AI 時代團隊的產品機會評估:一套現代化模板
Product Opportunity Assessment: A Modern Template for AI-Era Teams
Fathy Shalaby提出面向 AI 時代團隊的產品機會評估模板,將 Marty Cagan 2006 年的十個問題整合為五項。五項涵蓋問題、目標市場、方案、商業案例和風險;評估 AI 功能時,另須檢查模型可行性及競爭壁壘。文章以 AI 通話摘要功能作示例,提出以真實輸入評估模型表現,並把高影響、低信心風險列為測試對象。
Cheap code makes saying no harder, not easier. A product opportunity assessment forces five questions before you build, plus two checks every AI feature now needs.
A product opportunity assessment is a short document that decides whether an idea deserves discovery or build time, by answering a fixed set of questions about the problem, the customer, the solution, the business case and the risks. Marty Cagan's 2006 version had ten questions and ended in a go or no-go call. The modern version below uses five, plus two checks that AI-era products now need.
It isn't paperwork; it's the cheapest place to say no. An idea killed on two pages costs an afternoon. One killed after launch costs engineering time, support load and the roadmap slot it took from something better.
Key takeaways
- Marty Cagan's 2006 opportunity assessment asked ten questions, starting with "Exactly what problem will this solve?" and ending in a go or no-go call, to stop companies wasting money on poor opportunities (SVPG, 2006).
- Poor product-market fit was a factor in 43% of the 385 VC-backed shutdowns CB Insights could explain in its March 2026 analysis (CB Insights, 2026).
- Pendo's 2019 study of 615 software subscriptions found 80% of features rarely or never used, and 12% of features driving 80% of daily usage (Pendo, 2019).
- AI made building cheaper, which removed the natural pause that used to filter weak ideas. The assessment puts that pause back on purpose.
- For AI features, add two prerequisites: model feasibility (can a model do this reliably enough?) and the moat (why won't this be copied in a month?).
What is a product opportunity assessment?
It is a lightweight go or no-go document written before a team commits to discovery or delivery. Cagan described it on the SVPG blog in December 2006 as a quick way to catch bad ideas early (SVPG, 2006). His ten questions covered:
- The problem it solves (value proposition)
- Who has that problem (target market)
- How big the opportunity is (market size)
- Which alternatives exist (competition)
- Why we are best placed to pursue it (differentiator)
- Why now (market window)
- How we reach the market (go-to-market)
- How we measure success and make money (metrics and revenue)
- What is critical to success (solution requirements)
- The recommendation (go or no-go)
The questions still hold. What changed is the cost around them: in 2006 building was slow, so weak ideas often died of their own weight. In 2026 a prototype can exist before anyone has asked question 1.
An opportunity, in Teresa Torres's useful definition, is an unmet customer need, pain point or desire, not a feature (Product Talk, 2023). Assess the opportunity first and the solution second.
Why does an opportunity assessment matter more now that AI makes building cheap?
Because "can we build it?" is no longer the expensive question. "Should we?" is.
The waste is well documented. Pendo estimated that public cloud companies had invested up to $29.5 billion in features that are rarely or never used (Pendo, 2019). In CB Insights' 2026 data, poor product-market fit is second only to running out of capital (70%) among shutdown reasons (CB Insights, 2026). Money runs out faster when it funds products people don't need.
Cheaper code fixes neither number; it just produces more unused features per quarter. Our analysis of what the 2026 AI coding data shows goes through the speed evidence; the short version is that deciding what to build hasn't got cheaper at all.
That is why the assessment works as a "no" engine: if a team can't answer the core questions in a couple of pages, it isn't ready to build, however fast the build would be. Teams that skip it drift into a feature factory, measuring output instead of outcomes.
What questions should a product opportunity assessment answer?
The modern template compresses the ten into five, each answered with evidence, not intuition.
- The problem. What specific, high-friction pain are we solving? Is it a "hair on fire" problem for the user, or a mild annoyance?
- The target market. Who exactly is the customer, and how large is the reachable market (TAM, SAM and SOM)?
- The solution. Why are we positioned to solve this better than the alternatives people use today, including spreadsheets and doing nothing?
- The business case. How does this create measurable value: revenue, retention or operational efficiency? How will we measure it?
- The risks. What are the technical, market and execution risks, and how will we reduce them before committing?
How do the five questions map to Cagan's ten?
| Modern question | Covers Cagan's questions | What a good answer contains |
|---|---|---|
| Problem | 1 (value proposition) | A named pain, how often it occurs, the current workaround and its cost |
| Target market | 2, 3 (target market, market size) | One primary segment; a bottom-up market estimate |
| Solution | 4, 5, 9 (alternatives, differentiator, solution requirements) | Current alternatives and why ours wins for this segment |
| Business case | 7, 8 (go-to-market, metrics and revenue) | A success metric, a counter-metric and a route to customers |
| Risks | 6, 10 (why now, recommendation) | Top risks scored, the tests that would retire them, and a go, no-go or "discover first" call |
Keep "why now?" visible inside the risk section. Many good-looking opportunities fail on timing, which CB Insights lists as a factor in 29% of explained shutdowns.
What should an opportunity assessment add for AI features?
Classic assessments treat the solution as mostly an engineering question. For AI features that breaks in two places.
1. Model feasibility
Can an off-the-shelf model do the job reliably enough, or does the idea need fine-tuning, custom data or a human in the loop? Answer it with a small evaluation, not a demo: collect real inputs, define what a correct output looks like, and measure how often the model gets there. A demo shows the best case; an evaluation shows the rate. Our AI evaluation framework for product teams explains how to set that bar.
Then ask what a wrong answer costs. If it costs the user a minute, a 90% hit rate may be fine. If it costs money or trust, no rate may be good enough. Unsure it needs AI at all? Run the four-question AI test first.
2. The moat
Competitors can call the same models, so name the advantage that survives a fast follower:
- Proprietary data that improves the output and that others can't get.
- Network effects, where each user makes the product better for the next.
- Deep workflow integration that creates real switching costs.
If the honest answer is "we'll be first", treat the moat as missing and price that into the recommendation.
How do you use Working Backwards in an opportunity assessment?
Amazon's "Working Backwards" process is a good test for the solution question. As Werner Vogels described it in 2006, the team writes four documents before building: a press release, an FAQ, a customer-experience description and a user manual, starting from the customer and working back to the minimum technology needed (All Things Distributed, 2006).
For an opportunity assessment, the first two documents are enough. Write a one-page press release for the finished product and an FAQ with the hardest questions a customer or colleague would ask. If the release can't state the customer benefit plainly, the solution question isn't answered.
How do you separate assumptions from facts?
The most dangerous part of any assessment is the hidden assumption. Keep a two-column table inside the document: Assumptions (hypotheses about behaviour, market size or technical performance) and Facts (what interviews, analytics, experiments or market data have validated, each linked).
Every important assumption gets an experiment that turns it into a fact or kills it. If you can't design a test for a key assumption, you carry more risk than the team can see. Our free assumption map sorts assumptions by importance and evidence, and the guide to choosing a product experiment helps pick the right test for each one.
Cagan's four big risks give the table a useful structure (SVPG, 2017):
| Risk | Question | Owner in Cagan's model | Example test |
|---|---|---|---|
| Value | Will customers buy it or choose to use it? | Product manager | Fake door, presale, customer interviews |
| Usability | Can users figure out how to use it? | Product designer | Prototype usability test |
| Feasibility | Can engineering build it with the time, skills and technology we have? | Lead engineer | Technical spike or model evaluation |
| Business viability | Does it work for the rest of the business? | Product manager | Review with sales, legal, finance and support |
For what counts as enough evidence against each risk, see the four product risks and their evidence.
Worked example: assessing an AI feature in one afternoon
Say a 5-person B2B SaaS team, selling a customer-success tool to roughly 120 paying accounts, is considering "AI summaries of customer calls". This team and its numbers are illustrative.
Problem. In 8 of 12 recent interviews, CSMs said they write notes after every call; 5 said notes often go unwritten. Fact: the pain exists for these users. Assumption: it matters across accounts.
Target market. Primary segment: CSM teams of 3 to 15 people already using the product, about 70 of the 120 accounts. A bottom-up SOM is these 70 accounts, not "every company with customer calls".
Solution. Customers' existing meeting tools often summarise calls already. The possible edge: summaries that land on the account record next to health signals.
Business case. Success metric: share of calls with a summary attached within 24 hours rises from an estimated 30% to 70% in 8 weeks. Counter-metric: CSMs editing more than half of each summary, which would signal low quality.
Risks, scored on confidence and impact from 1 to 5:
| Risk | Impact | Confidence we're right | Next step |
|---|---|---|---|
| Customers already get "good enough" summaries elsewhere (value) | 5 | 2 | 6 more interviews focused on current tools |
| Summaries are wrong on product names and numbers (feasibility) | 4 | 2 | Evaluate on 40 real call transcripts |
| Moat is thin: competitors use the same models | 4 | 3 | Test whether account-record placement changes behaviour |
| Storing call data raises privacy questions (viability) | 3 | 3 | Review data handling with legal before beta |
Recommendation: go to discovery, not build, because two high-impact, low-confidence risks remain. One afternoon of assessment turned the next two weeks from a build into two tests.
The modern product opportunity assessment checklist
Use this checklist before you call an assessment done:
- Problem definition: states the customer pain explicitly, with its frequency and current workaround.
- Evidence of need: cites qualitative interview insights or quantitative survey results, linked. Ten interviews is a sensible starting point for a B2B problem, not a benchmark.
- Market size: a bottom-up TAM, SAM and SOM estimate backed by credible, recent data.
- Competitive moat: clear differentiation from incumbents and fast followers.
- AI checks: model feasibility tested on real inputs; failure cost described.
- Success metrics: one primary metric and at least one counter-metric (if adoption rises, does performance or support load suffer?).
- Risk matrix: each risk scored on confidence and impact, with a test for every high-impact, low-confidence risk.
- Recommendation: go, no-go or "discover first", with the evidence that would change it.
When should you skip the full assessment?
Not every idea earns two pages; cheap, reversible changes are often better tested than debated. Ron Kohavi and Stefan Thomke describe a 2012 Bing idea to change how ad headlines were displayed. It needed a few days of engineering, was ranked a low priority among hundreds of ideas, and sat untested for more than six months. When an engineer finally ran a simple A/B test, revenue rose about 12%, worth more than $100 million a year in the US (HBR, 2017). Debate kept that idea on the shelf; a cheap test settled it.
A workable rule:
- Full assessment for new products, new segments, pricing changes, AI features touching customer data, and anything over a few weeks of build.
- Five-question short form (one paragraph each) for mid-sized features.
- Skip it for small, reversible tweaks that can be A/B tested or rolled back in a day.
Common mistakes
- Justifying a decision already made. If the recommendation came before the evidence, it's a sales document.
- Top-down market sizing. "1% of a $10 billion market" says nothing about who buys first.
- Treating a demo as feasibility evidence. A demo shows the best output, not the error rate.
- Never revisiting it. Compare the launch outcome with the business case.
Where Specky fits
Specky keeps an assessment's evidence in one place. Calls, feedback, support tickets and analytics land in the Product Graph, so the problem section cites real sources, not recollection. The Opportunity Tree lays out outcomes, opportunities and solutions, RICE scoring ranks competing ones, and Alex, Specky's AI research agent, runs follow-up interviews from a shareable link when an assumption needs testing. The go or no-go call stays with you.
Frequently asked questions
What is a product opportunity assessment?
It is a one- or two-page document that decides whether an idea is worth discovery or build time. It covers the problem, target customer, solution, business case and risks, and ends with a go, no-go or "discover first" call. Marty Cagan popularised the format on the SVPG blog in 2006.
What are Marty Cagan's opportunity assessment questions?
Cagan's 2006 list has ten: the problem, who has it, market size, alternatives, our differentiator, why now, go-to-market, how we measure success and make money, critical success factors, and a go or no-go recommendation. The five-question template here groups those ten.
How is an opportunity assessment different from a PRD or business case?
An opportunity assessment comes first and asks whether to pursue a problem at all. A PRD follows a go decision and specifies what to build. A business case usually focuses on financial return. The assessment covers all four product risks, not just money, and is deliberately shorter.
How long should a product opportunity assessment take?
For most features, an afternoon to two days, ending in one or two pages. If it takes weeks, the team is doing discovery inside the assessment; split the two. The assessment names which assumptions need testing; it doesn't test them all itself.
What is the difference between an opportunity assessment and an opportunity solution tree?
An opportunity solution tree, introduced by Teresa Torres in 2016, maps one outcome to many customer opportunities, solutions and assumption tests; it's a continuous discovery map. An opportunity assessment evaluates one opportunity in depth at a decision point. Use the tree to find candidates and the assessment to decide on one.
Sources
- Marty Cagan, SVPG — Assessing Product Opportunities (2006)
- CB Insights — The top reasons startups fail (2026)
- Pendo — The 2019 Feature Adoption Report (2019)
- Marty Cagan, SVPG — The Four Big Risks (2017)
- Werner Vogels, All Things Distributed — Working Backwards (2006)
- Ron Kohavi and Stefan Thomke, Harvard Business Review — The Surprising Power of Online Experiments (2017)
- Teresa Torres, Product Talk — Opportunity Solution Trees (2023)
This story was originally published on 17 May 2026 at https://www.specky.space/blog/beyond-the-feature-factory-a-modernized-product-opportunity-assessment.
來源:DEV Community · dev.to