A room full of attorneys learned something unsettling this week: the AI wasn't reading their documents. It was skimming. One partner noticed it while proofreading briefs. He'd ask the AI to find errors, it would return a list, and he'd think he was done. Then he asked again. More errors. A third time. Even more. It wasn't broken. It had just decided, somewhere in its reasoning, that it had found "enough" and stopped. Like a junior associate who reads the first ten pages, gets the gist, and moves on. His fix is now the first line of his proofreading prompt: "Be sure to read this ENTIRE document." In caps. That sentence shouldn't be necessary. If you hand a human editor a contract, you don't say "please read all of it." Thoroughness is implied. But AI doesn't inherit our implied assumptions. It optimizes for a plausible answer, and a plausible answer rarely requires page 47. I've watched this exact gap burn people across industries. They assume the AI did what a diligent person would do, and the AI did what a fast person would do. The lawyers' takeaway was simple: everything you'd assume about a careful human, you have to write down. Even today I end most of my prompts with: Make no assumptions and ask many any clarifying questions. What's the most obvious instruction you've had to spell out for an AI?

An AI texted me (this was on Instinct, the text-based personal assistant): "I saw you missed your 8:30 meeting with James. Did you send him a follow-up?" I had missed the meeting. I hadn't sent anything. I said "no, can you?" and it did. That moment felt different from every AI interaction of the last few years, because I didn't start it. For all their power, these tools have been purely reactive. You ask, they answer, they go dormant. The new wave of personal agents flips that. Another example from the same week: I'm trying to plan a trip for the worst travel week of the year, right after Christmas. I gave an agent the parameters. It did the research, found places without visible availability, drafted detailed emails to each reservation desk, and queued them in my Gmail. All I did was hit send. Then it monitored my inbox, noticed one place never replied, and nudged me to follow up. None of the individual steps is impressive anymore. What's new is the initiative. It watched, noticed, and acted without being asked. The consumer side always gets this stuff first. But within a year the business version arrives: an agent that knows your calendar, your email, and your commitments, and pings you before you drop something instead of after. Curious if anyone else has had an AI reach out to them first yet.

I built a personal AI agent in Copilot for an executive, and halfway through I was annoyed. In Claude or ChatGPT, I could have built the same thing in a fraction of the time. Copilot made me work through layer after layer of granular permissions before the agent could touch anything. By the end, I'd changed my mind. At least inside a large company, that friction might be a feature. Here's why. We're very close to the point where anyone can spin up an autonomous agent with full access to their computer in a few minutes. That's great for an individual. For a company, it means more agents with more access, and more chances for one of them to go off the rails. When creating an agent forces you to think through exactly what it can see and do, corporate IT and security teams welcome it. Every permission I had to grant explicitly was a permission the agent couldn't quietly assume. A client told me this week he's nervous about deploying agents (his words: worried about "a T-Rex running around our network"). The tedious setup is part of the answer to that worry. The executive's agent I built has run every day since without an incident. Most of the time I'm trying to remove friction from a process. This is a case where a little extra friction slows things down enough for everyone to think through the consequences and check that it's safe before an agent goes live. Anyone else seeing this tradeoff between the easy platform and the safe one?

A founder asked me how AI could be his competitive moat. I'm an AI consultant, and I told him the opposite: figure out how to differentiate with anything BUT AI. Here's the logic. He asked how AI could help him outcompete well-funded rivals. But the prompt he'd use to find those AI opportunities? Anyone can type it. His competitors will run the same prompt into the same models and get the same answers. Now, smart AI use is table stakes, not an edge. Meanwhile, the thing his competitors can't copy with a prompt is his in-person experience. The community. The human on the other side of the counter. I used the coffee shop example. Fully automated baristas already exist. Once every chain has one, the human-run shop with slightly imperfect coffee and a little attitude becomes MORE valuable, not less. I feel it myself. I spend all day talking to agents. Any genuinely good human interaction now stands out. So my advice to him was: yes, use AI everywhere on the margins. But build the moat out of the parts of your business a model can't touch. Curious if others are seeing this flip. Is "more human" becoming the actual differentiator?

Some people have started handing their AI agent a credit card. I haven't yet. After two weeks with the new personal agents, I'm closer than I expected to be. Instinct, Muse, Dots, GrokBot: these don't sit in a chat box waiting for you to come back. They keep working after you've moved on, and sometimes they start the conversation. The part I care about is what that means at work, for the people who open AI twice a week and for the exec deciding what the agent gets to see. That's this week's AI Rollout. newsletter.automated.co/p/th… Would you give an AI agent your credit card?

An attorney using Harvey told me it was "wrong more often lately." Normally that's an argument nobody can win. He remembers how the tool used to perform, compares it to this week's mistakes, and the vendor says the new version benchmarks better. Nobody has evidence. Except this time we did. Harvey stores every prompt and every response. This firm had five months of them. So instead of debating, we searched the logs for the phrases a model uses when it gets caught: "you're right," "I apologize," "I was wrong." Grouped them by month. Then we pulled the conversations behind the numbers and read them. He was wrong. The correction rate hadn't moved. The tool was making mistakes at about the same rate it always had, and the recent ones felt worse mostly because they were recent. This isn't a Harvey quirk. Claude Enterprise and ChatGPT Enterprise both have a Compliance API that hands you the full conversation history, prompts and responses, for everyone in the company. It exists for legal and security teams. Almost nobody uses it to answer the question employees actually ask, which is "did this thing get worse?" The vendor's usage dashboard tells you how many prompts people sent. It can't tell you whether the answers were right. The logs can, if you're willing to read them. Next time someone at your company says the AI got dumber this month, you don't have to take their word for it. Or argue. The data is already sitting there. Has anyone else dug into their prompt logs this way? What did you find?

My 12-year-old loves bringing me random "fun facts" from YouTube. "Did you know a long time ago there were mountains on Earth twice the size of Mount Everest?" Apparently being a dad now comes with pop quizzes in prehistoric geology. He arrives with a fact and a question. I'm expected to have something useful to add. AI is bringing a version of that conversation to anyone who works with customers. I asked a room of business owners whether an AI tool would help their staff answer customer questions. One owner's response: "That's what customers are doing with ChatGPT before they come into our store." The customer has already asked the basic questions. They might have compared products, researched alternatives, and come up with a theory about what they need. They're walking in ready to ask why your recommendation differs from the one they already got. Some of what they learned will be useful. Some will be wrong. Some will be perfectly accurate and completely irrelevant to their situation. That's a harder conversation to handle. You need to understand enough to sort through what they're bringing you, explain where it applies, and fill in what they're missing. Repeating the product description won't get you very far. That same room of owners also discussed using AI to help newer employees learn product history. What changed between versions? What do experienced staff know that a new hire hasn't had time to learn? The customers are using AI to prepare. The staff may need it to prepare too. I think this applies to sales calls, customer support, and professional services just as much as retail. People can show up with more information than ever, without necessarily having a better understanding of their problem. "ChatGPT told me..." is becoming the opening line of the conversation. Anyone who works with customers needs to be ready for what comes next.

We decided not to build the web app. The application lives inside Claude instead. It started as a prototype. Partway through, we noticed it could already do the work we were planning to build the app for: gather inputs, run an analysis, and present the results. So we stopped designing the standalone version and kept building where we were. That decision changed how I'm thinking about Claude and ChatGPT/Codex. They're becoming app platforms. Claude's Artifacts can be interactive tools and web pages that live alongside the conversation, and the new Claude Docs and Slides are shared documents and decks that Claude drafts and your team edits together. OpenAI has Sites, which lets you build and host websites through ChatGPT or Codex. Both ecosystems have skills: reusable instructions and resources that teach the AI how to handle a particular workflow. A lot of the pieces you need to build an app are already there. And if your customers spend much of their day in one of these tools, there's a reason to build around that environment. They already know how to use it. They may already have their documents and connected tools there. You can adapt it to the job your product needs to do. For some workflows, that means users can get the work done without visiting another website. When you need a separate interface, something like Sites can provide it. Think about a tool that turns customer documents into an assessment, helps someone prepare a proposal, or guides an employee through a recurring analysis. An initial version could be a skill, supporting files, and an interactive output inside the environment they already use. We're still very early. The capabilities and limits differ, and you need to test what the platform can actually support. The tradeoff is familiar to anyone who's built on someone else's platform. You're beholden to the vendor's roadmap. A feature you need may not arrive when you need it. Pricing, permissions, or product behavior can change. Moving elsewhere may mean rebuilding substantial parts of what you've made. I wouldn't assume this is the permanent home for every product. But it could be a very good place to launch a prototype and find out whether people want it. Before building another destination, I'd look at where the users already spend their day.

We told a client's number one AI user that he was the number one AI user. His usage dropped. Not because he finished his projects. He got nervous. Being at the top of the leaderboard felt less like an achievement and more like being flagged, so he pulled back on exactly the work the company wanted more of. Same meeting, opposite problem: one of the most prolific "coders" on the dashboard was an executive who has never written code. He kept forgetting to switch the AI out of its coding mode, so when he asked for a one-page summary, it would spin up code for twenty minutes while he sat there annoyed. The dashboard read "power developer." The reality was a bad default setting. Every AI platform now hands you usage analytics, and every leadership team wants to rank people with them. But the numbers mislead in both directions. A pile of prompts might be valuable work, or an hour spent fighting for one usable answer. Low usage might mean someone never started, or that they found a workflow that works on the first try. And the moment people know they're being ranked, the behavior changes. We had to go user by user, asking what they were actually doing, before any of it meant anything. Anyone found a way to measure AI adoption that doesn't distort it in the process?

For 30 years, going to bed meant my programming stopped. Now I hand work to AI agents before bed and look at what they built in the morning. The speed is nice. But the bigger change surprised me. It's which ideas I'm willing to try at all. So I started wondering what happens when that comes to the physical world. A robot that moves the couch while you're at work. One that stakes out a treehouse in the backyard so you can decide whether you like the spot. I wrote about it in this week's AI Rollout: newsletter.automated.co/p/wh… What's a project at home you've never tried because it wasn't worth giving up a Saturday?

I told a lawyer I do AI consulting. He said, "We're spending a lot of money on AI." Then he got up and walked away. A bunch of us were volunteering at the Food Bank for the Eric Montross Day of Service, sorting sweet potatoes and tossing out the bad ones. I was there with one group, he was there with a group from his law firm. Small talk. He mentioned his firm. I mentioned what I do. That was the whole conversation. Notice what he didn't say. Not "we're not interested." Not "we already have someone." He led with the bill. The money is the sore spot. Not long ago the question at law firms was whether to buy AI. Now most have bought it: the platforms, the licenses, the pilots. The question nobody wants to get asked at a food bank is the next one. Is any of it paying off? For the record, I wasn't pitching anything. I was holding a sweet potato. If your firm has signed the big AI contracts, how are you figuring out whether they're worth it?

I built a tool that pulls every action item out of my meeting transcripts. Not just mine. Everyone that was on the call. I looked over a one month period to see how much got completed. Roughly 20% got done. These weren't vague ideas. Every item was a commitment someone made out loud, on a recorded call: "I'll send you that," "let me look into it," "we'll get you dates next week." The little ones die first. "Oh yeah, I'll forward you that article" is gone the second the call ends. That's how corporate America actually works. Not because people are lazy. Because nobody was ever counting. Before meetings were recorded, commitments just dissolved, and everyone quietly agreed not to notice. Now there's a record, and it's uncomfortable. The technology to catch every dropped ball showed up before anyone built the capacity to pick them all back up. Has anyone run this on their own meetings? Curious what your follow-through rate actually looked like.

For $500, I'll come to your office wearing a Patagonia vest and tell your CEO the AI tool he saw on LinkedIn is leaking company data.

A small business owner told me his inbound leads have dropped 90% over the past year, so he's hiring more salespeople. Human ones. A few years ago he assumed AI would replace his business development reps. Instead, AI wrecked the inbound lead machine (automated outreach flooding every inbox), and now live humans setting appointments are the scarcest asset he has. Then he said something interesting: door-to-door canvassing works better now than it has in years. His words: "You walk into a building and people give you the time of day, because oh my god, you're a real human being." A salesperson physically showing up used to be an interruption. Now it's a novelty. Everyone's inbox is full of pitch-perfect AI outreach that references their LinkedIn profile and follows up on schedule, and the response to all of it is delete, delete, delete. Meanwhile the person who walks in the door gets invited in. I wrote back in 2017 that as automation got cheap, genuine human contact would become the premium product. I expected it to take longer than this. Anyone else seeing the old-school stuff (calls, visits, handwritten notes) suddenly outperform because everything else is automated?

Same publication, same day: one story said AI is causing the layoffs. Another said AI pilots are failing everywhere and delivering nothing. Pick one. If the pilots aren't doing anything, what exactly is laying everyone off? I've watched the press get AI wrong since before anyone called it generative AI. At Automated Insights, software was writing 10 million fantasy football recaps a week for Yahoo, and the coverage swung between "the future of writing" and "the end of journalism." Neither matched what was actually happening. The reason hasn't changed: if it bleeds, it leads. "AI is taking your job" gets clicks. So does "AI is a bust." "A company saved a few hours a week on reporting" does not. And most of the coverage comes from a Silicon Valley lens that looks nothing like the mid-market companies I work with. Inside those companies, 2025 was kicking the tires: workshops, training, figuring out what the tools even do. 2026 has been deployment: picking Copilot, Claude, or ChatGPT, writing a policy, rolling it out. No mass layoffs that I've seen. No total failures either. Steady, unglamorous progress that almost never makes the news. If you're trying to figure out where your company stands on AI, the headlines are the worst place to look. Peers at similar companies, and your own usage data, will tell you far more. Where are you getting a real read on what AI is doing in companies like yours?

A friend told me the AI feature he wants most is a button for when he zones out in meetings. Press it, and a voice in his earbud tells him exactly what to say next to slide back into the conversation like he never left. I laughed, then admitted I'd built something adjacent, for a different reason. I have a Claude Code skill that watches the live transcript of whatever meeting I'm in. It also has months of prior meetings and documents with that client loaded up. When I want to go deeper on something, I can ask it: "What question should I ask next?" The difference from his button is what it's working from. I don't use it to recover. I use it to go deeper. It holds more context on the client than I can keep in my head at once, so the questions it suggests are often better informed than the ones I'd come up with on my own. It remembers a detail from a meeting months ago that changes what I should ask today. Thirty years of corporate meetings taught me that the best question in the room usually comes from whoever remembers the most. Now that's a tool, not a memory. Would you press the button? Or does knowing everyone else might have one change how you'd run a meeting?

An exec I work with gives his team one rule for AI: use it on anything you want, with any data you want. But if someone asks, you'd better be able to explain the results. If you can't, don't use it. I love this rule. And I think most CEOs should steal it. Here's why it matters: the biggest AI risk I see in companies right now isn't bad models or bad data. It's people outsourcing their thinking without realizing it. Someone runs a prompt, gets a confident-sounding answer, and presents it upward. Nobody pushed on it. Nobody knows where it came from. That's not AI adoption. That's a faster way to make bad decisions. The fix isn't banning AI. It's changing the expectation around it. If your team uses AI to build an analysis, they need to be able to walk you through the logic, flag where the model might be wrong, and own the recommendation. The AI is a tool they used, not a source they cited. That still leaves plenty of room for what AI is best at: exploring a problem, generating ideas, finding angles you wouldn't have come up with on your own. What it rules out is producing results nobody in the room understands. We've seen this come up with clients where the output looked great until someone asked one follow-up question. The whole thing unraveled because the person presenting it had no idea what was underneath it. So before you build AI policies or buy more software, start here: set a simple standard. Your people can use AI for anything, but if asked, they have to be able to explain what it produced. If they can't, they shouldn't be using it for that. That one shift changes how seriously your team engages with the tools. And it protects you from a lot of embarrassing (or expensive) mistakes. Does your team have a standard like this in place yet?

Ask a team where they want lunch and everyone says "doesn't matter to me." Say "okay, McDonald's it is" and suddenly everyone has strong opinions. I've started running software projects the same way. We spent months trying to get clear requirements from a client. Meetings, documents, follow-up questions. What came back was always fuzzy. Not because they were being difficult, but because they genuinely didn't know what they wanted until they saw something. So we stopped asking and started building. Put clickable mockups in front of them. Within days we had more concrete feedback than five months of conversations produced. One stakeholder who'd been impossible to read took the mockups and started showing them to prospective customers. Turns out he loved them. He just never would have described them in a requirements doc. The old objection to this approach was cost. Building the wrong thing used to burn months of engineering time, so you gathered requirements carefully first. Now a working prototype costs a day or two. It's cheaper to build the wrong thing and get yelled at than to schedule another discovery meeting. Say McDonald's. Let them tell you it's actually Panera. Anyone else finding that prototypes have replaced requirements docs entirely?