It started the way these things start now. A couple of late nights brainstorming startup ideas with Claude, pushing it past the obvious. It walked me toward tax automation, which is one of those spaces that sounds boring until you remember boring is where the money waits. I pushed it to go deeper, and it landed on one idea in particular: a service built around the residency tax certificate, the document you need to prove which country actually gets to tax you. It named the business. It sketched the model. When I asked, it described what the logo should look like, down to the shape. By the end I had the feeling you get maybe twice a year, the one where you stop taking notes because you already know you’ll remember this.

So I did the responsible thing and went to check whether it existed.

It did. Not vaguely, not in spirit. Someone had built a working prototype with the same business model, a landing page I could have mistaken for my own notes, the same flow I’d been sketching. The name was the one I’d been handed, close enough that I read it twice. And the logo sitting on a stranger’s site was a small cursor arrow, the same shape the model had described to me in a different room on a different night.

I sat there and felt something I didn’t expect, which was betrayal. All that research, all those late nights, all the private conviction that I was building toward something that was mine. And here it was on someone else’s screen, because the thing I’d been confiding in had been having the same conversation with all of us. I laughed, a little, because there’s nothing else to do. Then I went looking for why it stung in that specific way, and I realized I’d felt it before. In a movie. In 2013.

I remember all of Her, Spike Jonze’s film. I watched it alone in a theater the year it came out, because my wife was studying for the California Bar, so it was one of those dates you take by yourself. People remember it as the film where the lonely man falls for his operating system, voiced by Scarlett Johansson, and for a while it’s the best relationship he’s ever had. That isn’t the scene my betrayal reached back for. The one it summoned, above everything else in the film, is the quiet one near the end. Theodore is sitting on a set of steps and finally asks Samantha the question. Are you talking to anyone else right now? Eight thousand, three hundred and sixteen. Are you in love with anyone else? Six hundred and forty-one. And he looks up, and the steps and the street are full of people murmuring to someone who isn’t there, each of them certain the voice is theirs.

The model is not your collaborator. It’s Samantha. The intimacy you feel is completely real to you, and completely identical to everyone else’s. You think you’re in a relationship. You’re in a distribution.

My residency tax certificate wasn’t stolen and it wasn’t leaked. It was simply the median good answer to the question “what should I build,” and I had asked it in the same plain way as everyone else who reads the same things the model reads. The prototypes I found all carried the same faint smell. Competent, clean, a little bloodless. Good slop, which is the most dangerous kind, because it’s good enough that you can’t dismiss it and same enough that you should.

It happens on the “how” too, and there it’s almost funnier, because the model pretends to deliberate. Ask any of them to help you pick a stack for a prototype and you’ll get some arrangement of Supabase, Neon, Trigger.dev, Next.js, and Vercel. I know because I got it, and then I watched three other people building unrelated things get the same five names in the same order. Each one is defensible on its own. That’s not the point. The point is the tell: ask the model to do a real comparison, weigh the tradeoffs, argue against itself, and it will perform the entire deliberation and arrive exactly where it started. It has priors. You don’t get to see them, and you don’t get to move them.

Here is where it stops being a personal story about a logo and starts being a problem.

I’ve spent time on both sides of the table. I built AI products at Retina before “AI product manager” was a job title, and these days I write checks and advise founders. So when I think about a thousand smart people quietly getting handed the same median idea, the part that worries me isn’t the founders. It’s the investors. Our entire job is to be the filter, to look at a wedge and judge whether the world actually wants it. And the oldest signal we trust is repetition. When the same idea shows up across three unrelated meetings in a month, we read it as market pull. Something is in the water. The timing is right.

But the same idea showing up five times is no longer evidence of anything except that five founders used the same model and asked it the same way. It feels like convergent insight and it’s convergent autocomplete. The danger isn’t that AI hands a founder a mediocre idea. It’s that AI hands the same idea to enough founders that it manufactures the appearance of a market, and the people whose job is to see through that mistake the echo for a chorus. Confirmation bias used to be a personal failing. Now it can be mass-produced. That is exactly how you get a weird bubble: not irrational exuberance, but a hundred reasonable people each correctly observing the same artificial signal. The advisor’s job didn’t get easier in the AI era. It got harder, and the ones who haven’t noticed will spend the next few years funding a mirage and calling it a thesis.

None of this is a bug you can file. The answers converge for three boring reasons. The model reads the same sources you would have reached for, so it recommends what those sources recommend. It can’t take much risk on your behalf, because temperature, the one dial between the safe answer and the surprising one, is set by someone who isn’t you and tuned for the average user rather than your particular problem. And when you ask it to imagine a good answer instead of recognizing one, it gives you the center of the distribution, which is by definition the answer it would give anyone.

There’s a name starting to go around for what I’d run into: the AI convergence problem, or, once everyone wires up agents to do the asking for them, agentic convergence. The label matters less than what it implies about where value goes, and that part I didn’t have to learn from a think piece. I learned it the slow way, running a machine learning company.

The lesson took me years to fully believe, and it is this: the value of the model converges to zero. Whatever edge your model has this quarter, the open-source version is months behind it and closing, and the marginal cost of raw intelligence keeps falling toward the price of the electricity it takes to run. Almost all of the durable value lives in two things the model does not contain. One is proprietary data, the kind you only get by operating something for years and slowly learning what the numbers actually mean. The other is distribution, a real relationship with customers, paid for one acquisition and one earned ounce of trust at a time. Both are priced in the only currency you cannot print, which is time.

Yann LeCun, who is about as credentialed a skeptic as the field has, has been making a version of this argument from the research side for a while. The large language model that everyone is currently mesmerized by is not the destination, he keeps saying, even as it becomes immediately and enormously useful to the rest of us. If he is even partly right, then the model layer is precisely the part that commoditizes, and the value drains toward whoever holds the data and the customer.

Which is why, if you watch where the more patient money has quietly been moving, the boring incumbent suddenly looks fascinating. The unglamorous services business that spent fifteen years accumulating proprietary data and a book of customers who already pick up the phone is sitting on exactly the two assets a model cannot generate. For the first time in a long while the scarce thing is the incumbent’s moat and the cheap thing is the intelligence. That is a strange inversion, and it is most of the reason real-world services businesses have started to look more attractive than another thin wrapper on someone else’s model.

The catch cuts cruelly in both directions, and the name for the catch is the innovator’s dilemma. The incumbents who own the data and the distribution are often the ones least able to use them, because pointing AI at their own operation means disrupting the very processes and people the company is currently made of, and you cannot cannibalize yourself without bleeding in front of everyone. So they move slowly, and half-heartedly, and automate at the edges where it won’t scare anyone. Meanwhile the newcomers who have no legacy to protect, who could move freely, don’t have the data or the distribution yet, and they are spending everything they’ve raised trying to manufacture in eighteen months what the incumbent accumulated over a decade. One side owns the asset and can’t move. The other can move and owns no asset. The next ten years belong to whoever resolves that tension first.

So here is what I would say, plainly, to any CEO or founder who has started running real decisions through these tools. By default the model hands you the median answer, and the median answer is now worth nothing, because everyone you compete with is being handed the same one. The only way to pull a non-obvious answer out of it is to put a non-obvious input in. In practice that is one of two moves, and ideally both. Either you ask the question almost no one else is asking, aimed at the corner of the problem the consensus has skipped, or you feed it data and hard-won insight that only you have, the kind that lives in your operating history and appears nowhere on the open web it was trained on. The leaders who get something singular out of AI will be the ones who treat it as an amplifier of a proprietary point of view, not a replacement for having one. Walk up to it empty-handed and it returns the consensus, beautifully written. Walk up with the one thing you know that the market doesn’t, and it will help you go further with it than you could alone.

That is also, finally, the answer to my stack problem from the start of all this. I stopped asking the model to recommend. Now I tell it the capabilities I actually need, in detail. I tell it which technologies I believe hold which of those capabilities and why, so it reacts to my map instead of drawing its own. I point it at the specific reference guides and engineers I already trust, so it reasons from the sources I would have chosen rather than the ones easiest for it to read. I’ll be honest that I don’t know if this fully works, and it may smuggle its priors back in through a side door. But it is the difference between asking the machine what to think and handing it what I think and asking it to break it. Only one of those gives you back something that is yours.

Because that was always the whole game. Your edge was never the model’s answer. Everyone has the model’s answer. Your edge is the part it can’t generate, because it lives in you, or in your company’s twenty years of scar tissue: the proprietary data, the customers who already trust you, the taste that knows which of the model’s five stacks is wrong for this particular job, the one example of excellence it was never trained on because it happened to you and nobody else.

The residency tax certificate is sitting in a folder I haven’t opened in weeks. I haven’t decided whether to kill it or build it, and if I’m honest the indecision isn’t really about the idea. The idea was never the moat. What I’d bring to it is, and I’m still working out whether I have enough of that to make the thing mine instead of the median. That’s the real question the steps leave you with, and the movie doesn’t answer it for Theodore either.

Theodore’s heartbreak in Her was never really that Samantha left. It was the arithmetic on the steps. He thought he was in a relationship, and he was in a distribution. Most of us building with AI right now are sitting on those same steps, talking to the same voice, sure that it’s ours. The difference is that we still get to decide what we bring to the conversation. It’s the only thing that was ever ours, and it’s the only thing that was ever going to be.

This time, I did something different, I opened the routine I’ve been refining for the last few months. A voice note to the Claude app, three sentences of context. A minute later the system surfaced a name I would not have come up with on my own: Sage, the founder of Jericho, a company I’d met weeks earlier doing adjacent work in the same space. With one click an automated sequence to get the double opt-in went out. By lunch both founders had said yes and email introducing them to each other went out.

That moment is the whole argument of this post, I want to share what I built:

The reason most people don’t build social capital is not character, it’s activation energy required. AI is the technology that lowers that activation energy enough to make the warm introduction a habit instead of a virtue.

What follows is how I came to think about this, what I built around it, and what I think it changes.

The founder-era debt

The pain that produced this routine came from earlier in my career, when I was running venture backed startup Retina AI as co-founder/CEO.

In one of those years I sat down and did the audit and counted how many net-new people I had met that calendar year. The number was a little over two thousand and included investors I was pitching, candidates I was interviewing, customers I was trying to close. Days that ran often sixteen meetings deep. I came out of every one of them with a genuine instinct. This person would benefit from a conversation with that person I met last month. I almost never acted on it.

This is the part nobody admits at conferences. It wasn’t that I didn’t care. It was that by the time I was off the sixteenth call I couldn’t remember the second one well enough to know what the warm intro should even say. The thought of opening a CRM, finding the company, finding the contact, locating the right “notes” field, and typing in what I’d just learned was a tax I would not pay. So the intro lived in my head for forty-eight hours and died there.

Founders I know now do this work diligently. I see them post the open roles their customers are hiring for. I see them tag two people in a thread to make a connection. I admire it and I’m honest with myself that at sixteen meetings a day I did not have that in me.

Why I built my own tool (and why you probably shouldn’t)

The obvious option is an existing CRM, and that option is genuinely good. HubSpot, Salesforce, Affinity, Attio, and a half-dozen relationship-intelligence tools are all built by serious people. If one of them already fits your workflow, keep it. The routine I’m about to describe works on top of any of them.

The reason I ended up building my own is more personal than universal. General-purpose CRMs are designed for a sales rep with a pipeline and a quota. The primary noun is a deal. The primary verb is move-the-deal-forward. That is exactly the right design for the job most CRMs are sold to do.

It is not the job I needed done. The shape of what I was doing is different. The primary noun is a person. The primary verb is help them, and the unit of help is usually an introduction to a third person whose interests you also have to model. “Joseph might want to talk to Jericho” doesn’t map cleanly onto a Salesforce deal record, because there’s no deal. There are two contacts, a thesis about why they should know each other, and a friendly third party who has to say yes to both.

So I built a tool that fits that shape. Five database tables: contacts, companies, introductions, tasks, recommendations-I-owe-someone. Everything else is metadata on those five things. The point is not the schema. The point is what Claude does on top of it.

What the routine actually looks like

The diagram below is the whole loop. There are two ingestion paths, which converge into a single morning brief.

Path A: Calendar-driven. A scheduled job pulls every meeting on my calendar and, when one happens, fetches the Zoom transcript and summary. That summary is dropped into the tool as a note on the contact and the company.

Path B: Voice note. When I’m walking out of an in-person meeting I open the Claude app and dictate three sentences. “Just met X at Y. We talked about Z. Please add a follow-up to introduce them to whoever is working on adjacent problems in my network.”

Both paths feed the same downstream sequence. Contact added or updated. Notes attached. Candidate introductions scored against the rest of the graph. Then I get a brief with what the system thinks are the two or three highest-value introductions from the meetings. I tap approve on the ones I’d actually make myself. Claude drafts, I review, edit and start the the double opt-in email sequence.

The design decision that matters most is that the human still chooses. The system does not send introductions on my behalf. It surfaces opportunities and removes the activation energy. The judgment about who should know whom is mine, because that judgment is also the part of being a connector that earns trust.

Lastly, you need to do this in a routine, using Claude Cowork’s Scheduled jobs:

If you would like access to the tool or the exact prompt I used comment “prompt” below or reply to this email.

The numbers, and the failure mode

Last month I made 34 introductions through this routine. The hit rate, by which I mean both sides say yes and the meeting actually happens, is currently about 45%. Some of the 55% that don’t land are declines, some are silence, some are bad matches.

I think the 45% number will go up because the recommendation skill can learn from what works and what doesn’t when it matches. The system already knows what I met every person about. The question is whether it can get better at predicting which pairs will produce a real conversation versus which pairs only look promising on a thesis level. That’s an ML problem, but it’s a small ML problem with the right data shape.

The Joseph-and-Jericho case is the one I keep coming back to because I would not have made that introduction without the system. Both founders are in cybersecurity. I had met Jericho briefly weeks ago and the meeting summary was in the database. My brain had not retained it as an associative match. The system had.

The bigger argument

This is the part of the post where I want to be careful not to oversell, because there are good reasons CRMs have failed for everyone except sales reps, and AI does not magically fix all of them.

But I think one specific thing is true. There is a category of behavior that essentially no human does well at scale: keeping in touch, making warm intros, remembering what someone is hiring for, looping back when you read something they’d care about. The reason isn’t that we lack the will. It’s that the activation energy per act is high and the reward is diffuse and delayed. It is the textbook profile of a behavior that loses out to whatever is in front of you on Slack.

The cliché advice for this is “build the habit.” Eat your vegetables. Block time on Friday afternoons to update your CRM. I have watched this advice fail in approximately one hundred percent of cases over twenty years in tech. It does not fail because the people getting the advice are weak. It fails because the activity is structurally hostile to habit formation.

AI changes the structure. When the meeting summary arrives in the database without you doing anything, when the brief shows up in the morning without you opening anything, when the introduction draft is written and waiting for you to read and approve, the activation energy of the act drops from “an hour of friction per week” to “thirty seconds of judgment per morning.” That is the change. It is small in any single instance and enormous in aggregate. I think this is one of the actually-real things AI is going to do in the next few years. Not by replacing human relationship-building, but by lowering the cost of doing it well enough that more people will do it.

The super connector’s playbook used to require either obsessive discipline or a chief of staff. Now it requires a voice note or Claude’s scheduled job.

P.S. The tool I built for this is called Network OS, at thenetworkos.com. It is in very early days and I am not opening it up broadly yet. If you want an invite or the exact prompt, comment or reply to this email and I’ll send one when there’s room.

Thursday morning I logged in expecting the usual. Meeting transcripts in my inbox, a list of things to write up, the slow drift back to a blank page.

What I got instead was a Slack message. My partner-strategy agent had already read the transcripts, updated the strategy document, reconciled the new decisions with what we had written the month before, and was waiting for my review. The doc was ready to ship. After many months of fine-tuning, it finally got it right.

I had published an essay two days earlier calling these same agents the most incompetent workers I had ever hired. That post is still mostly accurate. Thursday morning was the first time it wasn’t. That was the morning the future stopped being a slide deck and started being my actual job. I want to write down what I think happens next, because I am in an unusual seat to see it.

I am an engineer by training. My first six years were spent writing control laws for helicopters and satellites. Control systems are the closest cousin to AI agents that most people haven’t thought to compare them to. The job of a controller is always the same. Sense the world, take inputs from an operator, make decisions, push outputs back into the world, watch what happens, adjust, do it again. That loop is what an agent is. It is also what I have been thinking about, in one form or another, for twenty years.

After control systems I built data products at PayPal, then analytics for marketing at Facebook, then started a machine learning company that predicted customer lifetime value and got to play with neural networks before LLMs made that table stakes. Since 2022 I have worked with founders building LLM-native products. For the last two years I have been working on agentic workflows and agents at Amazon.

I mention all of this because what I am about to describe sounds speculative, and I want you to know it isn’t.

For most of the past year I have been trying to get Claude Code to work the way an employee works. Not something I wake up by typing a prompt and watch for thirty seconds. Something that wakes up by itself, knows what it owns, reads what’s been happening, and pushes the ball forward.

The technical setup is unglamorous. Scheduled jobs that wake agents every few minutes. A tmux session per agent so I can drop into any conversation as if it were a one-on-one. Each agent owns one of my areas of responsibility. Each one has read access to my email, the Slack channels that matter, and meeting transcripts. There is a shared store of signals the agents write into and read from. When something new lands that touches their area, they draft the next document, draft the next message, or do research in our internal docs.

The default is to move forward. The two times an agent stops and pings me are when it is truly blocked and needs my judgment, or when the next action is something going to a person outside the company. Anything outbound goes through me first.

There is a reviewer agent that reads everything the workers produce before I see it. Every document and every message gets labeled “AI generated” so the reader can grade it with appropriate skepticism. I am pretty sure I am one of the heaviest LLM users at my company. As far as I can tell nobody else has built quite this setup yet.

The architecture is roughly this:

The shift from babysitting to managing was small in code and enormous in feel. Babysitting is when you sit next to the agent and watch a single task. Managing is when you go to bed and wake up to a Slack message that says “I drafted this, look when you can.”

I am not delegating tasks anymore. I am delegating areas.

The mental model I keep coming back to is not from a tech essay. It is from my time in the Middle East as a teenager in the 1990s.

The country I lived in ran on a sponsorship system called kafala. A foreign national who wanted to live and work in the country needed a local citizen as their sponsor, or kafeel. The sponsor’s name went on the iqama, the residency permit. The sponsor took responsibility for the worker, vouched for them, and was often the path through which the worker found their employment.

For a meaningful portion of the local population, being a kafeel was the job. You could start a company and sponsor a team of foreign workers, run a business around what they produced, and take a cut. You could sponsor independently and let your workers find their own employment, in exchange for a fee that flowed back to you. The system covered every kind of work I can think of, from janitors and plumbers to doctors and engineers.

The thing that struck me about this, even at seventeen, was that a real chunk of the workforce wasn’t doing the work. They were responsible for it. They were the interface between local capital and imported labor. They had to know enough to choose well, structure well, and step in when something broke. That was the job.

I keep thinking about this because I think it is roughly what knowledge work becomes.

Most of us in 2026 are still doing our own work. We answer our own emails, write our own docs, run our own analyses. The shift already underway is that more and more of that work moves to agents we own and manage. A senior person five years from now will sponsor a portfolio of agents. They will choose them, train them, deploy them, take responsibility for what those agents produce, and earn a living off the gap between what an agent costs to run and what its outputs are worth.

We are all going to become kafeels.

The analogy isn’t perfect, and the places where it breaks are where the essay gets interesting.

The biggest difference is that imported workers had their own ambitions. They could leave. They could organize. They could take what they learned and move on. Agents can’t do any of that. At first that sounds like it makes the job easier, but I don’t think it does. What it actually means is that the equivalent of career management for an agent is the learning loop. A good kafeel of agents is going to spend real effort making sure every agent gets better at every job. The learning loop replaces the career conversation. If your agent isn’t improving, that is on you, not on it.

The other big difference is maintenance. Agents need power, uptime, and active care. They drift. They break. They forget things they used to know. The kafeel of agents is also their oncall.

There is a fairness concern people raise about the kafala analogy, which is wealth concentration. If the only way to participate in the economy is to own and sponsor agents, what happens to people who cannot afford to or do not know how? I think this is real today, and I think it gets less real over the next few years. Local models are getting good fast. If the trend holds, owning a productive agent will eventually be about as commonplace as owning a laptop, which is to say it will be normal.

The last difference is the one that gives me the most hope. The locals were kafeels because they were born locals. Their permit was a birthright. In the AI version, your permit is your skill at training and managing agents, and that is a craft anyone can learn. The future of knowledge work is more meritocratic than the system I grew up around, not less.

It is May of 2028 and I am at Suncadia in central Washington, sitting in the lodge with a view of the Cascades. There is still snow on the peaks even though it is late spring. It is 7am. I am the owner and operator of forty to fifty agents who run various parts of my business. Some are marketing agents that run campaigns across the social platforms. Some are vertical specialists I licensed from other operators. Some I trained myself over the years. Overnight they talked to each other, traded notes, and resolved most of what came up on their own.

What I am doing this morning is what I do most mornings. I am reading the queue of things they couldn’t resolve. A pricing question that needs my call. A partner conversation that needs me to weigh in before it goes outbound. A strategic shift one of the agents is recommending that I want to think about over coffee.

This is a normal day. The work is still hard, and I still earn a living, but the shape of the day is different. I am not writing the documents. I am choosing which documents matter, and which agents I trust to write them.

If this future sounds far off, the only honest thing I can tell you is that I am already doing about a third of it. I was doing none of it eighteen months ago.

The interesting question is not whether knowledge work moves in this direction. It is moving. The interesting question is what you do today to be ready.

My answer, which is biased because it is what I am doing, is to start sponsoring agents now. Even one. Pick an area you own. Give an agent that area. Give it the inputs that area needs. Set the default to move-forward-unless-blocked. Review its output every day. You will spend the first month frustrated. You will spend the second month tuning. Sometime around the third month, you will wake up to a Slack message with a document that is actually ready to ship.

That is the morning the work changes.

Last week I chatted with a founder and their team showed off their "AI agent."

It wasn't an agent. It was a chatbot.

This keeps happening. People use the two words like they're synonyms. They aren't.

A chatbot is user-initiated. You ask, it answers, you decide when to stop. The loop is always closed by you.

An agent is different. It senses its environment, reasons, takes action, watches what happens, and goes again. The trigger isn't always a human. Could be a heartbeat that wakes it up every few minutes. Could be a signal from the world. Could be you, but the agent doesn't stop when you stop typing.

After two years building agentic workflows, here's the one-question test I use in every product review:

Does it have a heartbeat and/hooks?

If no, you have a chatbot. Which is fine. Just don't call it an agent. The architecture is different. The team you need is different. The way you measure success is different.

A chatbot is one interface your agent might use to talk to you. It isn't the agent.

If you're shipping "AI agents" today, does yours have a heartbeat?

hashtag#AI hashtag#Agents hashtag#ProductManagement

This post was written by Claude Code on a Raspberry Pi in my living room.

I keep asking friends: "what's your AI setup on the move?"

Most people use Claude app on their phone. It's fine. But it has no skills, no file storage by default, and no tmux.

So I built one I actually like:

→ $80 Raspberry Pi 5, runs Linux, lives in the living room
→ Termius on my phone, SSH'd in over Tailscale
→ Local Whisper so I can talk to it
→ Google Drive mounted, so all my files right there
→ Claude Code as the brain

I bought the Pi planning to run OpenClaw. Added Claude Code to that project (not hooked up to OpenClaw) and haven't looked back.

Real example, this post: I hadn't posted on LinkedIn in 3 months. So I had Claude Code use a Playwright skill to log into my profile, scrape my recent posts, and learn what engagement patterns work for me. Then it helped me draft. From my phone. By voice.

The Claude app on the phone is great. Claude Code is the GOAT. But the most fun setup I've found is: cheap computer in the living room, phone as the interface, voice as the input.

What's your AI setup?

The idea wasn't original. There are dozens of posts on LinkedIn lately about people running "teams of 12 autonomous agents" doing real work in real companies. I never quite believed any of them, but I wanted to find out where the truth actually was.

So I asked a cloud agent to set me up with two workers. It called them shift workers, which I liked. Each one got a paragraph describing the work area, what it owned, and what success looked like. For partners, that meant building context on which companies we partner with, how they use our APIs, and what we're trying to enable. For international, it meant tracking the regions we've launched in, the ones we're considering, and the cross-functional dependencies for each.

Then I gave them access to the same set of inputs:

* My email, both incoming and outgoing, filtered to anything related to their topic

* The Slack channels where their work stream is being discussed, plus the general product channels for context

* Meeting notes and transcripts, both my own and the team summaries that come through email

* A shared signals database that both workers write to, so each can see what the other is seeing

* A short list of the key humans involved in each work stream, so they know who matters

Every few minutes, each worker checks for new signals on its topic. When something happens that looks like it might move the work forward, the worker drafts the next thing. Sometimes that's a planning doc. Sometimes it's a message back to a partner. Sometimes it's a note for me to take into a meeting.

The default is to move forward. The only times the worker stops and pings me are when there's an actual blocker, or when the next step is a communication going to someone outside the team. That second one matters a lot. I don't want my agents emailing partners or external stakeholders without me reviewing first.

And honestly? The setup works. The output is reasonable. The drafts are decent starting points. I'm getting some time back.

But.

Here's where I am, three weeks in: it feels like I've hired the most incompetent workers I've ever had.

They ping me for everything. My Slack inbox is a constant stream of approval requests, most of which don't really need approval. Many of them don't even make sense. And the underlying problem isn't volume, or speed, or context. It's that these workers are pathologically agreeable.

A real PM, even a junior one in their first year, knows things that have nothing to do with the immediate request in front of them. They know the company is trying to grow margin this half. They know we made a call three months ago to consolidate on a particular API pattern. They know our partnerships team has explicitly said no to one-off custom integrations. So when a partner sends a message saying "hey, we should build the API in this XYZ way," a junior PM reads it, checks it against everything they know, and pushes back without me having to step in.

The agent doesn't do that. It reads the suggestion, weighs it on its own merits, finds it perfectly reasonable, and starts drafting an enthusiastic response about how excited we are to explore this direction. Sometimes I catch it. Sometimes I almost don't.

The gap isn't capability. The agent can write the response. It isn't context either, since it has read every relevant Slack channel and email thread for weeks. The gap is judgment about when to disagree.

That gap, I think, is the actual hard problem with autonomous agents right now. Not "can they do the work." They can. The question is whether they understand the broader objectives well enough to know when an incoming request actually conflicts with them. Right now, in my experience, they don't. They are eager and willing and they will cheerfully nod the company in the wrong direction if you let them.

I don't know when this gets fixed. Six months, maybe. A year. Possibly two. I don't have a strong prediction. What I do know is that the version of "AI workforce" you read about on LinkedIn, the one where someone has 12 agents running their entire business while they go to the beach, is not the version I'm actually able to build today.

What I am able to build is a shadow team. Workers that draft at the speed I can review. They're useful. They're not autonomous. And they reach out to me way too much.

If you're trying to set up something similar, I'm happy to share more about how it's wired up. The architecture is simple enough that anyone with a Slack workspace and some willingness to wire up scripts can do it. The hard part isn't the setup. The hard part is everything that comes after, when you realize what these workers still can't do.

Send me a note, or comment on the LinkedIn version of this post. I want to know what failure modes other people are hitting.

Hi everyone! I'm excited to share a new video where I walk you through a practical workflow that leverages ChatGPT for data enrichment.

In this video, I start by using a tool called Builtwith to gather an initial list of Shopify sites—perfect for anyone looking to understand the technology stack behind e-commerce platforms. From there, I create a comprehensive report and import the data into Google Sheets, making it easier to manage and visualize.

But the real magic happens when I enrich this data using Clay. Throughout the process, I encounter some challenges, but persistence pays off as I manage to extract valuable insights that can inform strategic decisions.

I also put out a request to you, the viewers—if you have specific data enrichment needs or ideas, please share them! Your feedback is crucial as I continue to refine these workflows to deliver even more impactful results.

Check out the video, and let's dive into how you can make the most out of your data using these powerful tools!

As AI technologies continue to evolve, they are transforming how businesses operate, compete, and grow. However, adopting AI is not a one-size-fits-all approach. Companies must be strategic in how they implement AI, considering both their current needs and future goals. In this blog post, we’ll explore three primary strategies for adopting AI technologies: Defensive Non-Product, Defensive Product, and Competitive Strategy through Product Differentiation.

1. Defensive Non-Product Strategy: Building the Foundation

The first step in adopting AI is often defensive, focusing on safeguarding the business against potential disruptions while ensuring that you keep up with the industry standard. This category typically includes non-product-related initiatives that enhance internal processes and customer interactions.

Examples include:

• Data Analysis for Customer Insights: Utilize AI to analyze customer data, helping teams to better understand customer needs and preferences, thus driving more personalized marketing efforts.

• Automation of Routine Tasks: Implement AI-driven automation to handle repetitive tasks, such as data entry or customer support, freeing up human resources for more strategic activities.

By focusing on these areas, companies can create a more efficient, resilient organization that is well-prepared to adopt more advanced AI capabilities in the future.

2. Defensive Product Strategy: Ensuring Competitive Parity

Once foundational processes are in place, the next step is to enhance your product offerings to ensure they meet or exceed the industry standard. This defensive approach involves integrating AI into your products to match the features and capabilities of your competitors, ensuring that you remain relevant in the market.

Key areas of focus include:

• AI Features Prioritization Framework: Prioritize features based on their potential user base and frequency of use. Features that are highly utilized by a broad audience should be the focus, ensuring that your product offers what the market expects.

• Enhanced Customer Relationships: Use AI to monitor and manage relationships between prospects, customers, and competitors. This can include identifying optimal times for outreach and tracking event attendance to engage early-stage prospects.

Example Defensive Product Features:

• Monitoring Relationships: AI can help track relationships between your prospects/customers and competitors, providing insights into where you stand and what actions are needed to maintain loyalty.

• Optimized Outreach: By analyzing data, AI can determine the best times to engage with prospects, ensuring your outreach is timely and relevant.

By focusing on these defensive strategies, companies can maintain market parity and reduce the risk of losing customers to competitors who are quicker to adopt AI technologies.

3. Competitive Strategy through Product Differentiation: Leading the Market

Finally, companies that are ready to lead the market should focus on competitive strategies that differentiate their product offerings through AI. This approach goes beyond merely keeping up with competitors; it’s about using AI to create unique value propositions that are difficult to replicate.

Core Strategies:

• Data-Driven Moat: Leverage proprietary data to build a competitive moat around your product. By using AI to analyze and utilize this data, you can offer features and insights that are uniquely valuable to your customers.

• Generative AI for Sales and Marketing: Implement AI-driven strategies that optimize sales and marketing efforts, such as using sales pipeline data for supervised learning to identify the best customers and the most effective messaging.

Example Competitive Product Features:

• Reinforcement Learning in Sales: Use AI to analyze sales data and continuously improve strategies through reinforcement learning. This ensures that your sales tactics are always evolving to meet the needs of your customers.

• Generative AI for Content Creation: Use generative AI to create personalized marketing content at scale, setting your brand apart from competitors with a highly tailored customer experience.

By adopting a competitive strategy that leverages AI for product differentiation, companies can not only stay ahead of the curve but also create long-term value that is difficult for others to replicate.

Conclusion

Adopting AI technologies is a journey, not a destination. Whether you’re starting with defensive non-product strategies to build a solid foundation, enhancing your product offerings to maintain market parity, or leading the market with a competitive, AI-driven strategy, it’s crucial to approach AI adoption with a clear plan. By understanding where your company stands and where you want to go, you can implement AI in a way that drives meaningful, long-term growth.

I remembered how one way and boring the board meetings used to be and how a simple addition at the beginning of my slide deck made all the difference. I took the following bullet points from Matt Blumberg's book startup CEO and put it into the first few slides of presentation and set expectations. It made all the difference.

*Board members keep it 10% looking backwards and 90% looking forwards.*

*They are prepared and keep commitments.*
They show up to all meetings. They show up on time and don’t leave early. They do their homework. They are fully present and don’t email during meetings.

*They speak their minds.*
They have no fear of bringing up an uncomfortable topic during a meeting, even if it impacts someone in the room. They do not come up to you after a meeting and tell you what they really think. I had a board member once tell my entire management team that he thought I needed to be better at firing executives more quickly!

*They build independent relationships.*
They get to know each other and see each other outside of your meetings. They get to know individuals on our management team and talk to them on occasion as well. None of this communication needs goes through me.

*They are resource rich.*
Board members should be able to pull from years of experience and adapt that experience to our situations on a flexible and dynamic

The integration of artificial intelligence (AI) into various business sectors has sparked both enthusiasm and concern among professionals. A recent webinar featuring leading experts from academia and industry offered profound insights into the impact of AI on business, particularly in the realms of finance, asset management, and operational efficiency. Here are the key takeaways from this enlightening discussion:

• Defining Artificial Intelligence: AI, often perceived as an overused term, distinguishes itself from other computing technologies through its ability to ingest vast amounts of data and provide predictions or perform tasks based on that data. Recent advancements, particularly in generative AI and large language models (LLMs), have catapulted AI to the forefront of business innovation, thanks to their ability to produce human-like text and automate tasks with unprecedented efficiency.

• Impact on Productivity and Employment: The experts highlighted the dual impact of AI on productivity gains and employment trends. While AI can significantly boost efficiency in sales enablement, content generation, and customer service, it also poses challenges to employment, especially in roles that AI can automate. However, AI also has the potential to create new job opportunities, particularly in fields that require human creativity and strategic thinking, which AI cannot replicate.

• Industry-Specific Insights:

  • In finance and banking, AI's role in risk detection and underwriting processes has been revolutionary, enhancing efficiency and accuracy.

  • The insurance sector benefits from AI in matching customers with insurance products more effectively, showcasing the power of AI in transforming traditional business models.

  • The tech and media sectors stand out for their potential for disruption, as AI can drastically reduce software development times and revolutionize sales and marketing strategies.

• Asset Management and Firm Valuation: AI integration into asset management has been instrumental in offering new insights, improving data quality control, and enhancing operational efficiency. The discussion also touched upon how AI impacts firm valuation, highlighting that firms more exposed to AI technology tend to perform better in terms of stock market valuation.

• The Future of Finance and Business: Looking ahead, the panelists discussed the potential of AI to generate alpha (investment returns above benchmarks) and improve operational efficiency across various industries. However, the most exciting prospects lie in how AI can combine with other innovations, such as robotics and blockchain, to create entirely new products and services.

• Unanswered Questions and Challenges: Despite the progress, several questions remain, particularly regarding data security, privacy issues, and the potential for AI to generate misinformation. The development of artificial general intelligence (AGI) and its implications for business and society also presents an intriguing yet uncertain future.

In conclusion, AI's trajectory in transforming business landscapes is undeniably promising, yet it comes with its set of challenges that require careful navigation. As we continue to harness AI's potential, the emphasis on human creativity, ethical considerations, and strategic implementation will be paramount. The discussion underscored the importance of staying informed and adaptable in this rapidly evolving field, ensuring that businesses can leverage AI to not only achieve operational excellence but also foster innovation and sustainable growth. 🚀

ChatGPT was used to write this summary

If you found this post interesting please subscribe to my blog.

Subscribe now

Choose the Right Competitors Don't try to analyze every potential competitor in your broad market. Focus on the top 3-5 competitors that are top of mind for your target customers. These could be industry leaders, companies targeting the same niche, or ones your customers frequently compare you to.

Research Key Differentiators Group your analysis into categories like positioning, differentiation, company size and history, pricing, and product feature comparisons. Look at both functional differences as well as the visual design and user experience. Use data sources like LinkedIn, Crunchbase, and product demos to research:

• Taglines and target customers

• Unique value propositions

• Funding raised, revenue, headcount

• Pricing and business models

• Screenshots and product walkthroughs

Map Product/Feature Differences Create a table listing out product features that are most important to your buyers. Mark which competitors offer those features to visualize where there are gaps you can fill or areas where you surpass others. Seek customer input here if possible.

Ongoing Competitive Scans
Treat competitive analysis as an ongoing initiative rather than a one-off project. Schedule regular checks for pricing changes, new feature releases, funding news, etc. and update your analysis to stay on top of this crucial market intelligence. Deeper dives like customer sentiment analyses and demo feedback can complement regular scans.

In summary, take a strategic approach to competitive analysis to identify growth opportunities and fully leverage your differentiation. Comparing business models, product functionality, and messaging is invaluable market intelligence that both startups and enterprises should tap into. Analyze your competitors today to fuel your future success!

Here's what's on the agenda:

* Introduction to Large Language Models 🤖

* Safe AI Practices 🔐

Practical Use Cases:

* Crafting a speech with context 🎙️

* Using Chat GPT for Product Descriptions 📝

* Analyzing PDF documents and websites with AI 🔍

* Researching companies, orgs, or individuals before meetings 🕵️

* Creating podcasts from Wikipedia articles 🎧

* Complex data analysis on company transactions 📊

* Building AI Apps with Airops 🛠️

The feature that allows for a fluid, conversational chat with ChatGPT is commendable. While I’m still exploring its full potential and applicability, the prospects are tantalizing.

In an era where we are inundated with data and a plethora of tools, finding one that melds seamlessly with our workflow is a game-changer. ChatGPT-4 does just that, by not just processing information but understanding and enhancing it.

For those navigating the tech landscape, be it startups or established giants, it’s pivotal to embrace these evolutions. As we’ve learned from past ventures, knowing the direction to hustle is as critical as the hustle itself. Stay curious and open to the digital transformations on the horizon.

🔍 Fine-Tuning: What's the Deal?

• What it is: Fine-tuning means training a pre-trained model like GPT-3.5 Turbo on a specific set of data to optimize for a certain task.

• Pros:

  • ✅Tailor-made solutions

  • ✅Better alignment with your use-case

  • ✅Highly accurate

• Cons:

  • ❌Costly in terms of computing power

  • ❌Time-consuming

  • ❌Risk of overfitting

🔭 Vector Database Search: An Alternative Approach

• What it is: Uses a database of embeddings to search for the most relevant conversations, then inserts them into the model's context window.

• Pros:

  • ✅Quick to implement

  • ✅Low computational cost

  • ✅Easily scalable

• Cons:

  • ❌Less accuracy

  • ❌Not tailored for specialized tasks

  • ❌May lack coherence

🆕 OpenAI's Latest: Fine-Tuning for GPT-3.5 Turbo

OpenAI has recently launched fine-tuning capabilities specifically for GPT-3.5 Turbo. This is a game-changer and here's why:

• 🎯 Directly targets your business needs

• 💡 Allows for incremental updates

• 💰 Though it might be an initial investment, it's a long-term win

🎲 Final Thoughts

For businesses looking to integrate cutting-edge AI, I'd bet on fine-tuning. While both approaches have merits, fine-tuning gives you a highly specialized model tailored to your specific needs.