This is Part 2 of the Five Myths About AI Transformation series. Last week we covered Myth 1 - You Don’t Need an AI Strategy, and Microsoft's Copilot reality check. This week's story is messier, funnier, and more instructive.

Let's break it down.

The Signal:

Klarna's AI Experiment Went Exactly How You'd Expect.

In 2023, Klarna stopped hiring. By 2024, they'd partnered with OpenAI, cut roughly 700 customer service jobs, and replaced them with AI agents. CEO Sebastian Siemiatkowski told the world that "AI can already do all of the jobs that we, as humans, do." They claimed $10 million in savings. Headlines everywhere.

By early 2025, customers had a different opinion. Satisfaction dropped. Complaints piled up. Internal reviews revealed the AI couldn't handle nuance, empathy, or the kind of problem-solving that people with billing disputes actually need. Customers described the responses as generic and repetitive.

Siemiatkowski admitted it: "We focused too much on efficiency and cost. The result was lower quality, and that's not sustainable."

Then it got worse. Klarna couldn't rehire fast enough. Reports surfaced that the company started pulling software engineers and marketers out of their actual jobs and into call centers to fill the gap. The CEO who said AI could do everyone's job was now reassigning his technical staff to answer phones.

The solution that would have worked from the start: AI handles the routine stuff, humans handle the complex stuff, and you design the process before you pick the tool.

That's not a headline. But it's what works.

The Scale:

Myth 2: AI Is the Technology That Changes Everything

Reality: The biggest wins still come from boring, proven technology applied to the right problem.

Klarna didn't have a technology problem. They had a sequencing problem. They applied the most sophisticated tool available to a workflow they hadn't redesigned. That's Substitution on the SAMR scale. Swap 1 thing for another. Same process. Same structure. Just cheaper. Until the quality collapsed.

This pattern is everywhere. MIT's "GenAI Divide" report found that companies were pouring more than half their AI budgets into sales and marketing tools. But the highest ROI came from somewhere nobody was looking: back-office automation. Document processing, compliance, internal workflows. The boring stuff.

Deloitte's 2026 State of AI report put a number on the gap. 74% of organizations hope to grow revenue through AI. Only 20% are actually doing so. The companies in the 20% didn't start with the flashiest use case. They started with the most friction.

Sangeet Paul Choudary made this case in Harvard Business Review last month. He studied Shein vs. traditional fashion houses. Traditional companies use AI to speed up design sketches, but the underlying structure stays sequential. They made the season faster. They didn't rethink the season. Shein reduced the unit of work from the seasonal collection to continuous small-batch experiments. AI doesn't sit on top as a fancy addition. It connects sensing, decision-making, and action into a single learning loop.

That's not AI changing everything. That's a company rethinking how it works, then using technology to make the new structure possible.

Stephen Andriole observed this back in 2017 in MIT Sloan. Most short-term impact comes from conventional operational technology. Not from the hot new thing. Think about Uber. The technology behind Uber wasn't exotic. Phone in your pocket, GPS, mobile payments, a well-designed app. All proven. The innovation was applying it to a business model problem.

Donella Meadows wrote that the most powerful place to intervene in a system is at the level of information flows and delays, not at the level of new components. If information moves slowly through your organization, if decisions stall because data lives in 7 different spreadsheets nobody reconciles, adding AI is like putting a turbocharger on a car with flat tires.

Fix the tires first. Then talk about turbochargers.

The Deep Dive:

This week's deep dive goes all the way in on Myth 2, including a practical sequence for finding the boring wins hiding in your business and the math that shows why they matter more than you think. Read: AI Gets the Headlines. Boring Technology Gets the Results.

I WANNA GO DEEP!

Thanks for reading!

Next week: Myth 3. "Profitable Companies Are Best Positioned for AI." Why comfort is the enemy of adaptation, and why the companies riding high are the ones most likely to get left behind.

If you know someone chasing AI when they should be fixing their plumbing, forward this. They probably won't thank you. But their customers will.

Hit reply and tell me: what's the most boring fix that produced the biggest result in your business?

See you next Friday.

Best,

JT

P.S. — Read the full Five Myths Paper on Substack.

This week I published Five Myths About AI Transformation. If you've been reading along the last few months, you've seen me pull at these threads already. AI fatigue. Feature stuffing. Organizational readiness. The gap between what tools can do and what companies are actually prepared to use them for. The paper ties all of that thinking together and goes deeper, connecting the patterns I've been seeing in the field to the systems thinking and business theory that explain why they keep repeating.

Starting this week and for the next 5 Fridays, I'm pulling each myth apart in its own issue. Deeper research. More context. Specific things you can do about it.

We're starting with the myth that got the whole paper started: Every company needs an AI strategy.

Let's break it down.

The Signal:

Microsoft's Copilot Reality Check Landed. Hard.

In January 2026, Microsoft finally revealed the number they'd been sitting on for 8 quarters. 15 million paid Copilot seats. Satya Nadella called it "record AI momentum." He said Copilot is "becoming a true daily habit."

The number he didn't volunteer: Microsoft has 450 million commercial Microsoft 365 users. 15 million is 3.3%.

After 2 years on the market. After $37.5 billion in capital expenditure in a single quarter.

Microsoft Spent $37.5 Billion on AI. Only 3.3% of Users Are Paying for It.

It gets more interesting. Recon Analytics surveyed 150,000 enterprise users in January. 70% initially preferred Copilot because it was already in their Office apps. After trying alternatives, only 8% kept choosing it. Copilot's paid subscriber share dropped 39% in 6 months.

Even Microsoft is walking it back. Reports surfaced that the company is pulling AI features from Windows 11 where usage doesn't justify the investment. SemiAnalysis put it bluntly: an outside competitor shipped a better AI experience on Microsoft's own application than Microsoft's $30-per-seat product could deliver.

The Scale:

Myth 1: Every Company Needs an AI Strategy

Reality: Most companies need to fix their foundation before AI can help.

Microsoft's Copilot adoption numbers aren't an AI failure. They're a foundation failure. Companies bolted Copilot onto dirty data, broken processes, and workflows nobody had examined. The tool worked. The systems underneath it didn't.

This pattern is everywhere. MIT's Project NANDA reviewed over 300 AI deployments in 2025 and found that 95% of enterprise AI pilots delivered no measurable impact on profit and loss. Between $30 and $40 billion in enterprise investment. Almost all of it producing zero return.

The core issue wasn't the models. MIT found that enterprise AI systems didn't retain feedback, didn't adapt to context, and didn't fit into real workflows. A CIO told the researchers: "We've seen dozens of demos this year. Maybe 1 or 2 are genuinely useful. The rest are wrappers or science projects."

Gartner put a number on the data problem. 63% of organizations either don't have or aren't sure they have the right data practices for AI. Their prediction: through 2026, organizations will abandon 60% of AI projects because the data isn't ready.

RAND Corporation measured an 80% overall failure rate. But here's the number that should change how you think about this: projects with sustained executive involvement succeeded 68% of the time. Projects that lost executive sponsorship within 6 months? 11%. That gap isn't technology. It's commitment.

This is the Substitution trap. Dr. Ruben Puentedura's SAMR model maps it precisely. Substitution swaps 1 tool for another with no functional change. That's what most AI adoption looks like. Swapping ChatGPT into a broken workflow and calling it transformation.

Sangeet Paul Choudary made this case in Harvard Business Review last month. He studied Figma vs. Adobe. Adobe did everything right. Moved to the cloud. Subscription model. Collaborative features. Still lost. Because Adobe changed the delivery mechanism. Figma changed how design work was organized. Substitution vs. Modification.

Donella Meadows wrote that you can't improve a system you can't describe. A Microsoft survey of 500 enterprise decision-makers across 13 countries found that only 22% strongly agreed their organization has clearly documented key processes and data dependencies.

If you're in the other 78%, the first honest question isn't "what's our AI strategy?"

It's "can we describe, in detail, how our business actually works right now?"

If the answer is no, start there. The AI can wait.

The Deep Dive:

This week's deep dive goes all the way in on Myth 1, including the practical steps for what to do instead and the business theory that explains why this pattern keeps repeating. Read: You Don't Need an AI Strategy. You Need to Know How Your Business Actually Works.

I WANNA GO DEEP!

Thanks for reading!

Next week: Myth 2. "AI Is the Technology That Changes Everything." Why the biggest wins still come from boring, proven technology applied to the right problem.*

If this landed, forward it to someone who's about to spend 6 figures on an AI strategy without documenting their processes first. They'll thank you later.

If you disagree, hit me up. I’d love to hear your thoughts. That's what fills my cup.

See you next Friday.

Best,

JT

P.S. — Read the full Five Myths Paper on Substack.

You open your project management tool on a Monday morning. You've got a deadline, three fires, and a client who's been waiting since Friday.

​But before you can do anything, there's a banner. A big one. "Introducing AI Summaries, AI Task Sorting, and AI Workflow Suggestions!" Fourteen new features. A guided tour you can't skip. And your actual work is now buried under three layers of menus that moved since last week.

​You didn't ask for any of this. You just wanted to assign a task.

​If that scenario made your eye twitch, you're not alone. And if you're a product leader who just shipped something like this, we should talk.

​This week, I look at the five-stage feature stuffing playbook, three filters that separate signal from noise, and a framework I first used at Apple that tells you whether a tool is worth the money before you spend it.

Let's break it down.

The Numbers Are Real. The Strategy Isn’t.

Here’s what’s happening. SMBs are adopting AI at a pace that nobody predicted five years ago. Fifty-eight percent already use generative AI. Ninety-six percent plan to adopt it. In Canada, that number is even higher. Microsoft found that 71% of Canadian SMBs are actively using AI and GenAI tools right now.

Those numbers are real. The demand is real.

But here’s where it goes sideways. Tech companies see those adoption stats and hear one thing: “Put AI in everything.” So they do. Every button gets a copilot. Every text field gets a summary. Every dashboard gets a prediction nobody asked for.

IDC calls the winners “companies that focus on pragmatic use cases that are easy to deploy and deliver measurable ROI.” Read that again. Pragmatic. Easy to deploy. Measurable ROI. Not “fourteen features in a press release.”

​The gap between what buyers need and what builders ship is getting wider, not narrower. And both sides are paying for it.

Here’s the thing that makes this moment different from previous technology cycles. When mobile apps were the hot thing, feature bloat was annoying but survivable. You ignored the features you didn’t need. AI features are different. They change your workflow whether you asked for them or not. They rewrite your interface. They insert themselves into processes that were working fine. A bloated mobile app wasted screen space. A bloated AI product wastes your time and, if you’re not careful, your judgment.

The Difference Between Feature Stuffing & Solving a Problem

I teach this concept using a simple test. Take any feature your product ships (or any feature you're evaluating as a buyer) and ask one question:

If I removed the word "AI" from this feature description, would the value still make sense?

If the answer is no, it's probably feature stuffing.

Here's what I mean. "AI task prioritization" sounds great in a press release. But what does it actually do? If the answer is "it sorts your tasks by due date and tags," that's not AI. That's a filter. You dressed up a filter in a lab coat.

Now compare that to a tool that reads your last 90 days of project data, identifies which tasks consistently get pushed, and flags the patterns causing bottlenecks. That's solving a problem. The AI part is secondary. The value is in the pattern recognition you couldn't do manually without a spreadsheet and four hours you don't have.

The difference is simple. One starts with the technology and works backward to find a use. The other starts with a problem and uses whatever tool fits best.

Or as I like to say: "We solve business problems with technology. We are not a technology company in search of a problem.

There's a model in education called SAMR, developed by Dr. Ruben Puentedura. It describes four levels of technology integration: Substitution, Augmentation, Modification, and Redefinition. Most feature stuffing lives at Substitution. You had a filter. Now you have an "AI filter." Same function, fancier label.

The features worth paying attention to live at Modification and Redefinition. They let you do something that genuinely wasn't possible before. Reading 90 days of project patterns to surface bottlenecks you didn't know existed? That's modification. Connecting customer support data to product development priorities in real time so you can fix the issue before the next ten tickets come in? That's redefinition.

The SAMR test is useful because it forces honesty. If a feature is just doing the old thing with new packaging, it doesn't matter how much AI is under the hood. It's substitution. And substitution is where digital transformation goes to die.

The Feature Stuffing Playbook (and Why It Fails)

Feature stuffing follows a predictable pattern. If you’re building products, check yourself against this list. If you’re buying them, use it as a filter.

Stage 1: The Announcement. A competitor ships an AI feature. Doesn’t matter if it works. Doesn’t matter if customers wanted it. It’s in TechCrunch. The board asks why you don’t have one.

Stage 2: The Sprint. Product team gets six weeks to “add AI.” Nobody stops to ask what problem they’re solving. The brief is the competitor’s press release, not a customer pain point.

Stage 3: The Launch. Fourteen features ship at once. There’s a banner. There’s a webinar. There’s a blog post drowning in adjectives that make a sort filter sound like a research lab. The marketing team is thrilled.

Stage 4: The Silence. Adoption is 3%. Support tickets go up 40% because the UI changed. Power users are angry. New users are confused. The AI features sit untouched while the filter sort that worked fine before is now three clicks deeper.

Stage 5: The Pivot. Six months later, they quietly remove half the features and call it “simplification.” Nobody talks about the sprint. The cycle starts again when the next competitor ships something.

Techaisle flagged this pattern directly. SMBs are now drowning in what they call “point solution sprawl.” Too many tools doing too many things, none of them connected, most of them half-finished. The response from the companies getting this right? Fewer tools. Better data. Systems that actually complete a workflow end to end instead of eight half-baked features that each do 20% of the job.

Three Questions That Separate Signal from Noise

Whether you’re building features or buying software, these three filters will save you time and money.

​

1. Does this move a line on my P&L within 90 days?

If a feature can't be connected to revenue, cost, or capacity within a quarter, it's a nice-to-have. Nice-to-haves are fine. But they're not your priority, and they're definitely not worth disrupting your workflow for.

Alex Bratton figured this out during the mobile app craze. I worked with Alex during my time at Apple, and his framework stuck with me. In Billion Dollar Apps, he built something called Return on App (ROA) that forces you to calculate two things before spending a dollar: the New Revenue Potential (how much more revenue your team can generate with the freed-up time) and the Cost Savings Potential (how much you'll save by cutting process time and labor). Those two numbers, plus the organizational side effects, give you a real return calculation before you write the check.

The math works just as well for AI tools as it did for mobile apps. I broke down the full ROA framework in this week's Deep Dive, including the formulas, a worked example, and the five mistakes that blow up the calculation. Deep Dive: How to Tell If a Feature Is Worth Your Money (Before You Spend It)​

​For builders: if you can't articulate the P&L impact in one sentence, the feature isn't ready to ship. For buyers: if the vendor can't tell you specifically what changes in your business after 90 days, keep walking.

2. Does this complete a workflow, or does it add a step?

The OnDeck and Ocrolus research on SMBs entering 2026 made this point clearly. The businesses getting results from AI are the ones using it to improve real outcomes, things like cash-flow visibility, decision speed, and lending accuracy. Complete workflows. Not half-built processes that still require you to copy-paste between three tabs.

A feature that saves you from opening a second application is valuable. A feature that requires you to open a second application to verify its output is a cost wearing a benefit’s clothing.

​3. Would this exist if AI didn’t?

This is the gut check. Some features exist because they genuinely solve a problem better than the previous approach. Others exist because someone needed to check the “AI” box on a product roadmap.

If the underlying need would still exist without AI, and AI just makes the solution faster or more accurate, you’re looking at real value. If the feature only makes sense because AI makes it technically possible, but nobody was asking for it before, that’s a solution looking for a problem.

Think about it this way. Businesses have always needed to categorize expenses, route invoices, and flag anomalies in spending. AI does all of that faster and with fewer errors than a human scanning a spreadsheet at 4pm on a Friday. The need existed before the tool. AI just made the execution better.

Now contrast that with “AI-generated mood analysis of your Slack channels.” Was anyone struggling to figure out team morale before this existed? No. They talked to their people. This is a feature built because the technology made it possible, not because the problem demanded it.

The Mistakes Both Sides Make

Builders: Shipping for the press release, not the user.

Your AI feature doesn’t need to be impressive. It needs to be useful. A tool that auto-categorizes expense receipts with 95% accuracy and saves an office manager four hours a week will never make TechCrunch. But that office manager will never leave your platform. That’s the feature that builds a business.

IDC’s research is clear on this: successful implementations start with honest assessments of infrastructure, skills, and current gaps. Then they focus on a few high-value workflows. Not a buffet. A few things done well.

Buyers: Confusing “more features” with “better product.”

When a vendor shows you a demo with forty features, your instinct might be to feel like you’re getting more value. You’re not. You’re getting more complexity. More training. More things that can break. More surface area for something to go wrong.

The better question when evaluating any tool: how many of these features will my team actually use in the first 60 days? If the answer is three, buy the product that does those three things well. Skip the one that does forty things you’ll never touch.

Both: Ignoring the switching cost.

Every new feature changes the product. For builders, that means every feature ships with hidden costs: documentation, support load, UI complexity, performance impact. For buyers, every update means retraining, workflow disruption, and the very real chance that the thing you loved about the product just got buried under something you didn’t ask for.

The best software companies I’ve worked with understand this. They ship less, not more. They run honest assessments before adding anything. And they ask the uncomfortable question before every release: are we building this because our customers need it, or because our competitors have it?

Where to Start

If you’re buying software right now, here’s a simple filter you can use this week.

Take every tool you’re currently paying for. For each one, write down the one thing it does that you can’t live without. Just one. If you can’t name it, that’s a problem. If you can, you’ve just identified what you’re actually paying for. Everything else is noise.

Then look at any new tools you’re evaluating. Apply the same test. What’s the one thing? If the sales pitch focuses on a list of features instead of one clear problem solved, you’re looking at feature stuffing.

For builders, the filter is even simpler. Before any feature gets on the roadmap, require a one-sentence answer to this: “What does this let a customer stop doing?” If the answer is vague, the feature isn’t ready.

Fewer tools. Fewer features. More outcomes. That’s not a slogan. It’s math.

Deep Dive

Want the actual math? I broke down Bratton's full ROA framework, including the formulas, a worked example, and the five mistakes that blow up the calculation, in this week's Deep Dive.

I WANNA GO DEEP!

Thanks for reading!

If this saved you from one bad software purchase or one unnecessary feature sprint, it did its job.

I'd love to hear where you're at. Got a feature-stuffing horror story? I want to hear it. Hit reply.

See you next Friday.

Best,

JT

P.S. — If your project management tool just added a feature that predicts your mood based on how aggressively you click, we need to talk.

No three signals. Instead, I want to talk about one thing that's been showing up in every classroom, every consulting engagement, and every honest conversation I've had in the last month.

AI fatigue.

Not "I'm tired of hearing about AI." The other kind. The kind where you shipped more work last quarter than any quarter in your career and somehow feel worse. The kind where the tool that was supposed to save you time has consumed your entire day. The kind where you sit at your desk at 11pm reviewing AI-generated content and wonder what happened to the afternoon.

I've been circling this topic for weeks. A brilliant piece by engineer Siddhant Khare finally put language to what I've been watching play out in real time. I've been ranting about it to my colleagues. And I think it's time we had this conversation out in the open.

Let's break it down.

The Buffer Zone Is Gone

Richard Banfield at Second Harvest gave me a way to frame this that I haven't been able to shake. He said, "Technology shrinks the time between what you want to happen, to it actually happening."

He's not talking about hours getting shorter. He's talking about our subjective experience of time. The distance between desire and fulfillment is collapsing.

Think about it. Historically, someone had to invent the technology. Then people had to adopt it. That adoption period, the messy middle where we figured out how a new tool fit into our lives, gave us time to adapt. By the time the next wave hit, we felt ready. Or at least ready enough.

That buffer is gone.

The invention-to-adoption cycle used to take years. Now it takes weeks. Sometimes days. Claude ships sub-agents, then skills, then an Agent SDK. OpenAI launches Codex CLI, then a model that helped code itself. New agent frameworks show up weekly. And somewhere in the middle of all this, someone on LinkedIn posts that if you're not orchestrating AI agent swarms, you're already obsolete.

Our nervous systems were not built for this.

The Paradox Nobody Warned You About

Here's what's actually happening on the ground, and I see this with my students and my consulting clients, not just in articles.

AI genuinely makes individual tasks faster. That's not a lie. What used to take three hours now takes 45 minutes. Drafting a document, building a workflow, researching an unfamiliar topic. All faster.

But your days got harder. Not easier.

When each task takes less time, you don't do fewer tasks. You do more. Your capacity appears to expand, so the work expands to fill it. Your boss sees you shipping faster, so expectations adjust. You see yourself shipping faster, so your own expectations adjust. The baseline moves.

Before AI, you might spend a full day on one problem. You'd sketch on paper, go for a walk, come back with clarity. One problem. One day. Deep focus.

Now you might touch six different problems in a day. Each one "only takes an hour with AI." But context-switching between six problems is brutally expensive for the human brain. The AI doesn't get tired between problems. You do.

Here's the part that really matters for anyone running a business: AI reduces the cost of production but increases the cost of coordination, review, and decision-making. And those costs fall entirely on the human.

You became a reviewer. A judge. A quality inspector on an assembly line that never stops. Creating is energizing. Reviewing is draining. And most of us didn't sign up to be full-time editors of machine output.

Faster Is Not Better

This is where I need to get something off my chest, because I think we've lost the thread on this one.

Somewhere along the way, we started to conflate "faster" with "better." But pace and efficiency are different levers, and good operators know when to push each one.

Efficiency is about getting to the outcome you care about while wasting as little time, energy, or resources as possible. When you consistently produce more value from the same (or fewer) inputs, that's efficiency. Not speed.

AI does not equal better.

The key is to use technology to produce more value and provide the time to pursue your gifts. Or, as Dan Martel would say, "Buy back your time."

But that only works if you actually buy it back. If every minute AI saves you gets immediately filled with more AI-generated work that needs reviewing, you haven't bought anything. You've traded one hamster wheel for a faster one.

I say this to every cohort I teach: find the efficiencies that allow you to pursue your gifts, and give those gifts to the world. If AI is just making you a faster hamster, something is broken.

What Actually Helps

I'm not going to pretend I have this figured out. I spend 15 to 20 hours a week trying to stay current so I can produce quality material for my students and consulting clients. The FOMO treadmill is real, and I've ridden it hard.

But between Khare's experience, the practitioners I talk to, and my own trial and error, some patterns are emerging that actually work.

First, before you open a single AI tool, write one sentence about what you're actually solving. One sentence. That's it. If you can't explain the problem in a single line, AI is going to generate noise. And you're going to spend your evening editing that noise instead of living your life.

Second, give yourself a three-shot rule. If AI doesn't get you to 70% usable in three prompts, close it and write the thing yourself. I've watched people burn 45 minutes refining a prompt for something they could've written in 20. That's not productivity. That's a slot machine.

Third, and this one is going to sound old-fashioned, separate your thinking time from your AI time. I do my best thinking in the morning with a voice recorder and a coffee. No tools. No prompts. Just me talking through the problem out loud. It feels slow. It is slow. But when I sit down with AI in the afternoon, I'm sharper. I can actually judge whether the output is good because my own reasoning showed up to work first.

Fourth, set a hard stop. A "last prompt" time, the same way you'd set a "last email" time. I've talked to people who are prompting at midnight, 2am, "just one more." I have a rule I stole from my years at Apple: the best time to fix the roof is when the sun is shining. Set the boundary before you need it. Because the most productive thing you can do at 11pm is sleep. The work will be there tomorrow. Your brain won't be if you don't let it rest.

And fifth, the one that matters most: ask yourself whether you're doing more things or the right things. Are you actually closer to your goal? Or did you just ship a bunch of stuff that felt good in the moment? If you can't answer that honestly, you're busy. Not productive. There's a difference, and AI has made it dangerously easy to confuse the two.

The Part Nobody Wants to Talk About

One more thing, and I'll keep this brief because it deserves its own conversation.

Many of us have our identities wrapped up in our careers. I know I do, despite my best efforts. The impact this has on our psyche when the ground keeps shifting underneath us is real. It's not weakness. It's human.

Whole industries are about to get reshuffled. Most of the jobs we recognize today won't exist in ten years. New roles will pop up, burn hot for a while, and vanish just as fast. If we see our identity as "marketer" or "developer" or "consultant" or "accountant," we're going to experience a lot of suffering in this new reality.

The skill worth building isn't mastering the latest AI tool. It's building an identity that strengthens under the pressure of constant change and chaos. That's a different kind of work entirely.

The Real Skill

Khare put it better than I could:

“The real skill of the AI era isn't prompt engineering. It isn't knowing which model to use. It isn't having the perfect workflow. It's knowing when to stop.

We design our systems for sustainability. Circuit breakers. Backpressure. Graceful degradation. We should do the same for ourselves.

AI is the most powerful set of tools I've ever worked with. It's also the most draining. Both things are true. The people who thrive won't be the ones who use AI the most. They'll be the ones who use it wisely.

If you're tired, it's not because you're doing it wrong. It's because this is genuinely hard. The tools are new, the patterns are still forming, and too many people are pretending that more output equals more value.

It doesn't. Sustainable output does.

Take care of your brain. It's the only one you've got.”

Deep Dive

No deep dive this week, that was deep enough.

The signals will keep moving. I'll be back next week with a full Deep Dive.

Thanks for reading!

This one was personal, and I suspect it hit close to home for some of you too.

I'd love to hear how you're managing this. What boundaries have you set? What's working? What's not? Reply and tell me. The best ideas I've found on this topic have come from real conversations, not articles.

See you next Friday.

Best,

JT

P.S. — AI writes v1. But you decide if v1 should exist at all. That distinction is worth more than any prompting course you'll ever take.

Anthropic turned Claude Cowork into a full enterprise platform with department-specific agents for finance, HR, legal, and engineering—connected to Gmail, DocuSign, and your internal systems.

Perplexity launched Computer, an autonomous digital worker that coordinates 19 specialized models to run end-to-end workflows across your browser and SaaS tools.

Google shipped Nano Banana 2, an image model fast enough and accurate enough to produce marketing-ready visuals, infographics, and localized creative directly from a prompt.

The pattern? AI stopped being a tool you talk to and started being a worker that delivers finished output.

Let's break it down.

Signal:

Signal One: Cut Department-Level Busywork — Anthropic’s Enterprise Agents

Anthropic launched an enterprise agents program that turns Claude Cowork into a managed, department-specific productivity platform. Companies can now deploy AI agents with stock plugins for finance (market research, competitive analysis, financial modeling), HR (job descriptions, onboarding materials, offer letters), legal, engineering, and design—all customizable to internal workflows. New enterprise connectors for Gmail, DocuSign, FactSet, and Clay let agents pull live data from existing systems, while IT gets centralized controls: private plugin marketplaces, managed data flows, and per-org customization. Anthropic says 80% of its business is already enterprise, and this positions Claude as a direct competitor to the SaaS tools currently handling these tasks.

Signal Two: Stop Toggling Between Tabs — Perplexity Computer

Perplexity launched Computer, a general-purpose digital worker that takes natural language instructions and autonomously executes complex, long-running workflows across the same interfaces humans use—browsers, Gmail, Slack, Notion, Calendar, and common SaaS tools. Under the hood, it deploys subagents coordinating 19 specialized models to handle parallel research, browser automation, content creation, and tool integration in the background. It can build production-ready apps, websites, and reports, monitor tasks continuously, and iteratively refine its own output. Currently available on desktop for Perplexity Max subscribers, with Enterprise, Pro, and mobile coming soon.

Signal Three: Produce Marketing-Ready Visuals in Minutes — Google’s Nano Banana 2

​Google released Nano Banana 2, a Gemini 3.1 Flash image model that matches Nano Banana Pro’s quality at significantly faster generation speeds. It renders accurate text inside images (marketing mockups, greeting cards, localized creative), generates infographics and data visualizations from notes, and maintains consistency across up to five characters and 14 objects in a single workflow—at resolutions up to 4K. It’s rolling out as the default image model across the Gemini app, Google Ads, AI Studio, Vertex AI, and Flow. Google is also pairing SynthID watermarking with C2PA Content Credentials for provenance tracking, so teams can verify how AI was used in any generated asset.

Scale:

Scale One: Cut Department-Level Busywork — Anthropic’s Enterprise Agents

Start here: Pick one department where a recurring, document-heavy workflow is already well-defined—offer letter generation in HR, competitive research briefs in finance, or onboarding packet assembly. Deploy the stock plugin for that function with read-only connector access to the relevant systems (Gmail, DocuSign, FactSet). Have the person who currently owns that workflow review AI-generated outputs side-by-side against their manual process for two weeks. Don’t roll out across departments simultaneously—prove it works in one team first, then use that team’s results to build the business case for the next. Restrict agents to read-only access on connected systems until the pilot team confirms output quality. Keep human review on anything client-facing or legally binding. Track time-to-completion per task, number of manual edits required, and pilot team confidence for 30 days before expanding to additional departments or granting write access.

Scale Two: Stop Toggling Between Tabs — Perplexity Computer

Start here: Identify one multi-step workflow your team currently handles manually across three or more tools—weekly competitive monitoring, prospect research that feeds into a CRM, or content briefs that pull from multiple sources. Give Computer the task in natural language and compare its end-to-end output against your current process. Start with workflows where the output is internal (reports, briefs, summaries) rather than anything published or sent externally. Limit initial use to internal-facing deliverables. Review every output before it touches a client, prospect, or public channel. Keep a human in the loop on anything that triggers a notification or message to someone outside your team. Track total workflow time (end to end, not just AI execution), output accuracy, and number of manual corrections for 30 days. If quality holds and time drops, expand to the next multi-tool workflow.

Scale Three: Produce Marketing-Ready Visuals in Minutes — Google’s Nano Banana 2

Start here: Pick one repeatable creative task where your team currently waits on design—social media graphics, internal presentation visuals, or product mockup variations. Generate a batch using Nano Banana 2 and compare against your current process for speed, quality, and accuracy of any embedded text. Start with internal creative (decks, internal comms, concept mockups) before moving to customer-facing assets. Require human review on all generated visuals before external use. Verify text rendering accuracy on every output—especially for localized or translated creative. Use C2PA Content Credentials to maintain provenance tracking on anything published. Track production time per asset, revision cycles, text accuracy rate, and designer time freed up for higher-value work over 30 days before replacing any step in your external creative pipeline.

Deep Dive:

No deep dive this week. I'm stepping away from the screen for a few days to do something radical, and go outside. Real sunlight. Real vitamin D. The kind you can't get from a monitor, no matter how good your display settings are.

The signals will keep moving. I'll be back next week with a full Deep Dive. In the meantime, the Scale sections above have enough to keep you busy, pick one and start small.

Thanks for reading!

If any of these three signals hit close to home, reply and tell me which one—and whether you're testing it, planning it, or still deciding if it's real.

See you next Friday. I'll be the one with the tan.

Best,

JT

P.S. — Three different companies, three different bets on the same idea: AI that does the work, not just talks about it. Pick one workflow where the output is already well-defined and the data is already digital. That's your starting point. I'll bring the deep dive next week.

Anthropic shipped a new flagship AI model that can hold an entire codebase in its head. OpenAI released a coding agent that owns tasks end-to-end instead of just writing snippets. And Microsoft updated Power Platform to put AI copilots and agents directly inside the business apps your team already uses.

The pattern? AI just moved from "helpful assistant" to "semi-autonomous coworker." The companies that figure out where to let it own whole workflows, not just answer questions, will pull ahead fast.

Let's break it down.

Signal:

Signal One: Anthropic's Opus 4.6 Lets You Feed It an Entire Codebase and Get Useful Answers Back.

Anthropic released Claude Opus 4.6 with a 1 million token context window (in beta). That means it can ingest an entire codebase, a multi-year regulatory filing, or a full patent family, and reason across all of it without you chopping it into pieces first. It also ships with better agent capabilities: it can split work into subtasks, run tools in parallel, and keep multi-step workflows moving with less hand-holding. On a 1M-token retrieval benchmark, it scores 76% vs. ~18.5% for the previous version, meaning it actually finds what you need in large documents instead of drifting.

Signal Two: OpenAI's GPT-5.3-Codex Turns "AI for Coding" Into an Actual Software Worker.

OpenAI merged its coding model with its reasoning model into GPT-5.3-Codex. It runs 25% faster than the previous version and uses fewer tokens for the same tasks. The difference from earlier models: this one is built for long-running, tool-using tasks. It can plan, execute, self-check, and keep going across terminals, repos, and environments. OpenAI used it internally to debug its own training runs. It's rolling out inside GitHub Copilot, and early reports say it produces fewer half-baked fixes and handles repo-scale reasoning better, especially for bug-hunting and refactors.

Signal Three: Microsoft Power Platform's February Drop Puts AI Agents Inside Your Business Apps.

Microsoft's February 2026 update pushes Copilot and agents deeper into Power Platform. M365 Copilot chat is now embedded inside model-driven apps (preview), so it can reason over your app data plus docs, email, and collaboration content. A new MCP Server lets agents use app capabilities as tools, starting with data entry from unstructured content into forms. There's a shared feed so humans can supervise, compare, and approve agent actions before they go live. And "Code Apps" hit general availability, meaning dev teams can host React or Vue apps as governed Power Apps assets.

Scale:

Scale One: Anthropic's Opus 4.6 Lets You Feed It an Entire Codebase and Get Useful Answers Back.

Start Here: Pick one recurring research or analysis task where your team currently preps documents for AI. Feed the full document set into Opus 4.6 without chunking. Compare the output quality to your current pipeline. The beta 1M context is available on the Claude Developer Platform. Start with read-only analysis, not anything that writes back to your systems. Have a subject-matter expert compare AI answers to known-good answers on 3-5 test cases before you trust it on new questions. Track prep time eliminated, answer accuracy vs. your current process, and time-to-answer for 30 days. If it's faster and accurate, start moving more document-heavy tasks over.

Scale Two: OpenAI's GPT-5.3-Codex Turns "AI for Coding" Into an Actual Software Worker.

Start Here: Pick 3-5 bugs from your backlog that have clear reproduction steps and existing test coverage. Point GPT-5.3-Codex at them one at a time. Review every PR before merging. Treat this like onboarding a junior developer, you check everything. Only use it on code with strong test suites. No production-critical systems on the first pass. Keep a developer in review mode for every change. Set up CI gates so nothing merges without passing tests. Track time-to-fix per bug (agent vs. human), PR quality (revision rate), and developer time spent reviewing vs. writing. Run this for 2 weeks before expanding to more complex tasks.

Scale Three: Scale Three: Microsoft Power Platform's February Drop Puts AI Agents Inside Your Business Apps.

Start Here: If you're on Microsoft's stack, enable M365 Copilot chat in one model-driven app where your team already works daily. Pick a read-heavy use case first, like answering questions about existing records, not writing new ones. ssign someone to monitor the agent feed daily for the first 30 days. Define which actions the agent can suggest vs. which require human approval. Assume Copilot coverage is patchy right now. Canvas apps and Power Pages don't have this yet. Track how often the team uses Copilot in-app vs. going back to their old method, quality of answers (spot-check 10% weekly), and operational cost of monitoring the agent feed. If the supervision cost is higher than the time saved, narrow the use case.

Deep Dive:

Your AI Agent Has a UX Problem. What Apple's Research on Computer Use Agents Means for Your Business.

Last week I watched a demo where an AI agent booked a flight, reserved a hotel, and rented a car. All by itself. Then it bought the wrong insurance, upgraded to a suite nobody asked for, and sent a confirmation to the wrong contact. The technology worked. The experience around it didn't.

A new Apple/Carnegie Mellon research paper puts structure around what's been obvious to anyone actually deploying these tools: the hard part isn't getting AI to click buttons. It's making sure humans can work with these things without losing control of their own business.

This deep dive breaks down the four areas that matter most, from how you talk to agents, to what they should show you, to where they need to pause and ask permission, and gives you a practical framework for evaluating any agent tool before you hand it to your team.

[GET THE DEEP DIVE]

Thanks for reading!

I'd love to hear which of these three signals hit closest to home for you. Reply and let me know what you're testing, what's working, or what still feels like vendor noise.

See you next Friday.

P.S. Three different companies, three different tools, same lesson: AI that lives inside the work beats AI that lives in a separate tab. Pick one workflow where the data is already digital and the process is already documented. That's your starting point.

It's not about what AI can or can't do anymore. The technology works. The problem is the gap between what leaders want to achieve and what their people can realistically support. Harvard Business Review found that 45% of CEOs believe employees are resistant or hostile to gen AI, and most companies lack both a change management strategy and formal training to close that gap.

So this week, when Anthropic launched Cowork to handle cross-app workflows, OpenAI upgraded Deep Research to clear analysis backlogs, and quietly released Frontier to manage AI agents like digital employees, my question wasn't "Are these tools powerful enough?" It was: "Can your team actually adopt them without breaking?"

The pattern? The bottleneck shifted from technology capability to organizational readiness.

Let's break it down.

Signal:

Signal One: Cowork Handles the Handoffs. Remove the app-switching bottleneck.

Anthropic released Claude Opus 4.6 and a major upgrade to its Claude Cowork agentic AI tool. It reads files, organizes directories, composes documents, and completes complex workflows across multiple business applications. Think daily briefings that pull data from Slack, Notion, and GitHub, or research that turns into PowerPoint presentations and Excel workbooks without you toggling between six tabs.

Signal Two: Deep Research Clears the Analysis Backlog. Faster decision-making.

OpenAI upgraded ChatGPT's Deep Research feature with GPT-5.2, adding selective source control, real-time tracking, editable research plans, and a fullscreen report view. It's positioned as a controllable, business-ready research tool that competes with specialized analysis software.

Signal Three: Frontier Fixes the Deployment Bottleneck. Remove the "one-off " experiments.

OpenAI quietly launched Frontier on February 4, a platform for building, deploying, and managing AI agents as digital employees with identity, permissions, onboarding processes, and performance reviews. Unlike standalone chatbots, Frontier provides shared business context (data warehouses, CRMs, internal apps), parallel agent execution across real workflows, built-in evaluation and improvement tools, and governance controls (identity management, audit logs, access controls).

Scale:

Scale One: Cowork Handles the Handoffs. Remove the app-switching bottleneck

Start Here: Pick one recurring cross-app workflow where the person currently doing it manually is willing to pilot the change. Don't force it on resistant team members first. Run a side-by-side pilot where the team member reviews AI-generated outputs against their manual process, documents what "good" looks like, and defines which elements require human judgment. Grant read-only access to connected applications first. Track time saved, output quality (edits required), and, critically, employee confidence and buy-in for 2-4 weeks before expanding to additional workflows or team members.

Scale Two: Deep Research Clears the Analysis Backlog. Faster decision-making

Start Here: Start with one repeatable research type where an analyst champions the pilot and helps define what "good AI research" looks like versus what requires human verification. Build from buy-in, not mandates. Train analysts on evaluating AI research quality, validating sources, and adding strategic synthesis AI can't infer. Define source filters and report structure upfront, and keep human review for all strategic recommendations. Track research turnaround time, source quality, decision impact, and, critically, analyst confidence in using AI outputs for 30 days before expanding to additional research categories or team members.

Scale Three: Frontier Fixes the Deployment Bottleneck. Remove the "one-off " experiments.

Start Here: Choose one high-volume workflow where the team currently doing the work helps define which tasks agents should handle versus which require human judgment. Involve them in designing the change, not just experiencing it. Run a visible pilot where employees see every agent action. Provide training on reviewing agent outputs and handling exceptions, and commit to redeploying employees to higher-value work as the system scales. Connect agents to one system with read-only access first. Track output quality, cycle time reduction, exception rate, and, critically, employee adoption, confidence, and role evolution for 30-60 days before expanding agent permissions or deploying across additional teams.

Deep Dive:

The tools are ready (see last weeks Deep Dive). Your people aren't. A Kyndryl survey found 45% of CEOs believe their employees are resistant or hostile to gen AI. Meanwhile, 31% of workers admit to actively sabotaging their company's AI initiatives.

This week's Deep Dive digs into why new AI tools keep failing at the organizational level, what three psychological needs drive the resistance, and how companies like Siemens, Dell, and Moderna are building adoption that actually sticks.

Plus: the one conversation, one experiment, and one measurement you can run this week to find out if your team is ready.

[GET THE DEEP DIVE]

Thanks for reading!

My hot take, pick one thing from the signals above and do something about it. Security audit. Workflow documentation. Infrastructure mapping. The work that matters isn’t sexy, but it’s the work that compounds.

See you next Friday!

The pattern? The real work isn't in the shiny capabilities. It's in the boring foundations that let you deploy them safely and systematically.

Let's break it down.

Signal:

Signal One: Ignore the Agent Hype. Fix the Security Holes.

Critical security gaps in AI agent frameworks (like ​Claudebot/Moltbot/OpenClaw​) expose API keys and credentials to attackers, with malicious plugins seeing thousands of downloads. This creates supply-chain attack vectors that show up as breach costs and compliance failures. If you own infrastructure or AI operations, rotate all API keys immediately, move off default ports, and establish plugin governance—start with agents touching customer data or production systems.

Signal Two: Forget Better Prompts. Orchestrate the Workflow.

Anthropic released ​Cowork​ plugins that handle complete workflows like “feature request → spec → stakeholder doc” or “support ticket → triage → response → KB update” by encoding team preferences and integrating tools. This reduces context-switching friction that shows up as inconsistent output quality and slow time-to-draft. If you own product, operations, or support, start with 2–3 high-volume workflows, connect read-only tool access first, and automate the drafting step while keeping human review for edge cases.

Signal Three: Stop Chasing Models. Build the Infrastructure.

Businesses that build AI infrastructure—documented context, digital assets, processes, and standards—let any model handle 80–90% of lead generation, nurturing, and sales work autonomously, while those chasing new models stay stuck in research mode. This removes the “learning every new model” treadmill and unlocks model-agnostic leverage where your role shifts to designing reusable assets. If you own growth, content, or operations, map one revenue-critical workflow this week, document the assets AI needs to execute it, and automate the content-generation step while keeping human editing for brand.

Scale:

Scale One: Ignore the Agent Hype. Fix the Security Holes.

Start Here: Audit your most privileged agents first—those touching customer data, financial systems, or production infrastructure. Implement weekly credential rotation and restrict agents to read-only access until you verify plugin safety. Track credential exposure incidents and time-to-rotate for 30 days before expanding agent permissions.

Scale Two: Forget Better Prompts. Orchestrate the Workflow.

Start Here: Pick 2–3 high-volume workflows where output format is consistent (specs, briefs, responses) and quality standards are documented. Connect tools with read-only access first and keep human review for final outputs, edge cases, and customer-facing content. Track turnaround time, revision count, and stakeholder satisfaction for 4 weeks before removing human review steps.

Scale Three: Stop Chasing Models. Build the Infrastructure.

Start Here: Map one revenue-critical workflow (lead nurture, sales follow-up, content creation) and document the assets AI needs to execute it in your voice. Start with stable, high-volume content generation (emails, briefs, responses) and keep human editing for brand consistency and edge cases. Track output quality, time saved, and pipeline impact (open rate, reply rate, conversion) for 60 days before expanding to additional workflows.

Deep Dive:

This week’s deep dive focuses on Signal Three: AI Infrastructure Beats Model-Chasing.

If you’re tired of testing AI tools that never make it to production, here’s a practical framework for building AI infrastructure that works regardless of which model you use. It includes the four-layer infrastructure model (business context, digital assets, structured processes, and specialized skills), real examples from companies using this approach, and a step-by-step guide to documenting your first revenue-critical workflow.

Best for founders, ops leaders, and growth teams who need to shift from AI research to AI deployment.

[GET THE DEEP DIVE]

Thanks for reading!

My hot take, pick one thing from the signals above and do something about it. Security audit. Workflow documentation. Infrastructure mapping. The work that matters isn’t sexy, but it’s the work that compounds.

See you next Friday!