Product Management
20+ discovery calls to a top-3 roadmap brief
· Updated · 5 min read
S
Get AI Workflow Teardown Template
The template I use to map AI workflows before writing a spec.
No spam, unsubscribe anytime. I only email teardowns.
The short answer
Discovery isn't call stacking; it's a ranked deliberation. At Upcore we ran 20+ prospect interviews across 12 verticals, distilled them into an opportunity brief, and surfaced the top-3 agentic-AI use cases by severity, ROI, and buildability. The scoring frame is the deliverable, not the call logs.
Twenty discovery calls will give you twenty compelling stories, and a roadmap can only hold three. That gap — between what you heard and what you're allowed to build — is where most discovery quietly turns into a highlight reel of the loudest prospect. Teresa Torres' continuous-discovery research makes the same point structurally: interview first, map what you heard into an opportunity space, and never let a solution skip the mapping step just because it's exciting. The fix isn't more calls. It's a decision stack ruthless enough that any executive can trace why use case #1 beat use case #7.
The interview phase
- 20+ structured calls with would-be buyers across 12 verticals (manufacturing, logistics, fintech, healthtech, SaaS, edtech, retail, real estate, professional services, media, energy, public sector).
- Each tracked: pain description, budget signal, who "wins" the decision, priority phrasing.
- Interview notes were signal-rich — each one landed as a claim, not a story. A note looks like: "VP Ops: manual dispatch causes 3-hr daily loss; budget approved for Q3; CTO owns; phrasing 'we need eyes on every truck'."
The interview script we actually used
We didn't wing it. Every call followed the same 45-minute structure:
1. Context (5 min)
"What does your week look like? Walk me through the last time [problem area] came up."
2. Pain mapping (15 min)
- "Describe the last time this hurt. What did you do? What did it cost?"
- "Who else feels this? How many people? How often?"
- "What have you tried? What worked? What didn't?"
- "If this vanished tomorrow, what changes?"
3. Budget & authority (10 min)
- "Is there budget allocated for this? Fiscal year or discretionary?"
- "Who signs? Who kills it?"
- "What's the trigger to buy — renewal, incident, mandate?"
4. Priority phrasing (10 min)
- "How do you describe this internally? What words?"
- "What's the competing priority that pushes this down?"
- "Show me the roadmap item / Jira epic / Notion page if it exists."
5. The magic question (5 min)
"If we built exactly what you described, what would make you not buy it?"
That last one surfaced more kill criteria than any feature list. It's the same instinct behind The Mom Test: people will happily bless your idea; they'll only tell the truth about their constraints.
Synthesis out of interviews
The synthesis wasn't thematic clustering — it was a scoring machine, close in spirit to RICE prioritization but reshaped for discovery instead of a backlog: no Reach factor, because at this stage every opportunity is still unproven, and no single Effort denominator, because feasibility varies too much by vertical to reduce to one number.
- Severity — how sore is this pain? (1–5, anchored to "would they stop work to fix it")
- ROI — "would they pay to make it go away?" (budget signal × frequency × time loss)
- Buildability — can we ship this without rebuilding the factory? (tech feasibility × data readiness × go-to-market fit)
Every opportunity got a triple score. The top-3 use cases by this triple are what a roadmap should be allowed to defend. Everything else is a tasteful "not yet."
The scoring rubric (what we actually used)
| Score | Severity | ROI | Buildability |
|---|---|---|---|
| 5 | Stops work daily; regulatory risk | Budget allocated, urgent | Core competency, data ready, GTM clear |
| 4 | Weekly blocker; workarounds expensive | Budget signaled, timeline defined | Adjacent to core, data accessible |
| 3 | Monthly annoyance; workarounds exist | Budget possible next cycle | New domain, data needs work |
| 2 | Quarterly friction; tolerable | No budget, maybe later | Significant R&D needed |
| 1 | Nice to have; no urgency | No budget, no owner | Research project |
We scored independently (PM + eng lead), then reconciled. Disagreements usually meant "we don't understand the build" — which is a signal itself.
The reconciliation meeting (where the real decisions happen)
After scoring independently, PM and eng lead sit down with the spreadsheet. The rules:
- If scores differ by ≥2 on any dimension, we discuss. Usually means one of us missed a constraint.
- If Buildability ≤ 2, the opportunity dies. No "we'll figure it out later."
- If Severity ≤ 2, it's a vitamin, not a painkiller. Deprioritize.
- The tie-breaker is learning velocity. Which 0→1 teaches us the most about the market?
This meeting is where the roadmap actually gets written. The scoring is just the prep.
What the brief said
- 20+ interviews produced a target pulse: agentic AI for scheduling, orders, and dispatch — the cross-vertical pattern.
- 12 verticals surfaced cross-functional flow, not a shadow play. The same agentic skeleton applied to manufacturing dispatch, logistics routing, and SaaS onboarding.
- Top-3 opportunity brief survived C-suite review, and both a pricing play and an intel play leaned on it.
The top-3 that survived
| Rank | Use case | Verticals | Severity | ROI | Buildability | Composite |
|---|---|---|---|---|---|---|
| 1 | Agentic dispatch (manufacturing/logistics) | Mfg, Logistics, Wholesale | 5 | 5 | 4 | 14 |
| 2 | AI routing & scheduling (logistics/field services) | Logistics, Field Svcs, Utilities | 4 | 5 | 4 | 13 |
| 3 | Agentic onboarding (SaaS/B2B) | SaaS, Fintech, Edtech | 4 | 4 | 5 | 13 |
The tie at 13 was broken by learning velocity: dispatch had the clearest 0→1 path.
The pricing play that leaned on the brief
The discovery brief wasn't just a roadmap artifact — it became the pricing framework. For each top-3 use case, we modeled:
| Use case | Pricing metric | Willingness to pay (median) | Model |
|---|---|---|---|
| Agentic dispatch | $/truck/month | $180 | Per-asset SaaS |
| AI routing | $/route optimized | $0.45 | Usage-based |
| Agentic onboarding | $/seat/month | $45 | Seat-based |
These numbers came directly from the budget signals in discovery. We didn't guess — we asked "what would you pay?" and the median answer became the anchor.
The slides that survived
- One slide: pain, outcome, budget signal, buildability — for each of the top 3.
- One ranking: "what we'd do 0–6 months" — sequenced by learning velocity, not excitement.
- No "insights" deck with a hundred bullet points. The brief fit in a 6-page memo.
What I'd do differently
- Run quant after qual. Interviews raised questions; a scorecard survey would verify them at 5× reach.
- Design the interview with scoring baked in. Asking "how bad is this, 1–10" per pain during the call beats post-hoc coding.
- Version the scoring rubric. The severity/ROI/buildability weights shifted mid-project; version the rubric so the brief is reproducible.
- Record the "no" conversations. The prospects who said "not interested" had the clearest signals about what would work.
Discovery's deliverable isn't what you heard. It's a ranking any skeptic can audit.
For the mechanics I keep alive, the deep-dive method covers evidence. The downstream scoring system this fed is in the Upcore lead-scoring case study.
Frequently asked questions
How many interviews do you need for product discovery?+
We did 20+ first-party, vertical-aware conversations; enough for pattern, not unanimity.
What separates good discovery from a spicy survey?+
Every interview ends as a signal, not a quote — scored for severity, ROI, and buildability.
How do you choose the top-3 use cases?+
Score each opportunity on severity, ROI, and buildability; the top three separate you from a dartboard.