Product-market fit in 2026: bespoke meets deterministic, and most AI products break in between

Product-market fit in 2026 has stopped looking like a single curve –  because it now splits into two user types that are adopting your product at the same time – human users and AI agents. And the irony is – human users now also expect something different now that AI tools have spoilt them by letting them “vibe code” whatever they need: that the software they use will be a lot more tailored to their use case. Otherwise – they threaten to cancel the subscription and vibe-code it themselves.

So how are SaaS products supposed to reach (and measure!) their product market fit if each custer expects a “bespoke” product tailored to their individual needs? 🤔

A second issue is the one most PMF guides still don’t acknowledge, and it shows up when AI agents start using your product as buyers in their own right. Agents don’t click or scroll, they call MCP tools and read JSON, and they treat your skill files and system prompts as instructions to interpret rather than as deterministic code to execute. A retention curve built around human session counts is not going to reflect agentic usage. You may think it’s not your problem yet – but ~80% of new signups at companies like Netlify are now from agents – who outnumber humans.

Third issue is with product teams rushing to “vibe code” their products themselves – without doing proper discovery, to match the increased engineering velocity expectations and the “productivity dogma”. Andrea Saez, Head of Product Marketing at Turtl, has been making a related case for a while now – that “you can’t vibe-code your way to product-market fit” . I agree with most of it, with one extension: even when AI can technically ship a tailored product on demand, it can’t ship a reliable one, and reliability is the aspect of PMF that nobody currently talks about enough.

So considering all of the above – how we define and measure PMF in 2026 has changed, so we need to discuss the following

• the bespoke-vs-deterministic paradox most teams are quietly losing to.
• Krzysztof Szyszkiewicz’s Service as Software reframe and why it’s the right shape for PMF in 2026.
• how PMF in 2026 splits into two measurable streams: one for human users, one for AI agents.
• rebuilding the classical 5-step PMF playbook (target customer, value proposition, MVP, test, iterate) for an agentic 2026 market
• what we’re shipping at Userpilot, including Agent Analytics, so SaaS teams can actually measure both streams.

If you only have the time to read one section, read the one on the two-stream PMF model.

Quick version, before we go deep

• What is PMF in 2026: The definition hasn’t changed (a product that meets strong market demand with enough value to keep customers), but the curve underneath it has gone from a single binary signal to a moving spectrum and Elena Verna’s “PMF treadmill” of re-finding fit every quarter is closer to the norm than the exception.
• Why it matters more in the agentic era: The cost of building has collapsed, the cost of being wrong has grown, and PMF is the only thing that tells you whether the next feature you ship is worth the runway it burns.
• The bespoke-vs-deterministic paradox: Human users now expect tailored agentic software AND deterministic reliability in the same product, which is the exact gap LLMs cannot close on their own.
• SaaS to Service as Software: Krzysztof Szyszkiewicz’s reframe is the right one, because winners merge SaaS’s deterministic spine with a per-customer agent layer that delivers tailored outcomes.
• The two-stream PMF model: The single retention curve is dead in agent-heavy products, so you need to measure human PMF (Sean Ellis, retention, NPS) and agent PMF (self-service rate, failure rate, top entry points) as separate signals.
• The 5-step playbook, rebuilt for 2026: Target customer, value proposition, MVP, test, iterate. The bones are the same and every step now has to account for two buyer types and a market that resets every three months.
• Why vibe-coding alone won’t get you there: Andrea Saez at Turtl says it best, and I agree, vibe coding is a sharp tool but not a strategy, and it cannot replace the discovery work that creates fit in the first place.
• The PMF stack at Userpilot: Lia, Workflows, the Userpilot MCP Server, and Agent Analytics are designed for exactly this two-stream measurement and delivery problem.

What is product-market fit, and what changed in 2026

Product-market fit is the point where you identify a strong target-market demand and your product fulfills it well enough that customers keep coming back. It leads to a validated product with the potential to grow without burning cash on demand creation. The definition has been stable for fifteen years and that part has not changed.

Product managers own this process, traditionally called product discovery, and as we’ll see it’s more than just coding a product. It involves human skills (and a kind of judgment about ambiguous data) that AI is still very far from replicating reliably. That part hasn’t changed either.

What did change in 2026 is the shape of the curve underneath the definition. Most teams still treat PMF as a binary they either have or don’t, and the Bessemer team’s PMF playbook for AI founders argues the opposite, which is that PMF in the AI era is a spectrum running from light-signal (a handful of users love it, retention is patchy), through moderate-signal (one segment is pulling hard), all the way to strong-signal (retention is high, word of mouth is doing the selling for you). Calling PMF a single binary state in 2026 is the same mistake as calling product feature analysis a single dashboard.

Speed is the second thing that broke. Elena Verna’s read is that “the era of hitting product-market fit, maintaining it, and scaling it is dead”, and most AI-native companies are now re-finding PMF on a quarterly cycle instead of a multi-year one. A single Sean Ellis survey result from last March is interesting historical context, but it tells you almost nothing about whether you have PMF today.

Who the buyer actually is has changed too. Kyle Poyar’s framing in Growth Unhinged says it plainly, that your next customer might be an AI agent. Gartner now projects 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025, and any product that serves agents on the buy side or the user side now needs a separate measurement story for them.

Why achieving product-market fit matters more in the agentic era

Finding product-market fit is still the primary goal of any product team, and it matters more in 2026 than it did before for an unintuitive reason. The cost of building has collapsed, which means the cost of building the wrong thing has collapsed with it, and the only thing that tells you whether your next feature is worth the runway it burns is whether the market actually wants it.

Concretely, achieving PMF gives a product team four things in 2026 that nothing else can replace:

• It identifies whether a product or feature meets a market’s demand, which is what makes the solution profitable and sustainable past the experimentation phase.
• It helps your team prioritize what users actually need, which optimizes your marketing and development resources and protects you from the product strategy drift that comes from chasing competitor features.
• It guides the positioning, messaging, and priorities in your product marketing strategy so the story you tell aligns with the voice of the customer rather than the voice of the loudest investor.
• It’s the basis of continuous product growth in a market where you have to re-earn that growth every quarter as new models ship and customer expectations re-anchor.

Roughly 34% of startups fail because they never find PMF, and a meaningful share of the rest fail because they thought they’d found it and stopped iterating. The 2026 version of that second failure mode is worse than the 2020s version, because the market resets faster.

The bespoke-vs-deterministic paradox no one is solving

The single hardest part of finding product-market fit in 2026 is that the same customer now demands two things from two different historical worlds. Customers want software tailored to their organization, data, and workflows (the way bespoke on-prem builds used to deliver), and at the same time they want deterministic reliability and security (the way SaaS has done it for twenty years). They used to pick one or the other, and now they expect both inside the same product.

LLMs alone give you the tailoring and break the reliability. As I wrote in my last newsletter, the model is probabilistic by design, which means the same instruction can produce different outputs across two runs, and the agent may not even load the same context every time depending on which skill file, retrieved chunk, or chat history it pulls in. Used to the deterministic old-school SaaS mindset, customers treat that variance as a bug instead of an architectural feature, and most “vibe-coded” replacements collapse under exactly that weight.

This is the loop I keep seeing repeated. A team decides their SaaS vendor is too generic and vibe-codes something tailored to their workflow over a weekend, it works for two weeks, then it breaks in a way no one on the team can debug because they have neither the engineering expertise to maintain it nor the subject-matter expertise to spec it correctly in the first place. Hubsessed has been tracking this pattern across HubSpot accounts, and the boomerang back to the original vendor is now common enough that there’s a small industry of RevOps consultants who specialize in cleaning up the migration.

What changed is that AI agents that live inside SaaS can now tailor the product to each customer on top of a deterministic spine, instead of replacing the spine entirely. One thousand customers can each get one thousand versions of the product while the SaaS vendor stays responsible for reliability, security, and the boring stuff that keeps the lights on. That is the merge of bespoke and deterministic that neither pure on-prem nor pure 2020s SaaS could deliver on its own.

A second shift I keep telling our team about is that the operator role is disappearing. Customers used to log into SaaS and operate the software directly, and now they create a project, tell the AI what they want, and evaluate or monitor the agent workflow that delivers it. PMF for that customer means the agent reliably delivers the outcome they asked for, not that the dashboard looks pretty when they log in to check.

From SaaS to Service as Software: Krzysztof Szyszkiewicz’s reframe

The cleanest reframe of what’s actually happening came from Krzysztof Szyszkiewicz at Valueships, who argues that the only way Software as a Service survives in the age of AI is by transforming into Service as Software. He sketches the timeline as three eras and once you see it you can’t unsee it.

On-premise software in the first era was tailored to each organization and great for fit, but heavy on cost and slow to evolve. SaaS came next and made software cheap, fast, and predictably reliable, at the price of never quite tailoring to any organization’s last mile. What’s arriving in 2026 is the merge of the two, where an agent layer tailors the product back to each customer on top of a deterministic SaaS spine.

The problem with SaaS was always the last mile. It’s easy and cheap to move something from China to Chile, the real challenge is shipping it to Avenida Libertador General Bernardo O’Higgins street. SaaS has been the same: cheaper than on-premise, much more accessible, faster shipping and staying up to date. But it was never tailored to your org. And most of the value created is created right here, at the last mile.

Krzysztof Szyszkiewicz, Partner and Co-founder, Valueships

The reason this matters for PMF is that the unit of value has shifted. Buyers used to pay for software seats and PMF meant enough humans logged in often enough to justify the seat count. In the Service as Software world buyers pay for outcomes, which means PMF now hinges on whether your agent layer reliably delivers what the customer asked for on top of a SaaS spine you can be held responsible for (and pricing follows the outcome, not the headcount).

This is also why “vibe-code your own SaaS replacement” almost always boomerangs back to the original vendor. There’s a reason SaaS companies hire product managers and backend engineers, which is to do proper discovery on user needs and then to make sure the resulting solution is reliable and secure. You can compress some of that work with AI, but you can’t skip it without paying for it later.

The two-stream PMF model: humans and agents are different buyers

If you take only one practical idea from this article, take this one: stop running PMF as a single signal stream and start running it as two. One stream measures whether human users find enough value to stick around. The other measures whether AI agents using your product (on behalf of humans or as buyers themselves) can complete the tasks they came to do.

Most of the classical human-side measurements still apply on the human PMF stream. Sean Ellis at ≥40% “very disappointed” still works in 2026 (and runs higher on tightly-segmented power-user cohorts, where Superhuman hit 58% before scaling), the retention curve still tells you whether usage flattens above zero by month six, and NPS plus word of mouth continue to do most of the work they always did. None of that has changed and none of it should be retired.

What’s new is the agent PMF stream, which needs its own measurement spine.

Five signals that are the most accurate PMF hallmarks in 2026:

• self-service rate (does the agent resolve the user’s need without human escalation),
• top entry points (where users actually call the agent first),
• top questions (what they ask repeatedly),
• failure rate (where the agent gives up or hallucinates),
• and satisfaction per conversation.

None of these surface in classical product funnel analytics, which is why most teams measuring agent products today are flying half-blind.

This separation also matters for sales and contract structure. As SaaStr has been pointing out, buyers of AI-agent products are increasingly refusing to commit beyond one year, which means gross retention has become equivalent to net retention for these contracts, and 85% or even 75% gross retention is plausible for vendors who haven’t built stickiness beyond the agent itself. Knowing your agent PMF stream is what tells you whether the renewal is going to happen.

This is why we built Agent Analytics as a separate stream inside Userpilot rather than retrofitting agent data into the human dashboards. It captures AI agent conversation logs and interaction data, measures the agent-side metrics above, and sits alongside (not inside) the human analytics views. One product, two buyers, two measurement streams.

💡 Read related blog posts: The PLG funnel in the agentic era: where the agent stream enters the human funnel

How to find product-market fit in 2026: a 5-step playbook

The five classical phases of finding product-market fit still hold, which is the part of the existing playbook nobody has managed to disrupt. What changes is that every phase now has a 2026-era twist, because you’re building for a moving market and two different buyers at once. Andrea Saez at Turtl puts the core point sharply when she says you cannot vibe-code your way to PMF, because vibe coding still needs a set of requirements that meet basic user needs, and that requires real understanding of the problem, market, and solution.

1. Determine a tighter target customer than you think you need

The first step is the one most teams shortcut and then pay for later. You have to pick a single, tightly-defined ideal customer profile and stay there long enough to learn the shape of their workflow before you generalize. Bessemer’s playbook is unambiguous on this, that PMF is clearest when you serve a narrow ICP first and land-and-expand from there rather than serving five buyer personas badly on day one.

For the target market work itself, do the boring stuff well. Run proper market research to find the gap, analyze third-party data (industry reports, benchmarks, databases) alongside it, and collect your own insights via surveys, interviews, or focus groups. The goal is one common problem inside a niche, named precisely enough that you can tell whether a given user is in or out of it.

From there, you can create a user persona that captures their full context, not just their job title.

The deeper persona work uses psychographic and behavioral data to understand the routines and constraints around their problem, the people they collaborate with, and the outcome they’re actually trying to reach.

After collecting enough data, the output of this step is a profile you can argue from, like this one:

2. Define a value proposition for two buyers (the human and the agent)

The value proposition step is where the 2026 version diverges hardest from the 2020s version. You now have to write two propositions, not one:

• Human proposition: the outcome the user needs and the time the product gives back to them, written around their actual customer pain points.
• Agent proposition: clear MCP tool definitions, predictable response shapes, and a reason for an AI agent to pick your tool over a competitor’s when it’s making the choice on the user’s behalf.

Geoffrey Moore’s template still works as a scaffold:

“For [target customer] who [statement of the need or opportunity], our [product/service name] is [product category] that [statement of benefit].”

Write that twice (once for each buyer), and make sure both propositions point at the same underlying outcome. AI can help brainstorm angles, and in the end you have to settle on a value proposition that defines your target market in detail, differentiates from existing solutions, creates a narrative around the users’ pain, and is feasible, specific, and motivating.

3. Build an MVP and instrument it like it’s already at PMF

The third phase is to build a minimum viable product using the persona and the value proposition you just wrote. Both inputs feed the build directly: the persona and market research insights help you focus on the outcomes your customer needs, while the value proposition determines which core features to ship, what technology to use, and how you’ll position the product in market.

AI is most useful at this stage for developing the prototype, not for replacing the planning that decides what to build.

The trap most teams fall into now is treating the speed gain from vibe coding as a substitute for instrumentation. An MVP that ships without measurement is one you can’t learn from, and learning is the entire point of the MVP phase. Plan the product experimentation framework before you write the first line of prompt, not after.

At Userpilot we follow the same rule internally. Our PM Abrar Abutouq has a standing practice that whenever we release a feature, we create a report and track meaningful events on day one to see usage and feature health, and when we shipped our email feature, the funnel showed a sharp drop at domain verification, so within a few hours she built a targeting tooltip directly inside Userpilot that highlighted the correct steps and closed the drop-off without involving the dev team. That ship-instrument-intervene cycle is what an MVP phase looks like when it’s working, and it’s the small-scale version of the product dogfooding we use to validate everything we build.

4. Test with both customer streams, not just human ones

The testing phase is where most AI-product teams in 2026 still under-invest. You need beta testers from your target persona running the human flow, real agent traffic exercising the agent flow, and feedback collection from both. Run in-app surveys for the humans, PMF surveys on the Sean Ellis question, and agent-side analytics on every conversation log.

Track in-app behaviors alongside conversation logs so you can connect what each stream did with what each stream actually said.

The signals to watch on the human stream are the classical ones: retention curve flattening above zero, ≥40% “very disappointed” on the Sean Ellis test, customer lifetime value tracking toward 3x customer acquisition cost (the standard SaaS sustainability benchmark), and word-of-mouth pull starting to show up in attribution. On the agent stream, watch self-service rate climbing, failure rate trending down, satisfaction-per-conversation holding above your bar, and feature-adoption-via-agent showing up in the data.

Don’t skip the qualitative side. Beyond bug reports, the feedback you collect from beta testers tells you whether there’s real demand for the outcome you’re selling.

The harder qualitative read is whether the product can actually satisfy customer needs in the messy way real customers run them, which a clean retention chart can hide for months before it shows up at renewal.

Mix qualitative data from interviews and small focus groups with the behavioral data from analytics, because either one alone will mislead you.

5. Iterate against the PMF treadmill, not against a fixed bar

Your MVP won’t be perfect. There will be value gaps, missing integrations, and rough edges that show up under real load. Elena Verna’s “PMF treadmill” framing changes how to think about iterating against those, because the old “find PMF then defend it” loop has been replaced by a continuous iteration cycle.

That means treating PMF as a habit, not a milestone. Run a quarterly cycle that refreshes the product roadmap against the latest signals from both streams, re-runs the Sean Ellis test on a recurring cadence, and watches second-bite usage rate (does the same user come back for a second project, not just a second session) as the early indicator that a usage habit is actually forming.

Anchor the cycle on a clear product vision so the iteration doesn’t drift into feature soup as new model releases shift the goalposts every few months.

This is the stage where you need product managers most. Someone has to understand the user experience, decide when to launch, run the discovery process continuously, and prioritize ruthlessly.

Reach for a feature prioritization matrix to make the trade-offs visible to the whole team rather than locked in one PM’s head.

For per-initiative scoring, a RICE score calculator works for the same job and is especially useful when you’re weighing reach and impact against effort across competing AI-and-human initiatives in the same quarter.

Why “vibe coding” alone won’t get you to product-market fit

In the era of vibe coding, you can create a whole prototype quickly with a few prompts. That dramatically increases the speed at which you can build, test, and iterate on products on the way to PMF, and some voices have used that speed to argue that product managers themselves are now redundant. Andrea Saez has spent the past year making the counter-case, and her version of it is the one I want to carry into the rest of this section, because she gets the underlying argument right.

You still need to understand the market and users to vibe-code a product

Even assuming AI could ship a perfectly coded product (and it doesn’t), you still need to know what to build and for whom. Andrea puts it like this:

“Vibe coding needs a set of requirements to meet basic user needs. This means having an understanding of the problem, market, and solution you’re experimenting with.”

Andrea Saez, Head of Product Marketing at Turtl

Current AI technology (LLMs in particular) isn’t designed to have initiative or to make ambiguous calls. If you want to generate any code worth shipping, there must be someone on the other side of the screen who knows the market gaps (or how to find them), understands the problem, and can articulate a vision of a successful solution clearly enough that the model can act on it.

Without those fundamental skills you can’t just ask a tool to build a SaaS with product-market fit. It will produce something that looks like a SaaS, and the thing it produces won’t have fit.

AI can’t replace all areas of the product manager’s role

AI is a serious tool that performs real parts of the discovery process. It can brainstorm product ideas, generate prototypes, and write the first draft of a PRD faster than most humans can.

It will also run feature validation against quick rubrics surprisingly well, which is useful as a sanity check before committing engineering hours to a build.

If you’re starting a PRD from scratch, a working PRD template plus a good prompt gets you to a defensible first draft in minutes rather than days, with humans editing the parts that need real judgment.

Andrea is direct about where that ends:

“If your role begins and ends at shipping tickets and validating features, yes… GPT-5 can probably do that better than you.”

Andrea Saez, Head of Product Marketing at Turtl

Where AI struggles is in the very core of what creates market fit and what makes product managers indispensable, which is the work of aligning a product with market demands as those demands shift. Going further into the PM role, there are several areas the current generation of models can’t handle reliably:

• Connecting user behavior tracking to actual business outcomes when the data is noisy.
• Managing stakeholder expectations across executives, engineering, sales, and support, where each group reads the same number differently.
• Aligning the product with the business strategy when leadership keeps changing its mind about what the business is.
• Crafting a narrative based on context, emotion, and empathy that lands with both buyers and end users.
• Determining the best course of action based on ambiguous data when the model would happily produce three contradictory recommendations with equal confidence.
• Running product operations across multiple squads, owners, and release trains without dropping commitments.

Vibe coding aids the discovery process but can’t produce a complete product

During AI discussions, people assume AI can yield positive results without technical expertise or prior research. The reality is the opposite. In Andrea’s experience, vibe coding helps with initial prototypes for landing pages and internal tools, and it’s helpful for communicating product ideas to a designer or developer so they can build a better version on top.

None of those vibe coding projects produces a complete product, let alone a solution you can sell to enterprise customers and stand behind for years. Andrea’s full take is the one I’d post on the wall of every product team running into the vibe-coding hype right now:

“Yes, vibe coding is real. Yes, it’s useful. But it’s not magic, it’s not strategy, and it’s not replacing your product team. It’s a tool. A sharp one. Just don’t wave it around like you invented UX. Remember that the job is to learn fast and build great things, not to move fast and break everything around you.”

Andrea Saez, Head of Product Marketing at Turtl

The line I’d add to Andrea’s is the deterministic one. Even when vibe coding produces something that works on Monday, the non-deterministic LLM at the heart of it can change behavior on Tuesday for reasons no one on the team can trace, and that’s the half of PMF that classical SaaS quietly held up for two decades. Service as Software brings the bespoke back without giving up the deterministic, and that’s what the merge is for.

How good product managers actually use AI in their PMF process

AI is a tool that will stay in most product teams, and the question isn’t whether to use it but how to use it without confusing it for the strategy. Vibe coding can save you serious time when you reach for it for the right jobs:

• Quick prototypes to test ideas, validate hypotheses, and confirm assumptions before you spend a full development cycle on them.
• Internal tools, dashboards, and automations that remove bottlenecks for your own team and never need to scale to a customer.
• First drafts of artifacts that humans will edit (PRDs, research summaries, positioning statements) where the AI gets you 60% of the way and the human delivers the last 40% that actually matters.
• Keeping you fluent in the technology so when you add AI features to your own product, your judgment about product design and development trade-offs is informed by hands-on experience rather than vibes.

If you want practical, example-based ways PMs can combine discovery, instrumentation, and lean builds to find product market fit faster and with less waste, join Andrea Saez’s Product Drive talk for free. The conference is free, the seats are limited, and the talk is the cleanest forty minutes you’ll spend on this topic this year.

The PMF stack we’re building at Userpilot

The reason I wanted to write this article under my own byline is that we’ve been building Userpilot for exactly this shape of problem, and I want our customers to know what they’re getting and why. Our PMF stack has four pieces, and each one maps to a part of the two-stream model above.

Agent Analytics is the agent-stream measurement layer. It captures conversation logs, usage trends, self-service rate, top entry points, top questions, failure rate, and satisfaction per conversation, and presents them separately from the human analytics stream so you can read both signals without conflating them.

Lia, our AI agent, is the operator-replacement layer. She builds in-app onboarding flows autonomously, monitors product usage signals continuously, surfaces growth drivers and churn risks proactively, and gives PMs a real shot at the operator-to-monitor shift I described earlier. You describe the outcome, the agent ships the experience.

The Userpilot MCP Server is the integration layer between the deterministic SaaS spine and the agent layer. Any AI agent your customers are using (Claude, ChatGPT, vertical agents inside their own products) can call into Userpilot via MCP and get product usage data, session data, and survey data in a structured form, without having to be retrained or wired up by hand. We use the MCP Server to track product depth across both streams so the agent and the analytics stay aligned.

Workflows is the orchestration layer that ties human-side engagement (in-app guidance, email triggers, mobile push) to agent-side events, so a user activated by an agent on Monday gets the right human follow-up on Tuesday without either side knowing the other ran.

It’s also where we surface product qualified leads back to the GTM team in a form they can act on, instead of letting agent-driven activations disappear into a different system.

Put together, those four pieces are how I think SaaS becomes Service as Software for real, and how teams measuring PMF in 2026 stop misreading their own data. If any of this maps to what you’re building, come see a demo, or reply to my next newsletter and tell me what we’re getting wrong.

FAQ

How do you measure product-market fit in 2026?

Run two streams in parallel. The human-side metrics still apply (Sean Ellis test at ≥40% “very disappointed”, retention curve flattening above zero by month six, customer retention rate above 40%, LTV ≥ 3x CAC, and organic growth via word-of-mouth), and the agent-side metrics sit alongside them (self-service rate, top entry points, top questions, failure rate, and satisfaction per conversation). Single-stream measurement is what produces false positives in agent-heavy products.

Is the Sean Ellis 40% test still relevant?

Yes, with two adjustments. Segment the survey by user type (Superhuman hit 58% by running it against power users only), and run it on a recurring quarterly cadence rather than once. The threshold still works and a stale answer is worth almost nothing in a market that resets every three months.

Can you vibe-code your way to product-market fit?

No, but vibe coding is a useful accelerator inside the PMF process. It compresses prototype cost, helps validate ideas with potential customers faster, and is great for internal tools that remove bottlenecks. What it cannot do is replace the discovery work, the subject-matter expertise, or the engineering rigor that keeps a product reliable enough to retain customers long-term.

How is PMF different for AI agents as buyers?

AI agents are typically refusing to commit beyond one-year contracts, which means gross retention becomes the meaningful retention figure (no multi-year buffer). Your PMF for an agent-buyer cohort lives or dies on the agent’s ability to complete its task reliably and on your product’s discoverability inside the MCP toolset the agent is choosing from.

How quickly does product-market fit decay in the AI era?

Elena Verna’s framing is the cleanest answer here: most AI-native companies are re-finding PMF on a roughly quarterly cycle. The classic “achieve PMF, then defend it” model has been replaced by “achieve PMF, then refresh it continuously.” Treat PMF as a habit, not a milestone, and keep a product roadmap template handy for the quarterly refresh.

About the author

Yazan Sehwail

Co-Founder & CEO

All posts