From dot-com to dot-AI: How we can learn from the last tech transformation (and avoid making the same mistakes)

On the top of the dot-com increase, including “.com” to an organization’s title was sufficient to ship its inventory worth hovering — even when the enterprise had no actual prospects, income or path to profitability. In the present day, historical past is repeating itself. Swap “.com” for “AI,” and the story sounds eerily acquainted.

Firms are racing to sprinkle “AI” into their pitch decks, product descriptions and domains, hoping to experience the hype. As reported by Domain Name Stat, registrations for “.ai” domains surged about 77.1% year-over-year in 2024, pushed by startups and incumbents alike dashing to affiliate themselves with synthetic intelligence — whether or not they have a real AI benefit or not.

The late Nineteen Nineties made one factor clear: Utilizing breakthrough expertise isn’t sufficient. The businesses that survived the dot-com crash weren’t chasing hype — they have been fixing actual issues and scaling with goal.

AI is not any completely different. It is going to reshape industries, however the winners received’t be these slapping “AI” on a touchdown web page — they’ll be those reducing via the hype and specializing in what issues.

The primary steps? Begin small, discover your wedge and scale intentionally.

Begin small: Discover your wedge earlier than you scale

One of the pricey errors of the dot-com period was making an attempt to go massive too quickly — a lesson AI product builders at present can’t afford to disregard.

Take eBay, for instance. It started as a easy on-line public sale web site for collectibles — beginning with one thing as area of interest as Pez dispensers. Early customers beloved it as a result of it solved a really particular drawback: It related hobbyists who couldn’t discover one another offline. Solely after dominating that preliminary vertical did eBay increase into broader classes like electronics, trend and, finally, nearly something you should purchase at present.

Examine that to Webvan, one other dot-com period startup with a a lot completely different technique. Webvan aimed to revolutionize grocery buying with on-line ordering and speedy dwelling supply — abruptly, in a number of cities. It spent a whole lot of tens of millions of {dollars} constructing huge warehouses and complicated supply fleets earlier than it had sturdy buyer demand. When development didn’t materialize quick sufficient, the corporate collapsed below its personal weight.

The sample is evident: Begin with a pointy, particular consumer want. Concentrate on a slender wedge you’ll be able to dominate. Increase solely when you might have proof of sturdy demand.

For AI product builders, this implies resisting the urge to construct an “AI that does every part.” Take, for instance, a generative AI instrument for information evaluation. Are you focusing on product managers, designers or information scientists? Are you constructing for individuals who don’t know SQL, these with restricted expertise or seasoned analysts?

Every of these customers has very completely different wants, workflows and expectations. Beginning with a slender, well-defined cohort — like technical challenge managers (PMs) with restricted SQL expertise who want fast insights to information product choices — lets you deeply perceive your consumer, fine-tune the expertise and construct one thing actually indispensable. From there, you’ll be able to increase deliberately to adjoining personas or capabilities. Within the race to construct lasting gen AI merchandise, the winners received’t be those who attempt to serve everybody without delay — they’ll be those who begin small, and serve somebody extremely effectively.

Personal your information moat: Construct compounding defensibility early

Beginning small helps you discover product-market match. However when you acquire traction, your subsequent precedence is to construct defensibility — and on the planet of gen AI, which means proudly owning your information.

The businesses that survived the dot-com increase didn’t simply seize customers — they captured proprietary information. Amazon, for instance, didn’t cease at promoting books. They tracked purchases and product views to enhance suggestions, then used regional ordering information to optimize achievement. By analyzing shopping for patterns throughout cities and zip codes, they predicted demand, stocked warehouses smarter and streamlined transport routes — laying the inspiration for Prime’s two-day supply, a key benefit opponents couldn’t match. None of it could have been attainable with no information technique baked into the product from day one.

Google adopted an analogous path. Each question, click on and correction grew to become coaching information to enhance search outcomes — and later, advertisements. They didn’t simply construct a search engine; they constructed a real-time suggestions loop that continuously realized from customers, making a moat that made their outcomes and focusing on tougher to beat.

The lesson for gen AI product builders is evident: Lengthy-term benefit received’t come from merely getting access to a robust mannequin — it should come from constructing proprietary information loops that enhance their product over time.

In the present day, anybody with sufficient sources can fine-tune an open-source massive language mannequin (LLM) or pay to entry an API. What’s a lot tougher — and way more worthwhile — is gathering high-signal, real-world consumer interplay information that compounds over time.

For those who’re constructing a gen AI product, that you must ask essential questions early:

• What distinctive information will we seize as customers work together with us?
• How can we design suggestions loops that constantly refine the product?
• Is there domain-specific information we will gather (ethically and securely) that opponents received’t have?

Take Duolingo, for instance. With GPT-4, they’ve gone past basic personalization. Options like “Clarify My Reply” and AI role-play create richer consumer interactions — capturing not simply solutions, however how learners suppose and converse. Duolingo combines this information with their very own AI to refine the expertise, creating a bonus opponents can’t simply match.

Within the gen AI period, information needs to be your compounding benefit. Firms that design their merchandise to seize and study from proprietary information would be the ones that survive and lead.

Conclusion: It’s a marathon, not a dash

The dot-com period confirmed us that hype fades quick, however fundamentals endure. The gen AI increase is not any completely different. The businesses that thrive received’t be those chasing headlines — they’ll be those fixing actual issues, scaling with self-discipline and constructing actual moats.

The way forward for AI will belong to builders who perceive that it’s a marathon — and have the grit to run it.

Kailiang Fu is an AI product supervisor at Uber.