Choosing a Problem Worth a Decade

The problem you choose sets the ceiling on everything that follows. A great team working on a weak problem usually loses to an average team working on a strong one. This chapter covers how to pick a problem that hurts enough to pay for, that you are well placed to solve, and that will still matter once everyone's tools get better.

Start With What You Know

The most reliable source of good startup problems is your own experience. Years inside an industry show you the workarounds, the spreadsheets held together by habit and the tasks everyone hates but nobody fixes. You also learn who buys, how they buy and what they are afraid of.

This is usually called founder–market fit. Chris Dixon, who popularised the term, defined it as founders having "a deep understanding of the market they are entering," and argued it is the best predictor of whether a startup reaches product–market fit. In practice it means you can get meetings others cannot, you speak the customer's vocabulary, you can tell when feedback is nonsense, and you stay interested when the work turns tedious.

The data supports this. A study of US administrative records by Azoulay, Jones, Kim and Miranda, published in American Economic Review: Insights in 2020, found that prior experience in the specific industry "predicts much greater rates of entrepreneurial success," and that the mean founding age of the 1-in-1,000 fastest-growing new ventures was 45. Time spent in an industry counts in your favour.

Paul Graham makes a related point in "How to Get Startup Ideas". He argues that the verb you want for startup ideas is "notice" rather than "think up," that the best ideas grow out of founders' own experiences, and that you should "live in the future, then build what's missing." An idea that came out of your own work will usually beat one produced in a brainstorm.

If you lack deep experience in a field, you are not disqualified. You will have to earn the insight, though, usually by spending real time with the people who have the problem.

Painful, Frequent and Expensive

Plenty of real problems make poor businesses. The strongest ones tend to share three traits.

Trait	What it looks like	Why it matters
Painful	People complain about it unprompted and have tried to fix it	Pain creates urgency, and urgency shortens sales cycles
Frequent	It happens daily or weekly, not once a year	Frequent problems build habits and retention
Expensive	It costs money, time, revenue or risk that someone can measure	Measurable cost makes it easy to justify a price

A problem with all three is rare and valuable, and two out of three can work. One that is mildly annoying, occasional and costs nothing obvious is a feature at best, and you will struggle to sell it.

The workaround is a useful test. If people are already solving the problem badly, with spreadsheets, manual processes, contractors or stitched-together tools, they have shown they care. If nobody has bothered to work around it, the pain may be smaller than it sounds.

Wedges Before Platforms

Many founders describe their vision as a platform, the place where an entire industry will run its operations. You may end up there. You should almost never start there.

A wedge is a narrow product that does one job very well for one type of customer. It is small enough to build and sell quickly, and sharp enough that customers pick it over the alternatives. Once you own that job, you have the trust, data and distribution to expand into adjacent ones.

Starting with a platform causes predictable problems. It takes too long to build, it is hard to explain, and it puts you up against every incumbent at once. A wedge gives you a clear message, a clear buyer and a short path to revenue.

Try to name your wedge in one line: who it is for and the single job it does. If the sentence needs "and" more than once, you are probably still describing a platform.

Where AI Changes the Economics

The most interesting openings right now are in markets where AI changes what is economically possible, rather than adding a feature to what already exists. Four patterns are worth looking for.

• Work that was too expensive to do at all. Tasks that needed skilled people to read, write or analyse large volumes of material were skipped, or done only for the biggest clients. Now they can be done for everyone.
• Services that could not scale. Businesses that sold expert labour by the hour, such as bookkeeping, compliance review, research or support, can now be delivered as software with a human checking the output. Foundation Capital calls this shift "service as software": the vendor sells the outcome rather than a tool. It sizes the opportunity at about $4.6 trillion, counting salaries in functions such as sales, engineering and HR plus outsourced IT and business-process services.
• Markets too small for incumbents. A niche that was never worth a large company's effort may now be worth a small one's, because building costs less.
• Messy, unstructured inputs. Documents, calls, emails and images that resisted automation are now workable, so industries that run on paperwork deserve a close look.

In each case, ask the same question: what can a customer now get, at a price they will pay, that was not available before? If the honest answer is "the same thing, slightly faster," the opportunity is smaller than it looks.

Avoid the Thin Wrapper

The fastest way to build an AI product is to put a simple interface on a general-purpose model. It is also the fastest way to become irrelevant. Anything the model providers or the large software platforms can add as a feature, they eventually will, and they already have the users. Andreessen Horowitz flagged this early in the generative AI wave, noting that many AI apps were "relatively undifferentiated, since they rely on similar underlying AI models," and that their gross margins were often only 50 to 60% because of inference costs (2023). Jasper, an AI writing tool, learned this the hard way: after ChatGPT gave its users a cheap alternative almost overnight, it reportedly cut its internal valuation by 20% and its 2023 revenue forecast by at least 30%, and refocused on marketing teams at larger companies (2023).

Put three questions to any idea.

1. If the underlying models got much better tomorrow, would my product get better or become unnecessary? You want the first answer. A product that depends on models staying weak is on borrowed time, and models are improving and getting cheaper fast: Stanford's 2025 AI Index found the cost of GPT-3.5-level performance fell more than 280-fold in under two years.
2. What do I have that a general tool does not? Good answers include proprietary data, deep integration into a specific workflow, domain expertise built into how the product works, regulatory approvals, or relationships with a hard-to-reach buyer.
3. Could the incumbent in this market add this in a quarter? If so, you need a reason customers would still choose you, such as serving a segment the incumbent ignores or doing the whole job rather than one step of it.

Defensibility usually comes from owning the workflow rather than the model. A product that becomes the system where a customer's work actually happens, holding their history and connected to their other tools, is hard to rip out even when competitors appear.

Write a One-Paragraph Thesis

Before you build or validate anything, write your thesis down. A thesis is a short, falsifiable statement of the bet you are making, and writing it turns vague enthusiasm into something you can test.

A good thesis answers five questions in plain language:

1. Who has the problem, described specifically enough that you could find them.
2. What the problem is and what it costs them today.
3. What they do about it now.
4. What you will offer and why it is meaningfully better, often because of something that recently became possible.
5. Why you are the right person to build it.

Keep it to one paragraph. If it will not fit, you do not yet understand the problem well enough. At AW3 we ask founders for this paragraph before any build conversation, because it shows quickly whether the idea is a business or a feature.

Treat the thesis as a draft. The next chapter is about testing it against real customers, and you should expect to rewrite it several times.