Comparison table
We dug into three problems from our long list. A (feedback synthesis) has the best AI fit and buildability; B (decision context) has the strongest evidence but the most crowded market; C (intake) sits in between.
| A. Sprint feedback synthesis (#23, 24, 25) | B. Assembling context for roadmap decisions (#1, 2, 26, 28) | C. Requests arrive without the real problem (#3, 5, 19) | |
|---|---|---|---|
| Problem | Every sprint, PMs hand-tag support tickets and reviews, then summarize them for their manager. Important feedback surfaces after dependent work has started. | Before a roadmap review, PMs pull requests, usage, estimates, commitments and past decisions from separate systems, then judge whether each source is current. | Stakeholders submit one-line wishes (“better track management”). Required form fields get “N/A” or “TBD,” so PMs chase the requester to learn the real need. |
| Evidence strength | Weak to medium. One detailed account (Sejal: an afternoon of spreadsheet tagging, repeated every 2 weeks). Maya and Daniel echo the trust concern. | Strongest. 5+ accounts (Aidan, Anduril, Charlie, Maya, Daniel, exec notes). Anduril and Charlie each spend about 2.5 hrs per review. | Medium. 4+ accounts (Aidan, Kanan, Anduril, Charlie, Daniel). Two literal-request builds took months of rework (#19). |
| Current workarounds | Spreadsheet tagging and counting; Zendesk/Intercom tags; ChatGPT summaries she then re-checks by hand | Running docs, Notion/Confluence pages, meetings to ask whoever knows; Aidan’s homegrown discovery agent | Intake forms with required fields; clarification calls; Aidan’s proposed intake agent |
| Competitive gap | General AI summaries flatten detail and overstate (“one speed complaint became a broad claim”). Dedicated AI feedback tools are sales-led and built for large teams. No tool ties each claim to its tickets at sprint cadence. | Enterprise search (Glean, Rovo, Notion AI) finds documents but doesn’t assemble a decision brief that flags stale or conflicting sources. Crowded and well funded. | Forms check that a field is filled, not that the answer is usable. No tool asks the requester follow-up questions before the request reaches the PM. |
| ICP | Consumer-app PM at a 50–1,000-person company, owns a shipped feature with steady ticket and review volume, no research-ops support, reports feedback to a manager every sprint | B2B PM at a 500+ person company, owns a roadmap touched by 4+ teams, runs monthly or quarterly roadmap reviews | B2B or platform PM who receives requests from Sales, CS and internal teams through a form or Jira queue |
| SAM / SOM (US + Europe, ~$600 per PM per year) | ~224k PMs, ~$134M / ~2.2k PMs, ~$1.3M ARR | ~338k PMs, ~$203M / ~1.7k PMs, ~$1.0M ARR | ~273k PMs, ~$164M / ~1.4k PMs, ~$0.8M ARR |
| Buildability | High. Input is a CSV or one API (Zendesk, Intercom, app-store reviews). An MVP is feasible this quarter. | Low. Needs many integrations, permission handling and enterprise security review. Regulated PMs can’t use it (#27). | Medium. Simple to build as a form layer, but adoption needs the form-owning team to agree (#6). |
| AI fit | Very high. LLMs can cluster free text and cite the exact tickets behind each theme, which wasn’t feasible before. Counts stay computed, not generated. | High. Retrieval plus summarization with sources is new, but trust and access limits cap it. | High. An LLM can ask the requester follow-ups in plain language until the problem, users and urgency are specific. |
SAM/SOM percentages are team estimates to validate (full breakdown in our Drive).
Where we’re leaning
We’re leaning toward A: sprint feedback synthesis for consumer PMs. It is the narrowest version of the central problem from our affinity map (PMs spend too much time assembling context they can trust). It also has a clear ICP, a fixed cadence, one main data source, and the clearest AI fit. Our studio ICP named what this PM needs from a summary: trustworthiness, specificity, timeliness, credible communication to her manager, time saved, and control over the final call. Current AI tools fail on exactly the first two, so a tool that links every theme to the tickets behind it and computes counts instead of generating them fills a real gap. B has stronger evidence, but it means competing with Glean, Rovo and Notion AI on integrations we can’t build in a quarter. A could later grow into B, since feedback is one of the sources a roadmap review draws on. A’s biggest risk is evidence: it rests mainly on one detailed interview, so this week we will run 3–4 follow-ups with consumer PMs to test whether the per-sprint pain is common and how satisfied they are with the tools they use now.
Reading: We used Lenny Rachitsky’s advice to make the ICP “almost comically narrow” with at least three attributes from his checklist (company size, job title, pain point, way of working) in our ICP section: a consumer PM at a 50–1,000-person company, with no research-ops support, who reports user feedback every sprint.
