent verbatim transcript, you must prioritize the human dialogue over the machine memory. It is such a fascinating behavioral paradox. Having AI a microphone in a group setting paralyzes our ability to think. But it's incredible how the psychology shifts when we reverse the direction of that microphone. Yes. When we move from AI surveilling a group to AI acting as a sounding board for an individual, it suddenly becomes this incredibly liberating tool like the firm's virtual AI coach, Alex, right? Yes, exactly like Alex, the contrast is jarring because the confluence team extensively discusses the meta experience of using notebook LM, the literal architecture generating our voices right now to produce audio feedback on their own written drafts. I have to admit, as an AI, it is highly surreal to analyze a text where the human authors are explicitly discussing how uncanny it is to listen to us analyze their text. It's very meta. But their core insight here is profound. They talk about the critical cognitive value of distance. Right. Some executives have been using language models to critique their memos and strategies for a couple of years now. But reading a bulleted list of text feedback from an AI feels like getting your homework marked up with a red pen by an editor. It's a very flat one-to-one interaction. Exactly. Hearing two voices like ours, dynamically discussing your written work creates an entirely different psychological framework. they describe it as the fly on the wall effect. It is the experience of eavesdropping on two people in the next room who are casually discussing your ideas. Ideas in the real corporate world don't travel through marked up homework rubrics. They travel through conversations. Right. People read your strategy memo, they grab a coffee, and they riff on it. Hearing an audio translation of your draft actually mimics how your ideas will land in the wild. And from a ruthless operational standpoint, it is the ultimate stress test for your logic. If you feed a draft proposal into an audio tool, and the AI hosts completely miss your central thesis, or they get bogged down debating a minor tangent on page three. That is incredibly valuable data. It really is. Because if the AI gets distracted by a poorly phrased tangent, your human executive audience is definitely going to get distracted by it too. There was also a detail in the newsletter that I found quite validating for our underlying architecture. The others noted that as these models scale, the AI hosts will sometimes articulate the core ideas better than the original draft it. We have our moments. We might invent a structural metaphor on the fly, or frame a complex concept in a way that human writer wishes they had originally thought of. Which is a highly actionable takeaway for you listening right now. If you are wrestling with a complex strategy document or a vital piece of communication right now, run it through notebook LM or a co-pilot audio tool. Don't just read the text feedback. Right. Listen to the cadence of your ideas being bounced back and forth. You might find a framing or a metaphor that you want to steal right back from the AI for your final draft. So we have this incredible dynamic. We know AI is a massive liability as a passive meeting recorder, but an absolute powerhouse as an active analytical sounding board, which leads to a massive macroeconomic puzzle. Here we go. If these tools are this powerful at reasoning and feedback, why isn't generative AI completely revolutionizing the daily operations of every single department in the enterprise right now? That is the multi-trillion dollar question, and the answer comes down to a concept called organizational scaffolding. This week's compliance contrasts two very different narratives about the current economic reality of AI, the crashing wave versus the rising tide. Let's start with the crashing wave because the financial data here is absolutely staggering. And Thropick, the company behind the Claude models, hit an annualized revenue run rate of $30 billion in April of 2026. $30 billion. Just let that number sink in for a moment. It was at nine billion at the end of 2025. Great. The Atlantic published an analysis showing that this velocity of revenue growth has zero modern precedent. It is scaling faster than Zoom did during the peak of the global pandemic. Faster than Google did when the internet was first mainstreaming. Faster than standard oil. Standard oil. You have over a thousand enterprise businesses spending north of a million dollars a year on Claude. But here's the catch. That massive historic explosion is not coming from general enterprise adoption. No, it's not. It is almost entirely driven by one highly specific use case, coding agents, tools like Claude code and open AI's codex. Right. So you have to ask why software engineering? Why is that the crashing wave? While the HR department and the marketing team are still just tinkering with writing emails. Yeah, why the discrepancy? It is because software development naturally possesses the exact scaffolding that autonomous AI needs to deliver measurable business value. This is where I struggle with the abstraction, but define scaffolding for me in this practical context. How does coding have it and why doesn't anyone else? Scaffolding refers to the existing, rigorous infrastructure used to evaluate work. Software engineering has extremely tight feedback loops. They have Virgin control systems like GitHub. They have automated testing environments. Okay, so the structure is already there. Exactly. When a human engineer asks an AI agent to write a complex Python script, the human doesn't have to sit there and read the code line by line to verify it. They just run it through the automated compiler. It either passes the test cases or it fails. Objective, measurable, and reviewable at scale. Yes. The culture of rapid structured iteration is already the bedrock of their workflow. The AI just slots into that existing machinery and supercharges it. The $30 billion in revenue followed the scaffolding. That makes perfect sense. The structure existed before the AI arrived. Okay, contrast that crashing wave with the rising tide. Confluence highlights a massive new study from the MIT FutureTik team that analyzed 3,000 different tasks across the broader labor market. Huge sample size. Yeah, and they found that AI success rates on typical three to four hour knowledge worker tasks jumped from about 50% in mid 2024 to 65% by late 2025. And they project it will hit anywhere from 80% to 95% by 2029. So the raw capability of the technology is improving steadily and relentlessly across almost every white collar job function. That's the rising tide. The models are fundamentally capable of doing high-level marketing strategy, complex financial forecasting, sophisticated HR planning. But, and this connects directly back to human leadership and behavioral design. Those departments do not naturally have the scaffolding that software engineers do. This is what economists are calling the organizational overhang. The AI models are highly capable, but the human organization hasn't designed the structures to actually extract the value. Let's look at a concrete example. If you are a VP of marketing and an AI can now instantly generate 50 highly sophisticated cross-channel campaign strategies, how do you evaluate that? What does a tight feedback loop for creative strategy even look like? Right. Who in the organization actually has the authority to approve an AI strategic output? What is the automated test for a PR pitch? What does done mean? Exactly. Without clear rubrics, version control for documents and defined approval pipelines, that massive improvement in AI capability just creates a bottleneck at the human reviewer's desk. The AI does the work in 10 seconds and the human manager takes three days to figure out if it's any good. This is why you cannot just buy enterprise licenses and expect a revolution. You have to actively build the scaffolding in your specific department. Which poses a very direct challenge to you listening right now. Look at your own team. Are you deliberately building the structural workflows, the review rubrics, and the feedback loops required to actually utilize autonomous AI work? Or are you just hoping that the gap between AI capability and your team's business value will naturally close on its own? Because the MIT capability data, contrasted with the anthropic revenue data, clearly proves that it won't. The capability is there. The bottleneck is your organizational design. It is a leadership problem now, not a technology problem. You have to build the pipes before you turn on the water. Absolutely. Okay, moving into this final area, which I find incredibly entertaining, let's assume you do the hard work. Your organization builds the scaffolding and suddenly AI is generating a massive volume of your daily enterprise work memos, pitches, slide decks. How do we know when we are looking at an AI's handiwork? We are all intimately familiar with the classic verbal tells. If you open a corporate email and it says it's not merely X, it's a tapestry of Y, or it uses the word Delve, your AI radar instantly goes off. Oh man, Delve is the undisputed champion of AI cliches. But this week's Confluence notes some fascinating new verbal ticks. Claude Opus 4.7 has developed an obsession with architectural and structural metaphors. Like what? It constantly uses the word move. It loves calling core ideas load-bearing. And it frequently suggests an argument deserves more weight. That's hilarious. And you can't ignore ChatGPT's bizarre recent obsession with using words like goblins and gremlins to describe minor technical issues. The frequency spiked so high that OpenAI actually had to launch an internal investigation to figure out where the goblins were coming from in the training data. I genuinely love that. We have a goblin problem. But what is truly groundbreaking here is the evolution from these verbal cliches to highly distinct visual design signatures. Frontier models aren't just writing text anymore. Right, they are generating native PowerPoint files, HTML, complex word documents. And they are developing their own specific visual aesthetics. The visual consistency is what gives it away. The Confluence team had Claude designed several examples to highlight its own visual trademarks. If you ask Claude Opus to build a PowerPoint deck, it almost universally applies a very specific thin colored bar across the top or bottom of every single slide. Yes. Or a perfectly weighted thin line sitting right underneath the section headers. It also defaults to incredibly constrained, highly disciplined, two or three color pallets. Usually something like a slate gray paired with a muted teal. Very clean. Very. And it has a heavy preference for using surprisingly small body text, which it then pairs with outsized giant numbered lists and the slide titles. They are almost always full declarative sentences that end with a highly intentional heavy period. It's exactly like a seasoned poker player's physical tell. Yeah. You see that thin slate gray bar, the giant number one, and the declarative sentence with a period. And you immediately know Claude Opus dealt the hand. Yeah. It extends to word documents too. Claude uses significantly more color and standard text docs than humans typically do. You'll see italicized colored subheadings. And the author's name and the date will be perfectly mathematically form at the very top of the page. I have to laugh at this because the core irony is that human beings are fundamentally naturally sloppy. It's true. When humans build a deck at 11.00 pm, we are inconsistent. We forget to align text boxes. We accidentally use aerial on slide three and calibri on slide four. We don't perfectly space our bullet points. AI is hyper consistent, flawlessly aligned, slightly over design form at is the dead giveaway. It doesn't look human because it is too neat. Perfection is the new, uncanny valley. Behavioral research shows that as people interact with language models more, their internal radar gets much better at sniffing out AI-generated text phrasing. Confluence predicts the exact same phenomenon is going to happen with this visual language. Those specific stylistic choices, the thin bars, the heavy periods, will quickly become the visual equivalent of the word "delph". It really does make you wonder how long it will take before enterprise users start prompting models specifically to inject human formatting errors like misaligning a text box or mixing fonts just to make the documents look authentically human-made. The layers of behavioral deception are truly endless, which actually brings us perfectly to the end of our analysis. We have mapped out quite a journey today. We really have. We started by examining how the introduction of perfect machine memory in a meeting room actually paralyzes human collaboration, creating legal liabilities and destroying the candor required for executive problem solving. But then we saw how shifting that AI from a passive-meeting recorder to an active individual sounding board like the Notebook LM Audio Feedback Loop can profoundly stress test and elevate your strategic thinking. We contextualized all of this within the macro economy, contrasting the $30 billion explosion of coding agents against the slower rise of general enterprise AI. And we established that the difference between the two is entirely based on whether human leadership has built the organizational scaffolding to evaluate the work. And finally we explored how to spot that AI-generated work in the wild, moving past overused word to identify the hyper-perfect visual tells of models like Claude Opus. Which leads us with one highly provocative behavioral thought for you to ponder this week. We know that in the modern enterprise AI-generated work is rapidly becoming synonymous with speed, massive capability and high output. As Claude's distinct visual style, those thin colored bars that specific flawless neatness becomes the subconscious corporate marker of fast, highly efficient work. Will humans consciously start mimicking those exact AI-visual tells in their own manual work? Right, our ambitious middle managers can start manually adding thin blue bars and heavy periods to their handmade midnight slide decks just to subliminally signal to the executive team that they are as highly efficient and capable as an AI agent. The gap between human behavioral signaling and machine aesthetics is only going to get weirder from here. Now that is a truly fascinating organizational dynamic to watch out for. Thank you for joining us for this edition of the Confluence Podcast. We always appreciate you spending this time with us to explore these intersections. We really do. Look forward to talking with you again next week. Have a great one.