I believe the opposite is true, that AI will actually usher in a golden age for cold outreach. Because unlike content marketing and other channels, AI solves the problem it creates.

The problem with cold outreach

Getting cold outreach sucks, for reasons that are pretty straightforward:

• Most messages are entirely irrelevant.

• The ones that are potentially relevant are mostly from low quality providers.

• There may be diamonds in the rough, but it takes too much work to vet them.

And yes, automation is going to make all of those things a lot worse in the short term. But relative to other channels, it also more clearly provides the solution.

Most companies die unnecessary deaths

Cold outreach is all about relevancy. If you can find the right message for the person at the right time, you don’t even need mutual connections or a recognizable brand — they will respond because it feels like you are reading their mind.

The need for relevancy creates the cold outreach problem. Especially without the tools to fine tune the process, the simplest solution is just to send a ton of messages. Hey {name} I think {your company} might like my product?

But that doesn’t mean outreach is tapped as a channel. Most startups will die before ever finding or contacting their ideal customer. And on the flip side, too many companies die unnecessary deaths before finding the right solution to their problems. Which is why AI is much more likely to scale it 1000X than drive it to 0.

The world of 1000 assistants

One way I like to imagine the future is to picture a world in which everyone has 1000 assistants. What would you have all of those assistants do every day?

It’s a useful frame because you would quickly run out of automating the things you are already doing, which are the jobs most people think of when they imagine AI agents. Maybe you take the first 100 assistants to book your travel and record your meetings and do the busywork in your job. What about the other 900?

The answer is fairly obvious — you would have those 900 assistants talk to the assistants of other people. Once you start thinking like that, you can actually imagine needing a million or a billion assistants, which is the right way to think of the potential scale of AI.

Now imagine each company has one assistant assigned to outreach each high potential customer. And that potential customer has one assistant assigned to triage each outreach message. That’s how you get the golden age of outreach.

The outreach renaissance

Interestingly, I find it harder to make the same rosy argument for other major channels like content marketing or advertising.

For one thing it’s just not clear why you would ever need those channels in a world where cold outreach is scaled one million times. Why would I create generic content when I could enter into a detailed conversation with each potential customer?

A more nuanced take is that outreach, content marketing, and advertising will all converge. Content marketing will be totally individualized and you’ll pay for outreach in the channels where customers live, just like we pay for advertising today.

But that final form feels much closer to outreach than any of the other channels today, and will emerge gradually via iterations. So if you want to bet on that future, the place to invest today is in sophisticated prospecting and outreach automation.

Go forth and outreach

I’ve heard peers argue that the flood of AI tools that have launched to do prospecting and outreach automation are basically performing market arbitrage — scooping up the value in outreach while they drive it out of existence.

There may be a weird transition period where the conversion rates go down, but cold transactions are simply too valuable to the economy for there to ever be a true cliff. New tools will crop up to solve the problems that others create. And in the end, we’ll all get our 1000 assistants.

We’ve been investing in outreach ourselves and building custom agents to do prospecting in biotech for customers, and my strong feeling is that we are much more likely to enter a golden age than an outreach winter — something I can’t say for content marketing.

P.S. if you’re curious about our stack, I’ve tested dozens of AI outreach tools in our research and strung together a few to make our own process work for biotech. Right now we’re using a combination of Clay, Apollo, Attio, and La Growth Machine — if you want to know how it works just reach out and I’m happy to share.

“The workforce of the future will include both humans and AI agents, and businesses that don’t learn to manage this incredibly complex reality will quickly fall behind.”

Workday realizes what is now obvious — real AI adoption at enterprises will happen via grooves laid by humans. Smart companies won’t sell AI software, they will help companies “hire,” “onboard,” “promote,” and “fire” digital workers. AI systems are being wheeled into organizations in Trojan horses shaped like employees.

As someone actively building and selling AI agents to biotech companies, I can assure you this reframe is immensely useful. And for what it’s worth it isn’t just marketing; the tech has advanced to the point where agents are behaving as humans would — outcome oriented, with agency to get there.

But there is an obvious drawback to thinking of AI in human terms — AI agents behave in ways that humans don’t (and can’t). So it’s worth asking, where does the analogy of AI agents to humans break? And what opportunities does that open to do things differently?

One critical difference is onboarding. Unlike humans, AI agents will need to be “onboarded” every time they go to perform a task, which creates a big challenge as well as a huge advantage.

The Best Agents Need to Be Onboarded (Just Like People)

Common wisdom is that employees only start delivering real productivity 6 months into their job — onboarding is a tricky problem for humans, and it will be for agents as well.

In niche industries, the onboarding challenge is even harder. As an illustrative example, walk through a biotech company and you'll find former scientists in unexpected roles — running business development, leading sales teams, even managing recruitment. Why? Because teaching complex scientific concepts to outsiders is so time-intensive that it's often easier to teach business skills to scientists than the reverse.

The need for domain and company knowledge also explains why niche industries struggle to adopt AI, even for workflows where AI is clearly a game changer (BD, sales, recruiting, content creation, etc).

We hear this directly from customers. For example, existing AI tools work great for recruiting if the job is just to parse LinkedIn profiles, but don’t work if you need to hire a scientist and compare their publications against your company’s research. Clay and other BD tools work out of the box if your prospects are easy to identify, not so much for the deeper analyses and opportunity identification BD people in biotech are doing. For those workflows, you need agents that are more deeply onboarded to the problem.

Onboarding in AI is an unsolved problem

The need for well-onboarded agents is obvious. the right way to build them is not.

For a while, RAG (retrieval augmented generation) was the gold standard for custom agents. But if you’ve used company support chatbots, you’ve experienced how RAG doesn’t feel as dynamic as a human. In fact, Hebbia did a study that claims RAG failed for 84% of real-world user queries in a chat based interaction.

A big step forward will come when we get better foundational models for niche areas, like Deepmind and BioNTech’s Laila, a “lab assistant” built on top of Llama 3.1. OpenAI released GPT 4b (bio) a few weeks ago. But having better foundational models for bio is like having a better talent pool to hire agents from — it doesn’t change the fact that you need to onboard agents to your company’s specific knowledge, especially since a lot of that information is private and inaccessible to the foundational models.

The best bet for how proper onboarding will work is reasoning agents like OpenAI’s Deep Research. If you haven’t used it or seen a demo, you should. Reasoning agents behave like a human would in an onboarding process —gather information, think through it, decide what new information to gather, and so on.

Here is where the analogy to humans starts to break though, because Deep research is onboarding itself for each new query in a matter of minutes. What it learns in a given search doesn’t carry over to the next one, which (perhaps counterintuitively) presents a huge and obvious advantage.

Unlike Humans, AI Agents Will Be Re-Onboarded Constantly

A huge problem for companies is that employees not only need to be onboarded, they need to be constantly re-onboarded as the technology advances.

The re-onboarding problem can be summarized as “good-enough is the enemy of perfect.” Why would a BD or sales person take the time to properly retrain when their existing knowledge — while outdated — is good enough for them to do their job? Companies collectively spend hundreds of billions of dollars trying to solve that problem.

The good news is that reasoning agents like Deep Research do not have the same issue. They get onboarded on the fly for each new task, along with context-specific information. It’s like being able to clone your best account rep and focus each clone on a different customer.

Onboarding agents will also use a very different set of tools under the hood — they will not just access your company Notion. In fact, the knowledge bases agents use will likely be unusable to humans.

How will Notion for agents work?

It’s amazing how ubiquitous and valuable knowledge management systems like Notion or Confluence are despite doing a horrible job of solving the problem they are meant to solve. This isn’t a dig at those companies — building knowledge management for humans is just a wicked problem.

An analogy to tech debt is useful here. Knowledge, like code, is cheap to create but expensive to refactor. As a result, knowledge and code bases tend to get more brittle and harder to use over time.

Knowledge management for agents will be vastly superior to knowledge management for humans for the same reason tech debt will kill fewer companies in the future. AI not only makes it trivial to rewrite a code of knowledge base, it also doesn’t care about relearning how to use it afterwards.

How will agents prune their own knowledge base as they perform new tasks? We can see interesting glimpses via companies like Shelf, which rebuild a company’s FAQs as humans interact with it via agents. Experimenting with how to solve that problem for biotech is occupying a lot of our time at the moment.

The talent war for agents begins now

This all sounds promising in theory, but where will the cost of training well-onboarded agents pay off first? Here are a few examples we have been looking at in biotech that can be generalized to any specialized industry:

• Recruiting: onboard the agents to the company’s unique tech, then have them go find and evaluate candidates that can expand/strengthen that focus.

• Market research: onboard the agents to a company’s unique tech and position in the market, then have them continuously search and distill insights as the market evolves.

• BD/Sales: onboard the agents to the company’s unique tech and opportunities for expansion, then have them search and prioritize partner opportunities.

The pattern is clear: while generic AI can deliver value out of the box, the real transformation will come from agents that deeply understand your company's unique technology and context. Just as the best human employees aren't just functionally competent but deeply aligned with your company's mission and technology, the best AI agents will need to be "company natives" – even if they achieve that status in minutes rather than months.

This isn’t a daring prediction so much as a foregone conclusion. Consider the analogous question of whether AI will make fake news worse. Duh.

Science is no different — it has it’s unique quirks but the general forces at play are the same.

So the better question is how to stem the tide in the short term and whether we’ll land somewhere far better on the other side. Those answers are far less obvious.

Science works backwards

One thing non-scientists often fail to realize is that science works from belief to data, not the other way around. People imagine scientists surfacing data in a totally unbiased way, without any preconceived notions of what it will show them.

With few exceptions, it is the opposite. A scientist comes up with a hypothesis and then works to prove it.

That should give you a feeling for why LLMs are such a powerful potential accelerant of irreproducible science. LLMs are notoriously easy to bias — frequently making up citations or entirely inventing the data underlying their claims.

Lowering the friction to write papers is also dangerous. I wrote an article for the Scholarly Kitchen during the peak of the COVID preprint “infodemic” about what we saw on the front lines at ResearchGate. Especially when it comes to lit review articles, the ability for scientists to build reputation by creating AI-generated papers could quickly drown any fact checking systems we have.

Incentives, Incentives, Incentives

While difficult to fix, the reproducibility crisis has never been hard to understand — scientists need to build proof of work every few years to achieve the next milestone, but it takes a few years to run a proper experiment.

When I was in the lab, I spent 2 years pretty much full time on a single experiment. I was lucky to get some good data and a few publications out of it, but suppose I hadn’t.

That, in essence, is the root cause of the reproducibility crisis.

Being a junior scientist is like getting a coin to flip and being told that you only get to flip it once or twice before deciding whether to publish which way it’s weighted. If you don’t publish anything, your career is shot. And every few years, you repeat the process.

Replacing scientists

At the risk of ruffling some features, I will say the obvious out loud – junior scientists who spend 95% of their time doing grunt work are going to be replaced by robots.

This may actually be the most optimistic take. For one thing, being a junior researcher kinda sucks. I spent basically two years dissecting mouse brains and developing back problems from hunching over the cryostat all day (being tall in the lab is not fun).

And for reproducibility, it seems obviously better if a given scientist more closely resembles a principle investigator (managing a lot of experiments in parallel) than a PhD student. To go back to the coin analogy, if you get to flip the coin a hundred times, then the pressure to publish half truths is much lower.

In the distant future, a given scientist might even resemble an organization like Arcadia — not just doing a lot of experiments in parallel but bringing discoveries closer to market. In that world the incentives to publish false data is basically zero because it only makes it harder for you to produce impactful products. Why spoil your own soup?

Whether there will be more or fewer total scientists in that world is very similar to arguing about whether there will be more or fewer total engineers as AI gets much better at code. If individual leverage goes up, do we get fewer individuals or a much larger pie? It’s an interesting question that smart people are asking about society in general, which is also why it is out of scope of this article.

Perhaps academic publishing can’t save itself

If you follow my logic here, I think it’s reasonable to conclude that AI can solve the reproducibility crisis but only when it can dramatically increase the leverage of any given scientist (unless all scientists can go work for organizations like Arcadia, but that would require a large increase in billionaires looking to fund basic research).

Massive increases in speed and efficiency aren’t coming to academia via a lot of the AI for academia we have so far, such as publication search or literature review — though those products are awesome and solve real problems for scientists.

The simple fact is that scientists spend the bulk of their time in the lab, and making that 100X more efficient is much more dependent on laboratory automation. And even that is a fairly wicked problem especially in biology — you can’t change the time it takes to raise a mouse, for example.

So in the short term, I’d expect the situation to get worse. We can build bandaids — better peer review, LLMs for science with strong safeguards on source data, and so on. And as a society we should celebrate and encourage many more experiments like Arcadia that aim to offer alternatives to the traditional academic path. But it’s really in the long term as we manage to automate the grunt work out of science that we can truly address the incentive issue.

As always, I’d love to hear if you see things differently. And I’m always happy to connect with others working to use AI to make the system of science better.

The challenge isn't that basic tools like ChatGPT and Claude aren't useful—they're exceptional for rapid analysis and knowledge synthesis. But unless you take control of the source data they tend to smooth out the exact irregularities and edge cases that often contain the most valuable insights.

For example, if you ask ChatGPT to give you the org chart for a typical biotech company, it will do an ok job, but there's a clear regression to the mean — it makes faulty assumptions based on "standard" org charts. So you get 80% right but the 20% is actually what you most need to understand because it's what makes those companies unique.

We've recently built some workflows that address this issue and that have helped us move 10X faster (at least). Here's a view of our current stack which is tuned for biotech but applicable to this kind of work more generally:

1. Cast a Wide Net

AI means you can process a basically limitless amount of information, so the first step is to gather as much quality source material as you can. We pull from everywhere:

• Granola to record customer calls and other meetings with experts

• Latest content from a select number of industry podcasts and newsletters (via RSS)

• Conference talks and presentations

• Technical blog posts and papers

• Market research databases (Apollo, Clay, AlphaSense)

• Academic research

• Company blogs

The goal isn't to manually review all of this, it’s just to aggregate a corpus of quality source material relevant to our work. We also don’t worry about tagging this ourselves, we use AI for that too (more on that in a sec).

It's not a perfect solution, but right now we're maintaining our knowledge base of source material in Airtable. The main appeal is that it's fairly simple to use and build workflows on top of - especially for semi-technical BD people.

2. Build an AI Intern to Screen Content

In general you don’t need to be precious about the content you store, but adding irrelevant information makes everything harder (and more expensive). So we maintain a prompt flow that’s basically an AI intern that screens for relevance on content we’re not sure is quality.

The prompts include things like:

• Key technical areas we're investigating

• Specific market dynamics we're tracking

• Types of customer problems we care about

When we're not sure about something, we chuck it at the prompt and get back a quick relevance score with reasoning. If it’s good, we add it to the database.

Maintaining this screening system has become a valuable exercise in itself. Writing down what we care about and why - then updating it as our understanding evolves - forces us to maintain really clear thinking across our team about our research priorities. As with a lot of AI workflows, the process of creating clear instructions leads to clearer thinking overall.

3. Enrich Content at Scale

Once you have raw source material in a database, you can enrich it in any way you want. For example, run a prompt over podcasts to extract examples of places where companies are currently using AI within regulatory compliance. Or a simple scraper that grabs information from company landing pages.

We use Pipedream for this - it hooks up nicely to Airtable and makes it easy to run content through LLMs or scrape additional data.

These all become searchable columns in Airtable. Nothing fancy - just automated enrichment that would be tedious to do manually.

4. Research Reviews at Lightning Speed

Here's where it all comes together - and where ChatGPT has been a game-changer. When we need to investigate something specific (like "What workflows do platform biotech companies use to source potential partnerships?”), we:

• Pull relevant source material/extracted insights from our database

• Feed it into ChatGPT with context about what we're trying to learn

• Iteratively refine the analysis, asking follow-up questions, adding more sources

• Summarize the analysis using ChatGPT’s document feature

The magic? I can then just share the chat history with others on our team — they don’t just get the summarized doc; they can also trace back how I got there.

Shortening the Path to Market for Bio

The above is a scrappy system we’ve whipped up for ourselves at Robo over the past month or so, but we’re also working on some more advanced versions with a few design partners in biotech. There’s low hanging fruit for any company, but especially as you get closer to biological data you need better evals and a bit more infrastructure.

What I find exciting about this work is that it’s a bit of an arbitrage — especially for cutting edge biotech, a big challenge is solving the needle in the haystack problem. There is a good partner or customer out there for you, but there’s also a good chance you don’t find them before time runs out. AI’s ability to process immense amounts of data means we can increase the odds of that successful match happening.

We’re just scratching the surface here. If you're doing market research in biotech - whether that's competitive intelligence, customer discovery, or opportunity analysis - we'd love to compare notes. We're particularly interested in what data you’re attempting to digest/synthesize and where current tooling falls short.

I now use stylesheets fairly religiously, so I'll show you exactly how to create one and share a tool we built to generate them automatically from any Substack.

Why Writers Need Stylesheets

Designers have known this forever. Stylesheets—sets of reusable building blocks—are the secret sauce behind cohesive, on-brand visuals across products.

A writer's stylesheet is no different. It's a toolkit for maintaining consistency, reducing editing time, and creating content that matches the vibe of everything else you've put out.

If you use AI to generate content, having a stylesheet is a gamechanger. Simply attach your stylesheet to ChatGPT or Claude or whatever you're using and spend a lot less time editing the output. Plus you can use AI to create one in a few minutes.

How to Create Your Own Writer Stylesheet

1. Feed Your Past Writing Start with what you already have. Got a treasure trove of old tweets, LinkedIn posts, emails, or even text messages? Pull those in. The goal is to build a sample set that reflects your natural style.

2. Extract The StyleSheet Here's a prompt that works well (but definitely experiment for yourself):

   "You are an expert ghostwriter who manages a team of junior ghostwriters. I am going to give you a bunch of sample content from a client you are working with, and I want you to extract a stylesheet. It should be something you can give to your junior ghostwriters so they can reliably produce content that sounds exactly like the client's voice. Pay attention to tone, style, and sentence construction. Avoid giving overly specific examples from one piece, as those might be over-applied. Instead, provide multiple examples that highlight a clear style or tone."

3. Refine and Test Once you've got a draft of your stylesheet, test it. Save it and attach it to your conversations with any AI model. Generate content and see how close it gets to the real you. Tweak as needed.

Generate a stylesheet automatically from any Substack

If you write on Substack, Austin Kennedy and I built a quick tool that can generate your stylesheet automatically. Try it here.

Here’s the first bit of the stylesheet it created for Lenny’s Newsetter, for example:

Bonus Options

One unexpected benefit of creating a stylesheet? You'll gain insights into your own tendencies as a writer—both strengths and areas for growth. For example, do you overuse parentheses? Tend to write overly complex sentences? A stylesheet makes these patterns obvious.

Stylesheets can also be aspirational. Ask ChatGPT for feedback on your writing style, then use that feedback to refine your stylesheet into a guide for the writer you want to be.

And of course, this process doesn't have to stop with your own voice. If there's a writer you admire, build a stylesheet based on their work and see what you learn. If they have a Substack you can use the tool above to build one in seconds.

If you end up experimenting with this, I'd love to hear about it.

-Joey

Build was our personal awakening, but 2024 seems to be the year everyone woke up to the idea that software is getting cheap. Even my mom spent time over Christmas worrying about whether my brother should abandon his job as a software engineer and go back to school.

So what becomes scarce if it isn't code? If software is cheap, then the most valuable thing to own is clear opportunities for software to drive impact. It's like electricity suddenly becoming 10X less expensive overnight - the winners wouldn't be startups, they'd be the companies where energy costs are currently limiting growth.

This sets up the central tension that will define 2025: Existing companies have backlogs of good problems to solve but struggle to transform how they work. New startups can build AI-native from the start but have to earn good problems to solve.

If existing companies are going to win the tug of war, they will have to increase their software capacity by a lot. But how?

Some argue AI means less software outsourcing — the internal team just uses AI to ship 10X faster. I couldn't disagree more. The surface area of what's worth solving with code is expanding by an order of magnitude.

It's not just that previously expensive solutions are now ROI positive — AI agents mean problems previously solved by humans can now be solved with software. Assuming an internal team can cover the breadth of problems a company should solve with custom code mistakes engineering capacity for expertise.

This is why 2025 will be the year of the specialized roboagency. Not general-purpose development shops, but teams focused exclusively on specific types of problems. Sales automation experts who've built dozens of sales agents. Support automation teams who deeply understand customer service workflows.

The best roboagencies will look like typical tech companies from the outside. They won't sell the service, they'll sell the value. Sierra isn't selling custom chatbots, it's selling automated customer support - the custom buildout is just a means to an end. Sierra’s goal is to build infra to increase the margins of serving the next customer. If they fully automate the whole thing then it becomes pure SaaS, but only time will tell if that's possible. Most will remain software and services, especially for larger deals. Enterprise SaaS is already not so secretly a service heavy model anyways.

Some people call these “AI enabled services” probably because it sounds more scalable to investors than roboagency (which I prefer just because it’s simpler), but under the hood it’s the same thing and they are popping up everywhere. 8090 is creating custom versions of enterprise SaaS, my friend Rick is building custom document automation for financial services firms at Strange Loop Labs, there’s versions for freight auditing and payment and countless others attacking different specific opportunities for custom code.

And that brings me to some news: We've decided to build Robo as a specialized roboagency to build agents for biotech companies. After starting my career in neuroscience labs at Johns Hopkins and UCSF and then spending years at ResearchGate and Benchling, it's the space where I understand the problems best and it’s where I met my cofounder Rapha. Cutting edge biotech companies have a unique opportunity to build custom agents to accelerate R&D and improve efficiency. We want to help them realize that opportunity so they can deliver world-changing innovations faster.

We have lots more to share about the agents we’re building for biotech, but in the meantime if you're interested in what we're building, or know someone at a fast-moving biotech company, I'd love to connect.

AI’s impact on software is obvious — it will get 100X cheaper. What happens downstream of that is difficult and mind-melting to imagine. So I’m going to think through my current vantage point, out loud, unfiltered.

Tech debt will kill fewer companies

One of the interesting things we’ve found is that it’s oftentimes easier to have the AI rebuild a product from scratch than trying to do a major refactor around a new feature. Part of this is due to a current weakness in AI at understanding the context of a codebase (which will disappear) but it’s also because AI removes the idea that writing quality code takes time.

Every engineer working at a big company dreams of rewriting the codebase, or thinks angry thoughts about the new feature the CEO steamrolled into the org that created a ton of tech debt. But given how good AI is at integrating and synthesizing lots of information, this will be far less common in the future.

Tech debt will always be a thing, but the pain of YOLOing that new feature or prioritizing improving customer value over code cleanliness will go down. There will be AI systems designed to clean up after you.

New bottlenecks will emerge

The other day someone asked me whether every company will have an internal roboagency that delivers new products at lightning speed. My simple answer is that I don’t think companies would know what to do with all of those products.

If you’ve worked in a tech company, you’re familiar with the marketing or sales team complaining that the product org isn’t shipping that new thing which would really make a difference to customers. In general, engineering velocity has always been the biggest bottleneck for startups. But that’s about to change — while all functions will be augmented with AI, the transformation in engineering will happen first.

So another place to look for startup opportunities is to think about the functions around product and engineering that will become bottlenecks. For example, if the product team can launch major new features or iterations every few weeks, the ability to train sales reps or educate customers on the latest version goes from being a somewhat annoying problem to a business critical problem.

Each company will have their own OS

I think we underestimate how much software has driven cookie-cutter company structure and operations vs the other way around. Finance, legal, sales, marketing — they are indeed separate needs but the degree of standardization in how they are achieved comes in part from the siloed nature of the tools we have today.

There’s an enormous opportunity in the age of AI to build the unique operating system that represents a company and enables an emergent constellation of custom tools on top.

This is why Palantir is so well positioned to ride the AI wave in defense, because they’ve been working for a decade on mapping customer ontologies. What everyone seems to be realizing is that the hardest and most complex work to build AI enabled software for companies is data work.

It seems unlikely for there to be one big winner here. For one thing, building an OS for a specific industry isn’t just a database problem. It’s about building a language that can be used to connect the pieces, one that makes trade-offs based on the characteristics of that market. Palantir’s customers value security and secrecy above all else, and that informs how they build the platform.

The obvious place to start is with the most important data — customer data — which is one good reason Salesforce is in the crosshairs for 8090 and all of the companies pitching custom software. Land in the data and you can expand to everything else. But making the leap from a high end software agency to a platform is not easy.

More companies will have tools like FAANG

Tech giants with an excess of money and engineering talent build a huge % of their own tooling. If you’re an engineer at Google you don’t just use standard off the shelf products like a startup, you use custom versions of everything.

One specific place to find specific product opportunities would just be to audit the tools FAANG employees use, and find ones where having the custom version is a huge win for the company. Then, build a company that delivers that custom version to a lot of other companies for 10 or 100X less.

If you made it this far

Believe me I could keep going, but I’ll stop there for now. If I missed or got something wrong, I’d love to hear about it.

I’m also hosting a few workshops with leaders of forward-thinking companies over the coming weeks to explore how AI-generated custom software will drive growth or reduce costs. If you know someone great I should add to that list, please let me know.

AI products typically need to solve two fundamental challenges:

• Aggregating and structuring a useful dataset

• Building the AI workflow/interface on top of the data

Not only is the data pipeline often the more complex part, it’s also why a lot of ill-conceived AI products have a severe cold start — the AI isn’t useful until the data exists.

To show how simple it can be when you take the right approach, we built something. Last week I wrote about how companies will increasingly use software as content, so we decided to showcase exactly what we mean by building a valuable AI interface in less than a week.

Introducing Robodocs

Today we’re launching Robodocs — an AI assistant that turns your Google Drive into a Github profile, giving writers the same insights developers have about their work. And it took Rapha a week to build, part time. Want to see it in action? Watch the demo and test it out for yourself here:

Robodocs already does a lot. It ingests your Google Drive database, encrypts and analyzes all of the documents and your activity, and helps you stay accountable and gain actionable insights on your writing. There’s even an AI assistant that you can chat with who has access to everything you’ve written. Plus it’s fully hooked up to Stripe for paid subscriptions.

Our aim wasn’t just to show what we can build with AI in a few days, it was to hammer home the idea that the hard part about most AI products isn’t the AI, it’s the data.

Products like Robodocs are really simple to build, and for most companies they are a much better place to start vs something entirely new.

Opportunities for AI interfaces

If you’re looking for more inspiration, here are two types of AI interfaces we’re seeing:

SMS/email agents for seamless data entry
Think about the channels people already use constantly — SMS, email, Telegram. What if these became the natural interface for updating a key product or workflow? AI makes this possible by handling messy, conversational data and translating it into structured updates.

Here’s a real example: We're working with a company that's transforming how field sales reps update Salesforce. Instead of pulling up forms while driving between meetings, reps simply leave a quick voice note about their customer interaction. An AI agent follows up with any crucial questions, then automatically updates Salesforce. For sales people on the go, it's a game changer for their ability to keep data up to date.

Assistants built for a specific niche on top of general productivity tools
Many failed startups have tried to build “Google Docs for writers” or “Notion for {insert niche}.” The reason they fail is because getting to and keeping up with table stakes functionality is too hard.

Whether as a wedge for a new startup, or simply as an app to automate some specific process, these products are also shockingly simple to build (see Robodocs). A lot of the data plumbing has already been built by the underlying product, and building on top of existing data means your AI product has clear utility from day one.

Picking fruit in a platform shift

AI is a platform shift, akin to the shift to mobile. That opens the possibility for entirely new products, but it also means that virtually every existing product will get an AI interface. And if you want to go looking for low hanging fruit to build, that’s where you should start.

If you’re unsure about whether you have a good opportunity to build an AI interface, just send me a quick note about the data you have and the problem you’re hoping to solve and I can give you a very rough assessment of how expensive/time consuming it would be to build.

Building Points for Pies in 2019 likely cost Domino’s millions of dollars. But what happens when products like that can be built for a few thousand bucks? We’re starting to find out. Since we launched Robo, I’ve been surprised by how many of our clients are founders of existing companies looking to build products to supplement or replace their existing marketing spend.

The opportunity is obvious once you see it. Instead of fighting a losing battle against influencers to produce viral media, companies can increasingly market via what they do best — building products.

Just look at ChatGPT. What began as OpenAI's product demo became the most effective marketing campaign of 2022, reaching 100 million users faster than any product in history. As AI makes software cheaper and traditional content marketing less effective, software will be the new form of marketing for forward-thinking companies.

Companies aren’t influencers

Traditional content marketing has been dying for a while, but AI is going to rip the plug out. For what it’s worth, I didn’t use AI to write this article (how old fashioned of me) but it’s already hard for you to trust me on that. Pretty soon it will be impossible to know.

It’s not just thought leadership type blog posts either — pretty much every channel companies use to drive leads via static content is going to get arbitraged to death. The impending efficacy free fall of cold outreach needs no explanation for anyone with an email inbox. Traditional PR is on the sinking ship of traditional media.

The result is that companies will continue migrating to social networks as their core focus. But that means competing with creators and influencers on platforms built for creators and influencers, which is … hard. Many companies are fighting the short-form content fight valiantly, adopting the Duolingo strategy of hiring interns to post unhinged content that makes board members age several years at each meeting.

Long term though, I expect companies to decide that waging war with creators on even footing is not very smart. The unfair game they can play is pretty obvious, because it’s what companies are built to do — create products.

Expanding the pie

Domino’s “Points for Pies” campaign itself wasn’t anything new — big brands have built marketing apps for years. IKEA launched “IKEA place,” an AR app to help you design your home, in 2017. Snickers created an app with Spotify to suggest a playlist based on a user-uploaded “hungry face” as part of their famous “You’re not you when you’re hungry” campaign.

But what's interesting isn't that these apps exist — it's how much the cost of building them has changed. Points for Pies clearly cost millions to build. There's a whole TechCrunch article about how Domino’s used Nvidia chips to train a neural network to detect pizzas from non-pizzas, and even got users involved in the labeling. Amusingly, the article is actually sponsored by Nvidia. Hard to imagine them paying for that basic level of awareness just a few years ago.

Building an app like that today, just a few years later, could cost 100X less via the right team. The computer vision in that app would be a 1 penny API call instead of a TechCrunch-worthy engineering project, and the rest of the experience could be whipped up pretty quickly using modern AI-enabled tools.

Let me put it this way — if someone came to me asking how much it would cost for us to build that product for them at Robo, I’d probably estimate it at $5-10k. Obviously building at Domino’s scale and security would add some cost, but even that is an increasingly solved problem.

The world hasn’t recalibrated to that reality, but it’s here.

As Don Draper famously said, “success comes from standing out, not fitting in” and modern companies will increasingly use software to stand out. Especially because traditional content marketing is dying fast.

- Joey

P.S. We’ll have some case studies launching soon but in the meantime if you want examples of the kinds of marketing products we are working with companies to build, just reach out and I’m happy to share.

Basically, the insight here is the same as what Amazon found with the PR FAQ format, which is that the most effective way to give people the answers they might have on a topic is to ask and answer those questions yourself.

That works when you’re writing a big strategy doc or giving a formal presentation, but if you’re firing off an email to a contractor or sending a Slack message to a coworker it’s pretty odd and potentially even insulting to say “here are the questions you probably have, and the answers to them.”

So what I’ve started doing is creating custom GPTs for different personas and having a quick conversation with them about whatever it is I need to communicate. You can also do this by just prompting ChatGPT/Claude each time, but getting the prompts right takes some time and I find myself digging for old conversations to copy/paste the prompt. For example, I created a custom GPT to help me talk through an idea I have for an article like this one, which you’re welcome to use yourself. I use it to create conversations I send to people I work with on writing.

This is also a big improvement on the typical voice-note-as-comms strategy many busy people employ. Instead of sending an 8 minute voice note to my cofounder when I’m on the go, I can just speak into a GPT and then send him the transcript which he can skim in a minute or two. I often ask the GPT to summarize the conversation at the end as well if it was especially rambling. If you want to get really fancy, you can pipe the whole conversation through another prompt and turn it into whatever you want, but again I think the conversation format is really effective as a form of communication on its own.

In terms of creating your GPT Rolodex, it evolves pretty naturally over time but generally there are two types of agents to create to get things started:

• Third party subject-matter experts you “work with” frequently. For example you can chat with a sample “tough customer” agent prompted to ask the hard questions about your product idea and then send the convo to your cofounder.

• FAQ oriented agents where the agent is the same persona as the person you are sending the chat to. For example the “content advisor” agent you chat with before sending the convo to the person who is ghostwriting for you.

Custom GPTs work great and are easy to set up, but if you want to get fancier and design more complex agents using other models it’s really easy to do on Vectorshift. I have a few Claude-based agents in my rolodex there, and the rest as Custom GPTs.

Hope something in here is helpful! If you have ideas for something I should write about next, send me a conversation :)

The drive to build moonshots comes from a genuine need for companies to transform their existing business, and because it’s fucking cool when it works. Skunkwork stories are the backbone of Silicon Valley lore — Bezos and the Kindle, Jobs and the iPhone, the endless breakthroughs of Bell Labs and Xerox PARC.

Many of the moonshot success stories come from companies that had tons of extra cash: Amazon, Bell Labs, Apple, etc. In fact, when people talk about building skunkworks projects or “the Bell Labs of the future” the assumption is often that the key is to allocate a big sack of cash and make sure it is insulated from the rest of the company. 

But what happens when the cost of developing a new product goes down by 100X or more thanks to AI? I think we’ll see a different scarce ingredient to launching moonshots — not cash cows, but distribution cows. And with any luck that means leaner startups will be able to follow through on their yearly goals of doing something big. 

One pizza slice teams

Amazon famously coined the “two pizza team” as a heuristic for the max size of a team shipping a new product — only the number of people that can be fed by two pizzas. Great teams definitely make the best products, but I think we’ll see a lot more “one pizza slice” teams shipping skunkwork projects at early stages. The mythical man-month is real and getting real-er.

To state the obvious, the speed at which AI is bringing down the cost of product development is truly nuts, especially in software. If you build your workflow end to end with a goal of using AI to move faster, the speed increase is already 10X or more. I know because we are literally doing this all day every day now, shipping complex products for people in a few weeks. It’s wild. Software is the tip of that spear, but speed improvements will tell the story of production in every industry over the coming years.

You would be right to point out that launching a new product requires a lot more than just building the product (it takes marketing, sales, etc) but it’s worth noting that historically those have not been the most expensive part. And on top of that, AI speed increases are coming for the work of building marketing content, creating contracts, and the other things you need for a launch.

So it may not actually be one person teams yet, but that’s where we are headed. As the world of indiehackers will show you, you can ship incredibly fast if you are fully autonomous and have a baked-in audience to shorten the iteration cycles. And the quality and complexity of products you can ship on your own or in a very small team is going up at a mind-boggling rate. I’m sure the amount of money successful AI-augmented indiehackers are being offered from companies to join full time is getting crazy — expect that to accelerate quickly.

There’s no AI for distribution

So if it doesn’t require a ton of money or employees to launch a successful moonshot, what does it require? Outside of a good idea and a talented person or two to ship it, the obvious answer is distribution.

Distribution has always been the silent killer of most new products, and that only gets worse as it gets cheaper to build. Things are about to get ridiculously crowded, and any solution that automates distribution will be quickly commoditized. So the Bell Labs of the future won’t need a cash cow business to spin out moonshots, they’ll need distribution cows.

The real unlock is that unlike cash, distribution doesn’t get “spent” when you use it to launch a cool new product. In fact it’s the opposite, which is an area where creators are actually driving the trend. The pace of new product launches for a creator like Mr Beast would put a lot of Silicon Valley startups to shame, and it’s because he is thinking from a distribution first standpoint — more products that drive distribution, more distribution, more ability to launch products.

So not only will companies with distribution be able to launch moonshots, they might be left behind if they don’t create new products to preserve and extend their distribution advantage.

When cash is counterproductive

When startup nerds talk about distribution advantages, they often point to Google or Microsoft and their huge existing customer/user bases and giant sacks of cash. But while there are some examples of those companies incubating moonshots, it’s hard to argue that they have been super successful in incubating new products relative to the money and effort they’ve put into it. What they’ve been better at is acquiring/fast following new products, which is a separate beast. 

It’s well documented that companies like Google with a money printer struggle to create the same urgency and culture around shipping new products to market. Pressure creates diamonds, and one of the things I’m most excited about is that startups that are strapped for cash should actually be better moonshot factories provided they don’t have to keep killing the moonshot projects when money gets tight. 

Building moonshot muscle

When I think back to the moonshots I helped plan, run, and often axe at previous startups, many of them still feel like innovative opportunities that just never made it to market at all. And maybe if we had the tools available today (or a great roboagency :) some of them would have found the light of day. That’s incredibly exciting to think about.

Of course, this isn’t just a tooling problem. The ability to ship products 100X cheaper will be a breaking change for many industries, and require a whole new cultural approach to what it means to be lean, focused, and fast moving. 

My suggestion to founders and companies is the same advice I’m giving myself. Just work backwards from the ability to ship new products 10X faster than you have been before. However you do it is fine, but the forcing function is what matters at the moment if you want to make sure to get into the right position as this huge wave crashes.

Unfortunately I can’t cure the FOMO because it is increasingly true you are wasting time and money if you aren’t using AI. Over the past few months we’ve stumbled into what is clearly the future for product development — developers pushing AI to build products 10 to 100X cheaper. We’ve seen products that were originally quoted for $100k get done for 5. The FOMO, if anything, should be bigger. 

At the same time, the people who are able to build stuff with AI 100X faster are using fairly dark magic given the current state of tooling. For one thing, none of the “no code” tools really work unless you can code. Yes if you’re smart you can get Claude to write code for you and become an engineer without knowing how it works, but the people who are successfully doing that are mostly A) product people that already mostly understand how code works and B) people with a lot of time on their hands. 

If you want to spend the next 6 months learning how to build stuff incredibly fast with AI, the flow the real pros are using can be boiled down to:

1. Generate most of the code with Claude/OpenAI/boilerplates

2. Massage it into the scaffolding of a product

3. Use Cursor to build the last 20% alongside an inline AI

4. Use v0 to generate components and shortcut the front end development

5. Test and squash bugs 

There are other products involved, but that’s the stack for 95% of good cyborgs right now. Every one of those steps is a deep rabbit hole of its own, and the people moving the fastest have been experimenting and building their own workflows for a year or two at least. Not to mention the fact that it is genuinely hard to build these workflows inside of a company because make a lot of tradeoffs that only work for certain kinds of products (especially standalone MVPs). If you are a bank and one of your engineers uses AI to whip up a product where any line of code might be a security risk, you will fire that engineer and I’m not going to stop you.

So of course, it’s nuanced. You should feel the pressure to use AI to move faster, but you should also appreciate that going full cyborg is not as easy as it might seem on Twitter or TikTok.

The roboagency revolution

People talk about how AI will democratize access to development, and it will, but it is also going to increase the delta between the top 1% of builders and everyone else. Working closely with those AI-augmented engineers over the past few months has convinced us that AI FOMO is about to get a lot worse. In fact, we’re betting the company on it. 

I haven’t done this in the past, but I want to share the update we sent our investors last week because it summarizes our insights and conviction on this emerging class of “roboagencies.”

We’re officially rebranding it to Robo and we’ll share more details as we continue to come out of stealth. A lot of people tell me this feels too good to be true, and it’s certainly the case that not all projects are a good fit for a roboagency right now — as the tech matures, that slice will increase. But if your MVP is a good fit a general rule of thumb is that we can take a project that would normally take 1-3 months and cost ~$50k and ship it in 1-2 weeks for ~5k. If you’re curious whether it would work for an idea you have, I’m happy to take a look even if you’re just curious to have that data point.

-Joey

Being scrappy is great, but it’s just not true that AI products are inherently more expensive. In fact, AI products are often cheaper/easier to build for non-technical teams than “traditional” products. 

You’re probably not building OpenAI

The reason people assume AI is so expensive/complicated is because they confuse AI products with AI infrastructure. They read stories about the billions of dollars OpenAI and Meta are spending training their models and assume they would need to do the same to build their AI-enabled marketplace or assistant or whatever. 

If you aren’t a machine learning engineer, odds are high the initial version of your startup idea doesn’t need a machine learning engineer. One of the truly amazing things about AI right now is how much power you can get off the shelf from OpenAI and the sea of other tools that have been built over the past few years. Talented engineers that are using AI themselves can build pretty slick AI products in a single day.

AI products can actually be thinner

In many cases, AI can actually decrease the cost/complexity of an initial product. As one specific example, I talked to a marketplace founder that wants to recommend products to new users after onboarding. Their vision “for the future” was to collect some unstructured text from the user, extract a few categories from a list of potential options via AI, and then serve the user products from those categories. But “to keep costs down” they wanted to build a longer onboarding flow that asked structured questions and then fed those answers into a rules-based scoring system to output the best categories. 

If the AI version of your product is something you can easily do in ChatGPT or Claude, it’s also something you can easily do in code! Asking the LLM to look at a block of text and assign categories from a list of options is so much simpler than building a long onboarding wizard and an algorithmic scoring system, not to mention being closer to the vision for the product.

Productize yourself first

Another mistake I see people making in AI is assuming that an initial product needs to be a website or app where the users/customers are interacting with the AI themselves. The main driver of cost/complexity in that case comes from needing to build a polished user facing app.

I wrote about “wizard of Claude” products last week but a really good and lean go to market for a lot of AI products right now is:

1. Sell the product 

2. Deliver it manually to start

3. Use AI to automate/scale your own processes

4. Once you’ve validated the idea and found the right iteration, remove yourself from the process entirely and build the user facing version

On top of being cheaper, it’s just so much easier to iterate quickly in that setup. There are amazing low/no-code tools for steps 3 and 4 as well (I’ve become totally hooked on Stack over the past week especially given how easily you can compare models across entire flows side by side). 

Of course if you want some help or advice building an AI product, we’d love to help. The roboagencies we work with are great at building AI products quickly/cheaply.

Other notes:

• Very cool demo from Yuwen Lu of Misty which allows you to use inspirational screenshots as input for code revisions. Remixing is a much more natural form of creation for many (most) people.

• Great side by side review of Rollout (new AI landing page builder) vs v0. Pretty clear that landing pages are going to be the first real chip to fall in terms of what you can build end to end with AI.

• GitHub Spark has entered the ring of the AI-built product race. Interesting and unsurprising that they highlight revisions and human collaboration as a key differentiator.

If there’s something that caught your interest here you’d love me to cover in more detail, let me know!

My mind is blown on a daily basis. I’ve always been obsessed with using tech to help people bring their ideas to life, and honestly there are moments these days where I am close to hyperventilating with excitement.

So I’m creating a new section of Flying Penguins called “Approaching Zero” where I can chronicle some of those learnings and open the door to others. Potential topics include:

• Case studies of products built faster and cheaper with AI

• Reviews of game-changing AI tools for code, design, and LLM ops

• Conversations with centaur builders and agencies pushing the pace on AI augmentation

• Hopefully some cool guest posts from other founders/explorers

I expect I’ll refine the specifics of the format/frame as I go, so on that note…

Wizard of Claude

Many startups come to us wanting to build AI products, but here's the truth: you probably don't need to build as much as you think. At Grids, we're running a remarkably simple but effective operation - we gather client requirements, feed them through some agents built on Claude, and output scope documents for our talent network. No complex infrastructure, just clever use of existing tools.

While we'll be launching some self-serve products soon, this Wizard of Oz type approach could scale pretty far if we didn’t have engineers. You can also use something like Vectorshift or Voiceflow to design more complex agents or multi-agent systems. If you need, you can use APIs to put a simple front end on top. 

Some of the demand for overbuilt MVPs is typical laziness/inexperience. But a lot of it is lack of awareness on how deceptively simple these products can be under the hood given current tools. People are getting rinsed left and right by agencies for that reason, and I’ve actually advised a number of people to use off the shelf solutions even though it means leaving a good amount of money on the table for us.

Unsurprisingly, there is a huge market (and even some marketplaces) of AI advisors cropping up right now to help people figure out how to build an AI product without overbuilding or doing something dumb. Many of those advisors are low quality but if you know someone you trust who is deep in the weeds that’s ideal. Or of course we’d love to work with you and we do have some advisors in our network we can recommend as well.

Other notes:

• This is a great podcast with Amjad Massad, CEO of Replit, where he talks about the new Replit agent that basically builds fully functional websites via a chatbot. It’s clearly best in class in that category now.

• None of the AI design tools are incredible yet, but I like this approach from Polymet of commenting on multiple aspects of the design before “sending to revision” because it approximates how people work with real designers. If you’re building a simple single page website it can get you pretty far.

• I firmly believe the simplest way to build a magical AI product right now is to be early on the voice wave that’s about to crash. We’ve done some early A/B testing of the same agents with voice vs text and the difference in UX is pretty nuts.

If you made it here I’d love to know what you're interested in reading (or writing) on Approaching Zero!

What is a Centaur Agency?

A centaur agency scales through technology, not headcount. Think of DesignJoy, where a single designer (Brett Williams) makes millions yearly by “productizing himself.” Or Marble, which is revolutionizing legal services by automating immigration and family law processes, reaching hundreds of millions in revenue in just two years.

This isn't entirely new — companies like Deel and Pilot have been productizing services previously billed hourly (HR and accounting) for years. But AI changes everything because it can now handle creative work where quality is subjective. That means centaurs can now disrupt creative industries that have been the most resistant to automation: design, content marketing, legal services, and now, product development.

Centaurs in Product Development 

The more creative the work, the smaller and more specialized the centaurs. While Marble can scale quickly by automating routine legal work (immigration and family law), product development will be different – expect a sea of specialized and highly augmented teams, each compounding leverage on their unique expertise.

Why aren't these centaurs dominating the product outsourcing market yet? The simple answer is that AI coding tools are still maturing. Devin or Cursor might make good engineers 2X more productive, with rare cases hitting 10X or 100X. But watch the buzz around Replit Agent, where people are building in 15 minutes what used to take days. That’s where we're headed.

Won't this automation reduce the need for engineers? Paradoxically, no. History suggests that when resources get dramatically more efficient, demand increases. And as AI makes basic development accessible to everyone, standing out by hiring experts will be even more important.

Why MVPs will be first

Just like with Marble and DesignJoy, product development centaurs will focus first on work that is small, repeatable, and where speed and cost is paramount — areas where AI can provide a dramatically better product. I would expect the centaur wave to crash first for MVPs (minimum viable products) where teams are testing demand for a new product before investing more fully.

Of course, teams with their own engineers will use AI to build MVPs on their own, and the people at YC will rightly tell you that building something yourself is the best way to learn about it. That said, I think the pendulum will swing proportionally towards outsourcing for a two primary reasons:

• The economics finally make sense. When an MVP costs $5-10k, the relative cost of training your internal team or hiring new developers becomes much harder to justify.

• The market will expand dramatically. As development costs plummet, we'll see a surge of non-technical founders who previously couldn't get funding for an idea or afford to bootstrap.

Each of those points could be its own essay. Personally, I'm most excited about helping non-technical subject matter experts launch companies — I believe unlocking new areas of domain expertise is the lowest hanging fruit for accelerating innovation. And we're already seeing this transformation begin.

The Next Wave is Coming

The shift to centaur agencies in product development isn't just a prediction – we're seeing it firsthand. At Grids, we're working with agencies that are pushing the envelope with AI, and connecting them with startups ready to move fast.

Are you:

• Building a centaur agency? We’d love to connect you with founders we’re actively working with.

• Looking to build your MVP? We can help you get it built much more efficiently without sacrificing quality.

Reach out to me at joey@grids.co if so, or if you have thoughts on this topic I missed or got wrong!