I have met a lot of startup founders. Nowhere in the world have I seen a younger professional community as sharp, as ambitious, and as energized as the one in Addis. This article is written for that community. Sharp people can handle hard truth, and the muscle of self-reflection is the one muscle a founder cannot outsource. With love and affection, I offer my short comments on seven hard truths about AI startups.
1. Economics is hard.
Most AI founders are happy to add multiple cognified* options to their product without thinking carefully about what each of those options costs to run. They show ARR projections without per-query cost, and they quote model prices that were true six months ago while assuming those prices will keep falling at the same rate. This is the first place where AI startups die. Every token your product generates costs you money, and every API call to a foundation model has a margin profile that decides whether your hundredth customer makes the company healthier or pulls it closer to insolvency. The pleasant fiction that you will figure out monetisation later was always a stretch in software, and in AI it has become a fast route to the graveyard, because the cost of goods sold is no longer a rounding error you can defer until the Series A. Your demo runs on the founder's credit card, your tenth customer runs on a free credit grant from a cloud provider, and your hundredth customer runs on infrastructure that has to clear gross margin every single month. Sometimes model prices fall faster than usage rises, and the unit economics improve on their own; more often they do not, and the founder discovers this only after the burn rate has already committed her to a path that no amount of growth can fix. Do the unit economics on day one, before the deck, before the seed round, before the celebratory dinner.
2. Moat (competitive advantage) is hard.
Most AI founders, when asked what their competitive advantage is, answer with a feature list. They describe what their product does and assume the description is the answer, but a list of capabilities is not an advantage; it is a specification, and any specification can be copied by anyone with access to the same underlying model. Most AI startups today are wrappers, which means a clever prompt, a thin interface, a foundation model doing the actual work, and a vertical label on top. This is acceptable as a starting point, but it is not yet a business, because the model you depend on is rented from a vendor who can rent the same model to your competitor tomorrow at the same price and on the same terms. Real competitive advantages in AI come from places software founders are not used to looking: data nobody else can collect, distribution into customer workflows that take years to displace, regulatory positions that took someone a decade to negotiate, and domain expertise so deep that the model needs you as a human partner to interpret what it produces. If your honest answer to "what is your competitive advantage" is "we got there first," then you do not have one, and the calendar will eventually prove this to your investors on your behalf.
3. The bulldozer is heavy.
Many AI founders are building their startups on a current weakness of the leading models. They notice that GPT cannot reliably do long-context reasoning, or that Claude struggles with a particular kind of tool use, or that no model handles citations without inventing sources, and they build a clever workaround on top of that gap. They raise a seed round on the strength of the workaround, and they ship a product whose entire reason for existing is that the underlying models are not yet good enough to do the job natively. Six months later, the next model release closes the gap, and the workaround that justified the company stops being a product and becomes a quaint historical artefact. Sam Altman has a word for what happens to those startups: they get bulldozed. The bulldozer is the lab itself, shipping a new version that absorbs the workaround into the model and eliminates the reason customers were paying you in the first place. The defence against the bulldozer is not to run faster, because you cannot outrun a foundation model lab on capability; the labs have more researchers, more compute, and more data than any startup will ever assemble. The defence is to build something the labs will not absorb even when they technically can, which means building a company whose value lives outside the model itself: in proprietary data the labs cannot legally or practically collect, in customer relationships that took years to earn, in regulated workflows where compliance matters more than capability, and in deep domain context where the model needs a knowledgeable human partner to do anything useful. If the only reason your product exists is that the current model is not yet good enough, then you are standing in front of a bulldozer, and the bulldozer ships on a regular release schedule.
4. Context is hard.
I often meet Ethiopian AI founders and hear a pitch that would not be out of place in Tel Aviv, London, or San Francisco. Somewhere around the third slide I realise that the founder has built a product for a market that does not actually exist in the country she lives in. These are smart, well-prepared people, but they have skipped the work of asking what their local market can afford, what infrastructure their customers have access to, and what problems are urgent enough that someone will pay real money to solve them this year. The Ethiopian customer is not the San Francisco customer with a smaller budget; she is a different customer, with different priorities, different constraints, and a fundamentally different relationship to software. She may be paying for connectivity by the megabyte, running her business on a phone rather than a laptop, working in Amharic rather than English, and choosing every month between tools that compete not with each other but with rent and inventory. A pitch designed for an enterprise buyer in California, dropped into this context unmodified, will not just convert poorly; it will fail to register as relevant at all. The founders I admire in Addis are the ones who have done the unromantic work of mapping their market honestly: who can pay, how much, for what, and on what payment rails. They build products that fit the answer to those questions rather than products that fit a slide template they saw at a Y Combinator demo day.
5. Products are hard.
Most founders ship a model with a UI on top and call it a product, confusing a working demo for a working product, when the two are not the same thing and never have been. The AI product map looks like Bosch's garden, full of bodies and beasts and fountains and small explanatory creatures, with no clear sense of what anyone is doing or why anyone is supposed to come back tomorrow. Most AI products are feature collections looking for a job, and they betray themselves in their usage data: they demo well, attract a wave of curious signups, and then lose ninety percent of those users in the first week, because nothing in the product solves a problem the user actually has on Tuesday at three in the afternoon when she has fifteen minutes between meetings. A product, properly understood, is something a user returns to without being reminded, and that simple test is harder to pass than most founders realise. Passing it requires a clearly defined job to be done, a workflow the product fits inside without forcing the user to redesign her day, and a reason to open the product instead of doing nothing or doing the same task in the tool the user already has open.
6. Smart communication is hard.
AI startups tend to style their communication to own preferences. They post on LinkedIn for the applause of their peers, refine the pitch with their co-founder over dinner, present at meetups full of people building the same kind of company, and then mistake the resulting warmth for market validation. It is not market validation; it is an echo, and it is the cheapest kind of feedback a founder can collect, because the people giving it have every reason to be encouraging and almost no information about whether the actual buyer cares. Founders fall in love with their own pitch the way Narcissus fell in love with his reflection in the pond, and they spend so long polishing the language that they forget the language was supposed to do work in someone else's mind, not in theirs. The actual buyer is not on LinkedIn applauding your latest carousel post; the actual buyer is a procurement officer at a mid-sized insurance company who has never heard of your category, does not subscribe to your newsletter, and has a stack of invoices to process before lunch. Communication discipline means writing for that person, in her vocabulary, about her problem, on her schedule. It means refusing the inflated language of "transformative" and "revolutionary" and "AI-powered" in favour of describing what the product does using the words the buyer would actually use to describe the difficulty she is trying to make go away.
7. "AI injera" does not sell.
Sometimes it feels as though founders in Addis are simply slapping the letters "A" and "I" onto products that do not require any AI inside, hoping that an investor or a grant committee will be too busy or too charmed to check. This strategy is bad, and it will bite you sooner than you expect. The funders are getting better at checking, the questions in due diligence are getting more technical year by year, and the local startup community is small enough that a reputation for AI theatre travels in days rather than months, sometimes faster than the founder can travel between two coffee meetings on Bole Road. Build something with real AI inside, where the AI is genuinely doing work the product could not do without it, or build something honest that does not pretend to be what it is not. Both options are respectable, both options are fundable on their own terms, and both options leave you with a story you can tell without flinching three years from now.
Talk to the chatbot.
Gary Vaynerchuk used to close his talks with a line: everything I just told you is available on a very exclusive platform, and I will spell its name for you, G-O-O-G-L-E. The joke worked because it was true; the wealth of tactical advice, best practices, and tips and tricks that conference speakers package and resell has been freely available online for years to anyone willing to type. The same move applies one generation later, one layer up. Everything in this article, every framework about competitive advantage and margins and context, is available to you inside any AI chatbot you can open right now.
Two prompts to start with.
First: "This is a list of hard truths about AI startups. Here is a description of my idea. Tell me how my idea relates to each truth, honestly, without flattery." Second: "This is a list of communication mistakes AI founders make. Here is a draft of my post. Tell me which mistakes it contains." Paste the seven truths above, paste your own idea or your own draft underneath, and read what the chatbot returns before the next investor call rather than after.
* Cognification is a term introduced by Kevin Kelly in his 2016 book The Inevitable to describe the ongoing technological shift of making objects and processes smarter by embedding artificial intelligence into them. I am using the term here to mean the process of adding LLMs to your application.