AI made feasibility cheap to prove. Value, usability and viability still need their own evidence, and a great prototype demo counts for less than it feels. The Risk Receipt shows what you've actually retired.

The four product risks are value (will customers buy or use it?), usability (can they figure it out?), feasibility (can we build it?) and viability (does it work for the business?). Marty Cagan's framing still holds, but AI changed the economics: feasibility is now cheap to test, while the other three still need their own evidence. A prototype demo says far less about value than it feels.

Key takeaways

  • Cagan's four big risks are independent: evidence for one tells you little about another (SVPG, 2017).
  • In CB Insights' 2026 analysis of 385 failed VC-backed startups with known causes, 43% cited poor product-market fit and 19% unsustainable unit economics. Those are value and viability problems, not build problems (CB Insights, 2026).
  • Five test users find about 85% of usability problems, so usability evidence is cheap to buy (Nielsen Norman Group, 2000).
  • Value evidence is behaviour with a cost attached, usually gathered before people see your solution.
  • A Risk Receipt keeps each claim, its evidence and an honest "counts / weak / unverified / unexamined" verdict together, so nobody mistakes four nice demo calls for validation.

What are the four product risks?

Cagan set out the four big risks in a 2017 SVPG article, and they have become the standard checklist for product discovery (SVPG, 2017):

Risk The question Owner in Cagan's model Strongest evidence Weakest evidence
Value Will customers buy it, or users choose to use it? Product manager Behaviour with a cost: pilots, pre-orders, workarounds they already pay for "I'd use this", demo enthusiasm
Usability Can users figure out how to use it? Product designer Unassisted task attempts by target users Stakeholder walkthroughs, "looks clean"
Feasibility Can engineers build it with the time, skills and technology we have? Lead engineer A working path on realistic data, with a list of what was skipped A happy-path demo on seeded data
Viability Does it work for the rest of the business? Product manager Cost-to-serve, pricing tested on a buyer, sign-off from sales, legal, support "We'll figure out pricing later"

In Cagan's words, viability is "whether this solution also works for the various aspects of our business". That covers pricing, cost to serve, sales, legal, compliance and support.

The four risks are independent. A feature can be trivially feasible, perfectly usable and still have no value. A feature can have obvious value and be unviable, for example when each use costs more in inference than the customer pays.

Why did AI change which product risks matter most?

When building was expensive, feasibility was the gate everything queued behind. Teams spent discovery time on the other three risks while engineering estimated. Now that a working prototype takes an afternoon, the queue is gone, and so is the forced pause where the other risks used to get examined.

Benedikt Kantus put it bluntly in an October 2026 essay: "AI has made feasibility almost trivial. It has not touched the other three." (Leading in Product, 2026).

Failure data agrees about where products die. CB Insights studied 431 VC-backed companies that shut down since 2023 and could identify causes for 385. "Ran out of capital" topped the list at 70%, but the report calls that the final cause rather than the root one. Underneath it: poor product-market fit at 43%, bad timing at 29% and unsustainable unit economics at 19% (CB Insights, 2026). And for products that do survive, Pendo's 2019 analysis found 80% of features in the average product are rarely or never used (Pendo, 2019). Value risk, not build risk, is the expensive one.

Is feasibility really solved?

For prototypes, mostly. For production, less clearly. In METR's randomised trial, 16 experienced open-source developers working on 246 real issues in their own repositories took 19% longer with early-2025 AI tools, even though they had forecast a 24% speedup and believed afterwards they had been 20% faster (METR, 2025). A February 2026 follow-up with 57 developers found only weak evidence of a speedup, with confidence intervals that include zero, and METR itself warns the data is an unreliable measure (METR, 2026). The honest reading: the first version is cheap; the edge cases, reliability and maintenance are not free.

What is a Risk Receipt?

The problem isn't that PMs stopped caring about the other risks. It's that a polished prototype feels like evidence for all four. Demo it to five people, hear "this is great", and the team walks away believing value, usability and viability are covered. Each risk has its own currency of evidence, and most teams pay in the wrong one.

A Risk Receipt is a short record you attach to any meaningful product bet. For each of the four product risks it answers three questions:

  1. What's the claim? One sentence, specific enough to be wrong.
  2. What evidence do we have, and what kind is it? Linked to the source, not paraphrased.
  3. Does that evidence actually count for this risk? This is the part teams skip.

The third question is the point. Evidence has an exchange rate. Some evidence is strong currency for one risk and close to worthless for another.

What counts as evidence for each risk?

Value: pay in behaviour, not compliments

Counts: customers describing the problem unprompted, with a recent, concrete example. Existing workarounds (spreadsheets, scripts, manual processes) that cost them real time. Commitments with a cost attached: a calendar slot, a pilot with their data, a pre-order, an introduction to their boss.

Doesn't count: "I'd use this." Enthusiasm during a demo. A prototype's click-through rate in a moderated session. Upvotes on an idea board from people who will never pay.

The tell: if the strongest evidence for value is someone reacting to your prototype, you've collected usability evidence and labelled it value. People react politely to anything competent-looking you show them. Value evidence usually comes from before they saw your solution, in how they talk about the problem. Cheap ways to buy it include a fake door test or a presale interest page.

Usability: pay in observed tasks, not opinions

Counts: watching a real target user attempt a real task without coaching, and recording where they hesitate or fail. Task completion on the first attempt. The questions they ask out loud.

Doesn't count: "It looks clean." A stakeholder walkthrough. Your own team using it every day (they already know where everything is).

The tell: if nobody outside the team has attempted a task alone, usability is unverified, no matter how good the prototype looks.

This is the one risk AI prototyping genuinely helps with. A clickable, data-backed prototype gets you real task observation weeks earlier than a static mockup. It's a real gain, just a gain for one risk. Our usability test template covers how to run the sessions.

Feasibility: pay in a working path, including the ugly parts

Counts: a prototype that runs against realistic data volumes, real permissions and real latency. An engineer's honest list of what the prototype skipped: auth, migrations, error states, the integration that's rate-limited.

Doesn't count: a demo on the happy path with seeded data.

The tell: ask "what did this prototype not have to handle?" If the answer is long, feasibility is partly retired, not fully.

Viability: pay in numbers someone outside product would sign

Counts: a cost-to-serve estimate (for AI features, cost per use at expected volume). A pricing hypothesis you've said out loud to a buyer. A check with whoever owns support, sales, security or legal. For B2B, whether a buyer's procurement process allows this.

Doesn't count: "We'll figure out pricing later." Assuming usage costs are negligible because the prototype was cheap to run for one person.

The tell: if no one outside the product team has seen the bet, viability is unexamined.

A worked example: four happy customers, one empty receipt

Say you're a two-person founding team building a B2B tool, and a customer call sparks an idea: an AI assistant that drafts weekly status updates from project activity. You prototype it over a weekend. It works. You show it to four customers on calls. All four say they love it.

Here's the honest Risk Receipt:

Risk Claim Evidence Counts?
Value Team leads will use weekly drafts instead of writing updates themselves 4 positive demo reactions Weak: reactions to a solution, not behaviour. No one described writing updates as a pain before you showed them.
Usability Leads can review and send a draft in under 5 minutes None observed Unverified: nobody tried it alone.
Feasibility Drafts stay accurate on real project data Weekend prototype on your own workspace Partial: no test on a messy customer workspace.
Viability Cost per draft stays well under what the plan charges Not estimated Unexamined

Four enthusiastic calls, and the receipt says you've retired almost nothing. That's not a reason to kill the idea. It's a reason to run three cheap checks before building the production version:

  1. Value: go back to two of those customers and ask how they write updates today, how long it takes, and what happened the last time one was late. Listen for a workaround.
  2. Usability: give one customer the prototype on their own data and watch them use it, unguided, for one cycle.
  3. Viability: multiply your per-draft model cost by realistic weekly volume across your customer base, and compare it with plan price.

The viability maths (illustrative numbers). Suppose a draft costs €0.06 in model calls on your own workspace, but a real customer workspace has 5 times more activity, so €0.30 per draft. A customer with 40 team leads generates 160 drafts a month, or €48 in model cost, against a plan priced at €60 a month. That is 80% of revenue gone on one feature. Better to know on day three than in month three.

Each check takes days, not weeks. That's the real dividend from cheap feasibility: you can afford to buy real evidence for the other three risks before you commit.

How many customers or users do you need for each risk?

  • Usability: small numbers work. Nielsen's analysis found 5 users uncover about 85% of usability problems, and recommended three rounds of 5 with fixes in between over one round of 15 (Nielsen Norman Group, 2000).
  • Value: the count matters less than the cost of the signal. Two customers who each give you a pilot with their own data outweigh twenty who say "sounds great". Look for at least 3 independent commitments before treating value as retired for a segment.
  • Feasibility: one realistic run, reviewed by the engineer who will own it, plus a written list of what was skipped.
  • Viability: one cost model with stated assumptions and one conversation with each function that has to say yes (sales, support, legal or security).

These are starting points for a small team, not benchmarks.

Why do receipts beat a discovery checklist?

You could turn the four risks into a checklist. Teams do. Checklists fail predictably: someone ticks "value validated" because a demo went well, and the tick hides what the evidence was.

A receipt keeps the evidence attached to the claim. Three months later, when the feature is underperforming and someone asks "why did we build this?", the answer isn't a memory or a slide. It's a record: here's what we believed, here's what we saw, here's which risks we knowingly left open. For the structure that keeps those records linked over time, see our decision graph for small AI teams. If your team ships without receipts at all, check it against the 10 warning signs of a feature factory.

That last part matters. Shipping with an open risk is often the right call, because small teams can't afford to retire everything. The failure is shipping with an open risk you thought was closed.

Common mistakes

  • Counting demo reactions as value. The single most common error. Reactions are usability evidence at best.
  • Testing usability with insiders. Colleagues and friendly customers already know the product's logic.
  • Treating a prototype as proof of production feasibility. Ask for the skipped list every time.
  • Leaving viability to "later". For AI features especially, cost per use can turn a loved feature into a loss-maker.
  • Trying to retire every risk. Pick the one or two that would kill the bet and buy evidence for those first. Our assumption mapping guide helps rank them.

How we run Risk Receipts in Specky

Customer signals (interview notes, support threads, call transcripts, analytics events) land in the Product Graph as typed nodes, each keeping its source. A drafted insight or decision cites the specific nodes it rests on, so "four customers loved it" can't quietly become "customers need this". For the value check, Alex, Specky's AI research agent, can run short interviews about customers' current process before they see your solution, and the transcripts come back as evidence. For early demand, a validation page measures whether people take an action with a cost attached. Judgement about which risks to leave open stays with you.

Start with one receipt this week

You don't need a tool to start. Pick the bet your team is most excited about right now, the one with the best prototype. Write four lines:

Value: [claim] | [evidence + link] | counts / weak / unverified / unexamined Usability: [claim] | [evidence + link] | counts / weak / unverified / unexamined Feasibility: [claim] | [evidence + link] | counts / weak / unverified / unexamined Viability: [claim] | [evidence + link] | counts / weak / unverified / unexamined 

If the most exciting bet has the emptiest receipt, you've found where your next week of discovery should go. You can also sketch the assumptions visually with the free assumption map tool.

Feasibility used to be the expensive question. Now it's the cheap one. The teams that win this cycle won't be the ones who build fastest. They'll be the ones who can show what they knew when they decided.

Frequently asked questions

What are Marty Cagan's four big risks?

They are value risk (will customers buy or choose to use it), usability risk (can users figure out how to use it), feasibility risk (can engineers build it with the time, skills and technology available) and business viability risk (does it work for the rest of the business). Cagan published them on SVPG in 2017 and assigns value and viability to the product manager.

Which of the four product risks is most often underestimated?

Value risk. A working prototype and polite reactions create a strong feeling of validation that isn't backed by behaviour. CB Insights' 2026 study found 43% of failed startups with known causes cited poor product-market fit. Viability is a close second for AI features, where cost per use can outgrow the price.

Does a prototype test value risk?

Rarely on its own. Showing a prototype mainly tests usability and some feasibility. Value evidence comes from behaviour with a cost: people describing the problem unprompted, existing workarounds, pilots, pre-orders or introductions. You can use a prototype in a value test, for example by asking for a paid pilot, but the reaction alone doesn't count.

How do you test business viability risk?

Write down the numbers someone outside product would need to approve: cost to serve per customer, expected price, support load, and any legal, security or procurement constraints. Then show them to those people. For AI features, multiply cost per use by realistic volume and compare it with plan price before committing to build.

Did AI make feasibility risk disappear?

No, it made the first version cheap. Prototypes now take hours, but production work on mature codebases is a different matter: METR's 2025 trial found experienced developers took 19% longer with AI tools on their own repositories, and its 2026 follow-up found only weak evidence of a speedup. Keep asking what a prototype didn't have to handle.

Sources

  1. SVPG, Marty Cagan — The Four Big Risks (2017)
  2. Leading in Product, Benedikt Kantus — vibe coding and product management (2026)
  3. CB Insights — The top reasons startups fail (2026)
  4. Nielsen Norman Group — Why You Only Need to Test with 5 Users (2000)
  5. METR — Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025)
  6. METR — Uplift update on late-2025 AI tools (2026)
  7. Pendo — The 2019 Feature Adoption Report

This story was originally published on 16 October 2026 at https://www.specky.space/blog/what-counts-as-evidence-four-product-risks.

Fathy Shalaby is a scientist by training (master's). He worked at LexisNexis and Canva, and writes How to Be a Product Engineer.