An opinion from someone building AI where the training data runs out.
Every AI product tutorial makes the same silent assumption: that your problem domain is already measured. Want to build a price estimator? There's a Kaggle dataset. A recommendation engine? Scrape the reviews. A valuation model? Kelley Blue Book publishes the ground truth and your model just has to approximate it.
Now try building the same product in Lagos.
I build AutoValue, an AI pricing platform for used cars in Nigeria, and I can tell you exactly what the data landscape looks like: there is no Blue Book. There is no clean listings API. The largest marketplaces are full of asking prices that everyone understands to be opening bids, posted by sellers who expect to be haggled down 20%. Two identical cars sit ₦8 million apart on the same page. The "dataset" for my domain is a negotiation culture, a WhatsApp group, and a dealer's gut feeling.
The standard takeaway is that markets like this are too hard for AI products. I think that conclusion is exactly backwards, and I want to make the case for why data-poor markets are the most interesting place to build right now.
Where there's no dataset, there's no incumbent
In the US, a startup doing AI car valuation is competing against Kelley Blue Book, Carvana's pricing engine, CarGurus' deal ratings, and forty years of accumulated market data. The AI is a marginal improvement on an already-solved problem. The moat belongs to whoever collected the data decades ago.
In Nigeria, the same product is not an improvement on the incumbent. It creates the category. Whoever produces the first trusted price reference doesn't compete with the Blue Book. They become it. And this generalizes far beyond cars: think rents in Nairobi, used electronics in Dhaka, farm equipment in rural Brazil. Enormous, active markets that trade daily without a single reliable price signal. Every one of them is waiting for its reference source, and the reference source is now buildable by a small team with an LLM, a search API, and domain stubbornness.
LLMs changed what "no dataset" means
Here's the technical shift that makes this possible. The classical approach to a pricing model is supervised learning: collect a hundred thousand labeled transactions, train, deploy. No transactions, no model. Data-poor markets were locked out by definition.
LLMs flipped the order of operations. My pricing engine doesn't learn from a historical dataset; it synthesizes a price at request time from live search results, structural knowledge (depreciation curves, trim hierarchies, mileage effects), and hard domain guardrails. The model contributes judgment, the search contributes freshness, and the database of resolved prices grows as a byproduct of usage. The dataset isn't a prerequisite anymore. It's the exhaust.
That inversion is the entire unlock. You can now bootstrap a data business in a market that has no data, because the first version of the product doesn't need the data the product will eventually accumulate.
The catch: the model thinks your market doesn't exist
I won't romanticize this. Building AI for a market the training data forgot means the model's priors actively work against you. Mine believed a dollar was still ₦450 and priced every car in the country at a third of its value, because its memorized world predated a currency collapse. Data that does surface in search is polluted with scam listings and aspirational prices. Payment rails that every Stripe tutorial assumes simply don't apply.
So you build differently. Retrieval isn't an optimization here; it's load-bearing. Domain guardrails aren't nice-to-haves; they're the only thing standing between you and confidently wrong output. Every architectural decision that's optional in a data-rich market is mandatory in a data-poor one. The upside is that these constraints force you to build the robust version of the system on day one, the version that data-rich builders only discover they need after their first embarrassing failure.
Who should be building here
If you're a developer in Lagos, Nairobi, Karachi, or Jakarta, you have an unfair advantage that no one in San Francisco can replicate: you can tell when the model is wrong. You know the street price, the negotiation dance, the scam patterns, the real exchange rate. Ground truth lives in your head, and in markets without datasets, ground truth in someone's head is the scarcest resource in the entire stack.
The standard advice says to copy proven Western products for your local market. I'd argue the opposite is now more valuable: build the measurement layer your market never had. Not another marketplace on top of chaos, but the price signal, the quality score, the trust index that makes the chaos legible. Marketplaces come and go. Reference data compounds.
The next billion users aren't waiting for translated versions of solved products. They're living inside unsolved measurement problems, and for the first time, the tools to solve them don't require a data moat you don't have. They require judgment about a market you already understand better than the training data does.
That's not a handicap. That's the entire edge.
I build AutoValue in the open at autovalue.tech, and I write about the engineering behind it, including the pricing engine and the currency bug that nearly sank it. If you're building for a market the datasets forgot, I'd genuinely like to hear what you're measuring.