AI Product-Market Fit: How to Know If Your AI Product Is Working | UData Blog

Simon Willison, one of the sharpest observers of the AI industry, published a post this week arguing that both Anthropic and OpenAI have genuinely found product-market fit — not the polite, PR-friendly version of PMF, but the real thing, measured in revenue per employee and organic retention curves that look unlike anything the enterprise software industry has seen at this speed. His argument is worth reading if you haven't. But the more interesting question for CTOs and founders building AI-powered products is not whether the frontier model companies have found PMF. It is whether you have — and if not, what the gap between your current state and real AI PMF actually looks like.

The signals that distinguish genuine AI product-market fit from polite early adoption are different from the signals that apply to conventional SaaS. Retention curves behave differently. The buyer is often not the end user. The product can improve week over week in ways that legacy software cannot. And the competitive dynamics move faster than anything a five-year-old SaaS company was built to respond to. This article covers the specific signals that indicate an AI product has found PMF — and the red flags that indicate it has not, even when early metrics look encouraging.

Why AI PMF Looks Different From SaaS PMF

The standard SaaS PMF indicators — net revenue retention above 110%, declining churn as cohorts mature, word-of-mouth as a meaningful acquisition channel — apply to AI products, but they take longer to appear and can be masked by patterns that do not occur in conventional software. The masking effect matters because it means AI products can appear to have PMF (strong early trial activation, high initial engagement, enthusiastic pilot cohorts) while actually being far from it (rapid churn after the novelty period, reluctance to expand beyond the initial use case, difficulty converting pilots to paid expansions).

The reason for this masking is the novelty premium. AI products — particularly products that demonstrate a genuinely new capability, not just an AI wrapper on an existing workflow — generate curiosity-driven engagement that looks like PMF in early dashboards. Users try the product extensively because it is interesting and novel, not because it is solving a problem they were previously solving worse. When the novelty premium expires, typically somewhere between week four and week twelve of active usage, the retention curve reveals whether the product is retaining users because it is genuinely useful or because it was entertaining to explore.

The second masking effect is buyer-user misalignment. Enterprise AI pilots are often sponsored by innovation-minded CTOs or digital transformation leads who have budget and enthusiasm for AI exploration. The pilots are not funded because the product solves a demonstrated pain point in the workflow of the people who will actually use it — they are funded because the company wants to be seen as exploring AI. These pilots generate favorable engagement metrics during the exploration phase but convert to renewals at low rates because the user-level value was never compelling enough to justify the organizational change required to integrate the product into standard workflows.

The Signals That Indicate Real PMF

The indicators that distinguish genuine AI product-market fit from the novelty-premium or innovation-budget patterns:

Users degrade significantly when the product is unavailable. The most reliable PMF indicator for any software product is user behavior when the product goes down. For an AI product that has found PMF, users who experience an outage notice immediately and tell you about it — not because the feature was interesting, but because their workflow stopped. The product has been incorporated into a task they complete regularly, and its absence creates a visible gap. If your AI product goes down for two hours on a Tuesday and you don't hear from users, the product has not been integrated into any workflow that matters to them.

Expansion happens at the user level, not just the account level. Conventional SaaS land-and-expand tracks the account: a team buys five seats, grows to twenty seats, then to fifty seats. AI products with genuine PMF show a different pattern: individual users expand their usage of the product within a single account, applying it to more tasks, more documents, more workflows — without any prompting or sales activity from the vendor. This user-level expansion is the signal that the product has found a genuine productivity multiplier in the user's work, not just filled a role in an organizational procurement process. Track per-user usage trends across cohorts, not just seat count.

Users describe the product in terms of outcomes, not features. When you ask users of an AI product that has found PMF what they use it for, they describe what they accomplish — "I use it to get through my contract review backlog," "it's how I prep for customer calls," "it handles the first draft of everything my team writes." When you ask users of an AI product that has not found PMF what they use it for, they describe the feature — "it summarizes documents," "it generates content," "it answers questions." The outcome framing indicates that the product has been integrated into a workflow and is delivering a result the user values. The feature framing indicates that the user knows what the product does but has not found a workflow where it consistently delivers value.

The question to ask in every user interview is not "what do you think of the product?" It is "when was the last time you used it, what were you trying to accomplish, and what did you do differently because of what it gave you?" The answers to those three questions tell you more about PMF than any retention dashboard.

Non-technical users use it without help. Enterprise AI products that have found PMF are used regularly by people who did not participate in the pilot, were not part of the onboarding cohort, and did not receive direct training on the tool. These users found the product through word of mouth within the organization, figured out how to apply it to their work on their own, and returned to it because the value was apparent without a guided demonstration. If every active user in your enterprise accounts went through your onboarding process, the product has not yet found organic internal spread — which is the mechanism that drives NRR above 120% in AI SaaS.

The retention curve flattens after week eight. The signature PMF retention curve for AI products — as opposed to the novelty-premium curve — shows a retention drop in the first four to six weeks (as the curiosity-driven users who were never going to integrate the product into their workflow churn) followed by a flattening of the curve from week eight onward. Users who remain after eight weeks are the ones who integrated the product into their work. Their retention from that point forward should look similar to well-retained conventional SaaS — high but not universal, with gradual attrition from role changes and organizational shifts rather than from value disillusionment. If the retention curve is still declining steeply at week twelve, the product has not found PMF in the cohort you are measuring.

What to Instrument and Track

The Build-Measure-Learn Loop for AI Products

The challenge for teams building AI products is that the feedback loop for PMF is longer and noisier than for conventional SaaS. A conventional SaaS product can run A/B tests on conversion flows, read the results in two weeks, and iterate. An AI product's PMF depends on whether the product has been integrated into the user's workflow — a process that takes weeks, not days, and that involves organizational and behavioral change that no A/B test can measure.

The build-measure-learn loops that work for AI PMF iteration:

Use case specificity over breadth. AI products that try to serve many use cases before they have served one use case deeply rarely find PMF. The products that find PMF earliest are the ones that serve a specific, well-understood workflow so well that users cannot imagine going back to their previous approach. Document review for lawyers, contract analysis for procurement teams, customer call preparation for account executives — these are specific enough that the product can be built and instrumented to measure genuine value delivery. A general-purpose AI assistant that can do many things adequately rarely creates the workflow integration that produces genuine PMF.

Track time-to-value, not activation. Activation metrics — "the user completed onboarding and ran their first query" — measure whether the product was set up, not whether it delivered value. Time-to-value metrics measure the interval between a user's first session and the moment they completed a task that would have taken them significantly longer without the AI product. Identifying this moment and instrumenting it — first contract reviewed, first call summary generated, first code review completed — gives you a leading indicator of retention that activation metrics miss. Users who reach a genuine value moment in their first session retain at substantially higher rates than users who complete onboarding without a value moment.

Interview churned users, not just retained ones. The most useful information about AI PMF gaps comes from users who activated, engaged for two to four weeks, and then churned. These users were interested enough to try the product genuinely but did not find a reason to continue. Their answers to "what did you use it for, what was the result, and why did you stop?" are the most direct signal of what the product is missing for PMF. Retained users will tell you what they like. Churned users will tell you what the product actually needs.

Product Decisions That Accelerate AI PMF

The product decisions that consistently move AI products toward PMF faster, based on patterns across companies that have found it:

Reduce the distance between the AI output and the user's existing workflow. The AI products that find PMF fastest are the ones that deliver output in the format the user already works in — inside the tools they already use, in the document formats they already share, in the review interfaces they already navigate. An AI that produces a well-formatted summary that still requires copying, reformatting, and integrating into a separate document has a higher workflow friction than one that produces the output directly in the user's Google Doc or Notion page. Every integration that reduces the distance between AI output and the user's existing workflow reduces the behavioral change required to retain the product — and behavioral change is the primary friction that prevents AI products from finding PMF.

Give users control over the output quality signal. AI product quality is not uniformly perceived across user populations. A summarization quality that one user considers excellent another user considers insufficiently detailed. Products that give users meaningful control over output parameters — detail level, tone, format, domain-specific terminology — and that learn from user corrections over time build higher retention than products that deliver a fixed output style with no adaptation path. The retention signal from user-configurable output is not just about preference — it is about the product becoming more tailored to the user's specific workflow over time, which makes it progressively harder to replace with a competitor product or a general-purpose alternative.

Build the workflow integration before the enterprise sales motion. The most common pattern in AI companies that struggle to convert pilots to renewals is that the product was built to demonstrate well in a pilot context — impressive capability demonstration, clean user interface, polished onboarding — but without the workflow integrations that make the product genuinely useful in the enterprise environment where most users work. Email integration, calendar access, document system connectivity, CRM write-back — these are table-stakes for an enterprise AI product to achieve genuine workflow integration. Building them before the first enterprise sales motion, not after the first failed renewal, is the difference between a company that scales enterprise AI adoption and one that runs repeated pilots that never convert.

At UData, we work with teams building AI-powered products — both as a dedicated development partner and through our automation and AI services. The PMF measurement conversation is one we have at the architecture stage, not after six months of building — because the instrumentation decisions, the integration priorities, and the use case specificity choices that determine whether an AI product finds PMF are engineering decisions as much as product ones. See our case studies for examples of AI product work where these principles were applied from the start.

What OpenAI and Anthropic Got Right

The Willison argument about OpenAI and Anthropic's PMF is worth examining through this lens. The signal that separates their position from companies with the appearance of AI PMF is not the model quality, impressive as it is. It is the workflow integration depth of the API products. The companies that are paying OpenAI and Anthropic the most are not buying access to a chatbot — they are buying a reasoning and generation capability that they have embedded so deeply into their own products and workflows that the cost of switching has become prohibitive. That switching cost is not created by contracts or lock-in mechanisms. It is created by the depth of integration: fine-tuned behaviors, workflow-specific prompting infrastructure, user experience built around specific model capabilities, and the accumulated institutional knowledge of how to get the best output from a specific model family.

This is the PMF that AI products should be building toward: not "users like our product" but "our product is integrated deeply enough into our customers' operations that its absence would require significant reorganization of the workflows that depend on it." That depth of integration takes longer to build than a polished onboarding flow, but it produces the retention and expansion dynamics that the genuine PMF numbers reflect.

Conclusion

AI product-market fit is real, measurable, and different enough from conventional SaaS PMF that it requires different instrumentation, different iteration loops, and different product decisions to achieve. The companies that have found it — and the signals from OpenAI and Anthropic confirm that it is achievable at scale — have built products where genuine workflow integration has occurred, where users notice the absence of the product immediately, and where usage expands organically as users find more of their work that the product handles better than their previous approach.

The path there is not mysterious, but it requires discipline: specific use case focus before breadth, time-to-value instrumentation before activation metrics, workflow integration before enterprise sales motion, and user-level expansion tracking before account-level expansion. The teams that ship AI products with these principles as first-class engineering and product considerations will find PMF faster than the teams that build for impressiveness and iterate toward usefulness later.

If you are building an AI product and want to think through the architecture, instrumentation, and integration decisions that affect PMF trajectory, reach out. The engineering decisions that determine whether an AI product finds PMF are made in the first few sprints, not after the first renewal cycle.