The Embedded AI Thesis
The biggest AI wins will be invisible
In 1882, Thomas Edison opened the Pearl Street Station in lower Manhattan and began selling electricity to fifty-nine customers. Within a decade, entrepreneurs across America were building "electric products" - electric pens, electric belts marketed as cure-alls, electric hair brushes, electric cigar lighters. Most of these products failed. They were novelties, solutions looking for problems, built by people who saw a powerful new technology and assumed the way to profit from it was to make the technology itself the product.
The actual electricity revolution looked nothing like this. It looked like a factory in Cincinnati replacing its steam engine with an electric motor and discovering that output increased by thirty percent - not because electricity was inherently better than steam, but because electric motors could be distributed throughout the factory floor, each machine powered independently, eliminating the elaborate system of belts and shafts that had constrained factory layout for a century. The factory owner did not buy "an electricity product." He bought a factory that worked better. The electricity was invisible.
I keep watching the AI industry and seeing electric pens everywhere.
The Demo Problem
Every AI startup has a remarkable demo. I have sat through hundreds of them. A chatbot that answers complex questions with uncanny fluency. A vision system that identifies defects a human would miss. A coding assistant that generates working applications from a paragraph of description. These demos are genuinely impressive, and they are the reason the founders got funded, got press, got their first wave of signups.
They are also measuring the wrong thing entirely. A demo measures what a technology can do in a controlled setting, with a prepared input, in front of an attentive audience. It does not measure what a person will actually change their behaviour to use, unprompted, on a Tuesday afternoon when they are behind on three deadlines and have seventeen browser tabs open. The distance between those two measurements - impressive demo and daily-use product - is where most AI startups go to die.
The pattern is remarkably consistent. Strong launch. Enthusiastic early adoption. A usage curve that peaks in week two and declines steadily thereafter. The founder looks at the retention numbers and assumes the product needs more features, or better onboarding, or a redesigned interface. Sometimes those things help. Usually the problem is more fundamental: the product is asking people to do something new, and doing something new has a cost that no amount of interface polish can eliminate.
The Behaviour Change Tax
The best technology disappears. The user does not admire it. The user does not even notice it. They just notice that things work better than they used to.
Every standalone AI product carries a hidden cost that never appears on a pricing page or a financial model: the behaviour change tax. This is the accumulated friction of learning a new interface, developing a new habit, integrating a new step into an existing workflow, and - critically - trusting a new system enough to rely on it for work that matters.
The behaviour change tax is the single most underestimated force in software. It has nothing to do with the quality of the technology. People are not resistant to AI because they are luddites or because they do not understand it. They are resistant because they are busy. They have workflows that function well enough, and "well enough" is an extraordinarily powerful force in human behaviour. A tool that is twice as good but requires you to change how you work will lose, repeatedly, to a tool that is marginally better but requires you to change nothing.
Embedded AI pays zero behaviour change tax. The user's workflow stays identical. The interface stays identical. The habits stay identical. The tool simply works better than it did yesterday. The invoicing software catches an error it would have missed last month. The CRM populates a field that used to require manual entry. The scheduling system accounts for a constraint it used to ignore. The user may not even notice the improvement consciously. They just notice, over time, that things go wrong less often.
This is not a minor advantage. It is the entire game. The history of technology adoption is, at its core, the history of behaviour change tax: technologies that demand new behaviour are adopted slowly and painfully; technologies that embed themselves into existing behaviour are adopted fast and irreversibly.
The Infrastructure Phase
What I am describing is not a novel observation. It is a pattern that has repeated with every general-purpose technology of the past hundred and fifty years, and it always follows the same sequence.
Phase one: the technology emerges and entrepreneurs build standalone products around it. Electric pens. "Internet companies." Cloud-native applications. The products are designed to showcase the technology. The technology is the value proposition. Some of these products succeed, but most do not, because the market for "a product built on exciting new technology" is small and fickle. The market for "a product that solves my problem" is large and durable, and those are different markets.
Phase two: the technology becomes infrastructure. Electricity wired into every building. The internet beneath every application. Cloud computing under every service. The technology disappears from the user's awareness entirely. It is no longer the product. It is the invisible layer that makes products better. This is where the overwhelming majority of the economic value is created, and it is created by companies that nobody writes breathless profiles about.
AI is deep in phase one. The press covers AI products - the chatbots, the image generators, the coding assistants. These are the electric pens of our era: impressive, visible, and (in most cases) addressing a market far smaller than the technology's actual potential. The real value of AI will be created in phase two, when AI capabilities are embedded so deeply into existing software that users stop thinking about AI entirely. When your accounting software is simply better at catching errors, and you do not know or care whether it uses a large language model or a rules engine or a trained hamster to do it.
What Embedded AI Actually Looks Like
The numbers: Forty hours of manual re-routing per week became six. Adoption time was zero - because there was nothing new to adopt. The team that built this will never be on the cover of a magazine. They built something that is, by design, invisible.
Let me be specific, because the embedded path is easy to dismiss as unambitious when described in the abstract.
A logistics company I advised was spending roughly forty hours per week on manual shipment re-routing - staff checking weather reports, port congestion data, customs delay notifications, and adjusting routes accordingly. They evaluated three AI startups that offered AI-powered logistics optimisation as a standalone platform. Each required migrating to a new system, retraining the operations team, and rebuilding integrations with their existing tools. The behaviour change tax was enormous. They chose none of them.
What they did instead was work with a small team that built an AI layer directly into their existing routing software. The interface did not change. The workflow did not change. The system simply began suggesting better routes, with brief explanations, inside the tool the operations team already used every day. Adoption was instant because there was nothing to adopt. The forty hours per week dropped to six.
The pattern scales across industries. The accounting firm whose audit software now flags anomalous transactions before a human reviewer sees them. The hospital whose scheduling system learned to account for procedure duration variance without anyone configuring it to. The law firm whose document review tool quietly improved its relevance ranking until associates stopped complaining about search quality and nobody could pinpoint when it changed. None of these are AI products. They are products that use AI, and the distinction determines everything about how they are adopted, how they are retained, and what they are worth.
The Distribution Advantage
The hardest problem in startups is distribution. Not the hardest technical problem - the hardest problem, full stop. Building something valuable is necessary but not sufficient. You also have to get it in front of people who will use it, repeatedly, and pay for it. Most startups that fail have a product that works. They fail because they cannot solve distribution.
Standalone AI products have to solve distribution from zero. They are a new product in a new category, competing for attention against every other new product in every other new category, asking users to develop a new habit when existing habits are serving them adequately. This is why customer acquisition costs for standalone AI products are so high and why the retention curves are so consistently discouraging.
Embedded AI has a distribution cheat code: it rides existing channels. If you can get your AI capability into a platform that already has two million users, you have bypassed the problem that kills most startups. You do not need users to find you. You do not need them to sign up for anything. You do not need them to change a single thing about their day. You are already there, inside the tool they open every morning. Your challenge is now purely technical - make the product measurably better - and technical challenges, unlike distribution challenges, yield to engineering.
The infrastructure layer will be where this plays out most dramatically. The companies building APIs, middleware, and orchestration tools that let existing software add AI capabilities without rebuilding from scratch - these will produce some of the largest outcomes in AI. They are selling picks and shovels, the old saying goes, but that metaphor undersells what they are actually doing. They are selling invisibility. They are selling the ability for any software company to embed intelligence without becoming an AI company, and that is worth enormously more than any single AI product.
The Funding Gap
There is an obvious question: if embedded AI is so valuable, why is most of the money going elsewhere?
The answer is structural. Venture capital is optimised for stories, and standalone AI products tell better stories. "We built an AI that does X" fits on a slide. It demos well. It photographs well. It makes sense to a limited partner who reads about AI in the Financial Times and wants to know what the fund is doing about it. "We made this enterprise invoicing platform fifteen percent more accurate at anomaly detection" does not have the same ring. It is not going to be the lead example in the fund's annual letter.
I am not criticising venture capitalists, who are responding rationally to their own incentive structures. I am observing a market gap. The embedded path is underfunded relative to its economic value, and underfunding creates opportunity. There is less competition, more room to build, and often more willingness from customers to pay, because embedded AI improvements are directly measurable in operational terms the buyer already cares about. You do not have to convince a CFO that "AI" is worth investing in. You have to show her that the error rate in invoice processing dropped by fifteen percent last quarter. That is a conversation she already knows how to have.
The founders willing to do this work - the unglamorous, deeply technical work of understanding someone else's workflow well enough to improve it invisibly - are building on a path with better unit economics, lower churn, and a customer base that does not need to believe in AI to benefit from it. The companies that electrified American manufacturing did not put "electric" in their names. They were manufacturing firms that happened to adopt a better power source. The most valuable AI companies of the next decade will follow the same pattern. They will not have "AI" in their pitch decks, or if they do, it will be on page twelve, after the slides about the customer, the problem, and the margin improvement. The technology will be real and important and completely beside the point.