Customer validation doesn't end with AI research: four columns to record
AI research only drafts whom to meet and what to ask; whether someone will buy shows in what they do when you meet.
Key points
- AI drafts hypotheses; meeting tests them
- Meet after shipping; users may differ
- Record in four columns, kept separate
- Read intent from actions, not praise
Contents
AI research takes you as far as hypotheses about whom to meet and what to ask
A&A perspective
Customer validation for a new business means meeting the buyers you have in mind to confirm that they really spend money and time on the problem. AI research covers the step before that: deciding whom to meet and what to ask.
A&A perspective
Point AI at public information and it will quickly list candidate companies, the words people in that industry use, and the situations where they seem to struggle. It can also help you put into words the questions to ask and the condition that would prove a hypothesis wrong, such as "if we see this reaction, the hypothesis is off".
A&A perspective
Three things, however, cannot be settled from documents. Will the company pay out of this year's budget? Who decides on the payment? Who actually does the work? None of them is clear until you meet the people or they try a prototype.
Hypothetical example
Say the new business helps people put together quotes. AI can suggest a scene like "sales reps at a building-materials trading company probably write quotes while hunting for similar past jobs". Whether those reps really write the quotes, and whether it takes them long, stays unknown until they show you a quote.
A&A perspective
Automate & Augment (A&A), which writes this article, treats the output of research not as a conclusion but as the order in which to meet people. Once you can write the names of the next three companies, with one assumption about each that might turn out wrong, stop researching there.
A&A perspective
When your manager or the board asks "are there really customers?", the amount of research is not an answer. The answer is a record of whom you met and what they did. In this article, what the other side actually did is called an "observed action".
Your idea of who uses it can change after you ship a prototype
A&A perspective
Assumptions about who will use a product can change not while you research but after you ship a prototype and people arrive. Two customer stories published by Anthropic, a US company that develops AI, describe how that happened.
From the sources
The first is Lovable, a service that turns a plain-language description of an app into that app. According to the case study, the founder had long heard from people who said they had something they wanted to build but could not code.
From the sources
Even so, what the founder built in early 2023, as a weekend experiment, was a tool to help developers work faster with AI. Released for anyone to use, the experiment spread. But the people who came to use it were not the developers he had built it for, the case study says.
From the sources
The second is Lex, a service that uses AI to help people who write. Its founder says the tools writers and editors use are not nearly as good as those of programmers and designers. The team first worked on fixing the interface, such as awkward track changes.
From the sources
The team then realised that generative AI could improve the writing process itself, the case study says. Lex is reported to have had 25,000 new sign-ups within 24 hours of launching its AI features. The case study does not say what the figure covers or how it was counted.
A&A perspective
A&A reads the Lovable story as the assumption about who the user is changing after a prototype shipped and people arrived. The requests had been coming in for a long time, yet the first thing shipped was for developers. What you hear and who actually turns up for what you ship are different pieces of information.
A&A perspective
At Lex, what changed was the assumption about what to fix. A&A reads it as the target moving, while the team worked on it, from the screen controls to the writing process itself.
From the sources
Paul Graham, who has invested in and advised newly founded companies, wrote in July 2013 that the most common time-consuming thing founders have to do at the start is to recruit users by hand. You cannot wait for users to come to you, he also wrote; you have to go out and get them.
A&A perspective
AI research is a tool for narrowing down where to go and cutting preparation time. It does not replace going.
Split hypothesis, question, action and offer change into four columns
A&A perspective
Write up each meeting in four columns: hypothesis, question, observed action and offer change. An observed action is what the other side actually did. Keeping the columns apart lets you trace later why you changed the offer. Mix them in one column and within days you can no longer tell what the buyer said from what you made of it.
A&A perspective
In the hypothesis column, write who struggles, in which situation, and how they cope today. In the question column, note which part of the hypothesis the question tests, and keep the buyer's answer there too. In the observed-action column, write what the buyer did, not what they said. In the offer-change column, write which of target, scope or price you changed, and how.
Hypothetical example
Say an in-house new-business team wants to sell help with writing quotes to a building-materials trading company with a dozen or so employees. This is a hypothetical example, not a real company. The hypothesis: "Sales reps spend more than an hour on every quote, searching for similar past jobs."
Hypothetical example
The question: "Could you show us the last three quotes you wrote, in the order you built them?" The observation: of the three, two were written by an office administrator rather than a sales rep, and on one the owner rewrote the unit prices at the end.
Hypothetical example
Three offer changes follow. Move the target from sales reps to the administrator. Take price-setting out of what gets automated. Keep a column where the owner checks the quote last. This single meeting showed two things at once: the person who uses the tool is not the person who decides, and one step is better left unautomated.
A&A perspective
The four-column record also works, as it is, in a report to your manager or the board. Instead of "we researched and concluded this", you can explain line by line: "this buyer did this, so we changed the offer like this".
A&A perspective
If all three companies you met matched the hypothesis, question the form of your questions before deciding the target was right. A question that tests something is one where the next step changes whether the answer is yes or no.
Read buying intent from actions where the buyer spent their own time or authority
A&A perspective
What comes closest to buying intent is an action in which the buyer spent their own time or authority. A remark like "that sounds handy" tells you nothing about where you rank in their budget. Keep remarks as answers in the question column, not in the observed-action column.
A&A perspective
A&A recommends writing actions like these in the observed-action column. They showed you internal documents. They booked the next meeting on the spot. They passed the conversation to someone in another department. They already pay for a substitute today. They named, on their own, what should be left out of scope.
A&A perspective
Conversely, hearing "I will send you the documents" and then receiving nothing is an equally useful observation. Not acting also shows the buyer's priorities.
From the sources
Some teams keep observing after launch. Each time a new version of the AI it uses comes out, Lovable measures how often the system produces an app that is broken or is not what the user asked for. According to the case study, it has run this measurement from the start, because many of its users cannot read the code themselves.
A&A perspective
Counting how often the result differed from what was asked picks up dissatisfaction that customers do not voice. In a new business too, setting up a column for this count before the first customer is decided makes offers to the second and later customers faster.
A&A perspective
This reading fits some situations less well. Both companies build services anyone can use, so once they launched they could see many users at once. A new business selling to companies can only meet them one at a time. In that case, count not the number of cases but how many kinds of departure from the hypothesis appeared.
A&A perspective
As an A&A rule of thumb, if the same kind of departure appears twice across three companies, that is a reason to review the target.
A&A perspective
The evidence has limits too. The two stories are customer stories published by Anthropic, so they are the companies' own accounts. They are not independent verification, and the two companies' growth cannot be treated as the effect of meeting customers. The four-column record and this rule of thumb are also A&A's provisional tools, not methods that have been measured and confirmed.
Summary
A&A perspective
Use AI research to build hypotheses about whom to meet and what to ask.
A&A perspective
Your idea of who uses the product can change after you ship a prototype, so ship one and meet people.
A&A perspective
Record each meeting in four columns: hypothesis, question, observed action and offer change.
A&A perspective
Read buying intent from what the buyer did, such as showing you documents or booking a date.
Use AI research to build hypotheses about whom to meet and what to ask. When your manager or the board asks whether there really are customers, the answer is not how much you researched but a record, split into four columns, of whom you met and what they did.
FAQ
Can I use AI research alone to narrow down candidates before meeting anyone?
You can use it to rule candidates out. In industries with few buyers, where public directories show the whole field, ruling some out before meeting is reasonable. But whether those who remain will pay is not something research can settle.
How many companies should I meet before I can report to my manager?
No source gives a set number. A&A suggests stopping research and going to meet people once you can write the names of the next three companies and, for each, an assumption that could prove wrong. In the report, rows with all four columns filled say more than the number of companies.
Don't I need to record what the buyer says?
You do record it, but as the answer in the question column, not in the observed-action column. The observed-action column holds only what the buyer did, such as showing documents or booking a date.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Lovable helps anyone create software 20x faster with Claude
Anthropic · n.d. (no date shown on page)
Accessed 2026-10-22 - Lex streamlines the writing process with Claude
Anthropic · n.d. (no date shown on page)
Accessed 2026-10-22 - Do Things that Don't Scale
Paul Graham · 2013-07
Accessed 2026-10-22