idea-hunt — find a painful workflow AI can replace, then prove people will pay

Core thesis (read first)

Do not invent an idea. Find an existing, painful, already-paid-for workflow that AI can now do as completed work — and kill it fast if real money doesn't show up.

The best AI businesses in 2026 are not new problems. They are old problems whose cost structure just changed. The winning shape is Service-as-Software: instead of selling a tool that helps a human do the work (per-seat SaaS), you sell the finished outcome the human used to produce. Buyers are shifting from "we need 50 licenses" to "we need 5,000 tickets handled / 200 listings written / 40 intakes qualified." Vertical AI agents are eating horizontal SaaS precisely because they own the outcome, not the seat.

Someone is already paying — in money, hours, or dread — for a workflow AI now collapses. Your job is to locate that workflow with evidence and confirm willingness-to-pay, not imagine a product.

Three signals a workflow is a real target:

1. Money already flows (a painkiller, not a vitamin). People pay a tool, agency, VA, freelancer, or plugin to do this today. If nobody pays for today's bad solution, AI won't make them start. A business owner paying a freelancer $500/mo for a task you can automate for $50 has already told you the price.
2. The pain is repetitive, specific, and now. It recurs weekly/daily and has a narrow owner ("clinic front desks", "Shopify stores > $10k/mo", "immigration paralegals").
3. AI changes the unit economics for real. The core task is language, extraction, classification, drafting, routing, or judgment that LLMs now do reliably — not "AI-washed" decoration over work that's actually offline, physical, or human-mandated.

If a candidate fails any of the three, reject it and move on. Rigor is what you throw away.

The stat that governs this whole skill: 42% of startups fail because there is no market need (CB Insights) — the #1 cause, ahead of running out of cash. And 60–70% of AI "wrappers" earn $0, with ~90% expected dead by end of 2026. This skill exists to make you fail those ideas on paper in an afternoon instead of in the market over six months.

The pipeline (do not skip stages)

STAGE 0   Frame the hunt        → pick the search surface
STAGE 1   Pain mining           → last30days + review/gig-site complaints (real verbatims)
STAGE 2   Candidate slate       → 5–8 workflows, each an existing-solution-being-replaced
STAGE 3   Kill-gate scoring     → apply the 6 gates; most candidates die here
STAGE 4   Demand-reality test   → office-hours' six questions + Mom Test discipline
STAGE 4.5 PROOF OF WALLET       → design the pre-sell that proves willingness-to-pay
STAGE 5   Pick ONE              → single highest-EV wedge, justified against rejects
STAGE 6   Build blueprint       → only now, the full execution plan

Never jump straight to STAGE 6. The whole point is that most ideas die at STAGE 3–4.5.

STAGE 0 — Frame the hunt

Establish the search surface before mining. If the user gave a niche/workflow, use it. If open-ended, ask (via AskUserQuestion) for one anchor — enough to aim, not to stall:

• A domain they know (their job, an industry they've worked in, a hobby with money in it). Insider knowledge is the single biggest edge — it lets you see painful workflows and, later, gives you distribution into a niche horizontal players won't target.
• A customer they can reach in 72h (who do they already have access to?).
• A constraint (B2B vs B2C, budget, geography — e.g. India-first vs global).

Write the frame to docs/idea-hunt-<slug>.md (create docs/ if needed). This file is the running artifact for the whole hunt.

Bias: B2B and prosumer beats consumer for cold-start-in-72h. Businesses pay fast, pay monthly, and feel workflow pain acutely. Consumers churn and won't pay a stranger.

STAGE 1 — Pain mining (evidence, not imagination)

Goal: surface verbatim complaints and proof of spend about existing workflows/tools. You want people saying "I hate that I still do this manually", "X is too expensive/clunky", "I pay someone $__ to…". Founders who win mine complaints; they don't brainstorm.

Run these in parallel and log every hit into the artifact:

1. last30days on the domain and on incumbent tools — e.g. last30days pain points [workflow], last30days [incumbent tool] alternatives, last30days what [role] hate about their job. This is the freshness engine: what people complain about this month, with engagement signal.
2. Review-site feature-gap mining (highest-signal source). Read 1–3 star reviews on G2 / Capterra / App Store for incumbent tools. These are paying users telling you, unprompted, exactly where the current software fails. B2B categories (accounting, HR, legal-tech, healthcare admin) are especially rich.
3. Follow the money = willingness to pay. Search Upwork / Fiverr gigs for the task. A posted job is not a complaint — it is a purchase order: proof the pain is worth paying to fix right now. Note the going rate; that's your pricing anchor.
4. Community lurking — the subreddits / Slacks / Discords where the ICP already vents.

Capture per complaint cluster: the workflow, who owns it, what they pay today, the incumbent solution, and the verbatim pain. Quote real language — the copy for the eventual landing page is hiding in these quotes.

If STAGE 1 finds no one paying for the current bad solution and no gig/spend evidence, stop and re-frame (STAGE 0). No existing spend = no wedge.

STAGE 2 — Candidate slate (5–8 workflows)

Turn pain clusters into candidates. Each MUST be phrased as replacing a specific existing solution and selling the completed outcome — not a new invention:

"Replace [incumbent: agency / manual process / clunky tool / freelancer] for [narrow owner] with an AI system that delivers [the finished outcome]."

Bad candidate: "AI tool for marketing." (invented, vague, no incumbent, sells a tool) Good candidate: "Replace the $200–500/mo freelancer who writes + schedules Amazon listing copy for Indian D2C sellers with a self-serve AI that ingests the SKU sheet and delivers compliant, keyword-optimized listings." (incumbent = freelancer, owner = D2C seller, task = drafting, sells the outcome, money already flows)

List all candidates in the artifact before scoring. Resist committing to one yet.

STAGE 3 — Kill-gate scoring (most candidates die here)

Score every candidate against 6 gates. Any hard-fail = rejected. Do not rescue a favorite by fudging a gate.

Gate 6 — the durable moats that actually hold in 2026 (model access is NOT one of them):

• Workflow depth — you live inside the multi-step process, not a single prompt.
• Proprietary data flywheel — usage produces data a competitor can't get, making the product better over time (the model providers don't have the hospital's notes or the firm's case archive; you do because you sit in the workflow).
• Niche distribution — you own a vertical horizontal players won't bother targeting.
• Switching costs / integrations — you're wired into their systems of record.
• Compliance — auditable HIPAA/SOC2/GDPR handling is itself a moat.

Realistic caveat: for a solo cold-start, speed of execution + distribution into your niche is the day-1 "moat"; the data flywheel and switching costs are what you build toward. Favor permanence (being the obvious default in a niche) over a fragile technical moat.

Score each gate 0 (fail) / 1 (weak) / 2 (strong). Reject anything with a 0 or total < 8/12. Write the scored table into the artifact and explicitly list what you rejected and why — the rejections are as valuable as the pick.

Anti-slop check: "ChatGPT with a system prompt and a login" hard-fails gate 6 unless there's a data/workflow moat. A wrapper is a feature, not a business — 90% of them are dead by end-2026.

STAGE 4 — Demand-reality test (office-hours + Mom Test)

Take the 1–2 survivors and run them through office-hours (startup mode). Its six forcing questions are the real filter:

1. Demand reality — is anyone desperate for this, or is it merely reasonable?
2. Status quo — what do they do today, and why is it painful enough to switch?
3. Desperate specificity — who is the one customer type that would buy tomorrow?
4. Narrowest wedge — the smallest thing you can sell that still delivers value?
5. Observation — have you watched this pain happen, or are you guessing?
6. Future-fit — does it ride a trend (AI cost curve, regulation, platform shift)?

Apply Mom Test discipline to any user/customer conversation you plan:

• Ask about past behavior and actual spend, never about a hypothetical future ("Would you use this?" is worthless — people lie to be nice).
• Don't pitch the idea. The moment you pitch, honest exploration turns into polite praise.
• Good questions: "Walk me through the last time you did X." "What do you use today?" "What did that cost you / who do you pay?" "What happens if you don't fix it?"

A candidate that can't answer #1 and #5 with real evidence is not validated — send it back to STAGE 1 or kill it. office-hours saves a design doc; fold it into the artifact.

STAGE 4.5 — PROOF OF WALLET (the gate everyone skips — don't)

Opinions are not demand. Before writing production code, design the smallest test that extracts real willingness-to-pay. Pick the strongest one the situation allows:

• Pre-sale / deposit — take money (or a refundable deposit / paid pilot / annual-upfront founder's-rate) before the product exists. Cash is the only unfakeable signal.
• Smoke test — a landing page with a real price and a "Buy / Start" button that captures intent (and ideally a card). Measure click-to-checkout, not "signups."
• Concierge / manual-first — deliver the outcome by hand for the first 3–5 customers (you + AI in the loop) and charge for it. This proves demand AND writes your automation spec from real cases before you build software.
• Letter of intent — for higher-ACV B2B, a signed "we'll pay $X/mo when you ship Y."

Define the pass bar up front, e.g. "3 paying pilots or $X collected from 30 qualified conversations." Miss the bar → the wedge is wrong; return to STAGE 1. Hit it → proceed. This one stage is the direct antidote to the 42%-no-market-need failure mode.

STAGE 5 — Pick ONE

Choose the single highest expected-value wedge. State plainly:

• Why this one beats the other survivors (EV ≈ pain intensity × reachability × build-ability × moat × proven willingness-to-pay).
• The three closest alternatives and the specific reason each was rejected.
• The narrowest possible v1 — one customer type, one workflow, one outcome.

Do not present a menu. One recommendation, defended.

STAGE 6 — Build blueprint (only for the survivor)

Now — and only now — generate the execution plan. Decision-grade, not a 44-item wall.

1. Wedge & UVP — one sentence each. Lead with the outcome sold, not the feature list.
2. ICP & where to reach them in 72h — name the exact community / list / channel.
3. MVP scope — the thinnest thing that delivers the outcome. Concierge-first is allowed and often smarter (your STAGE 4.5 manual cases are the spec). Cut everything else.
4. Stack — pragmatic (Lovable/Bolt/Next + Supabase + Razorpay/Stripe + n8n + an LLM API). Use codex / rapid-prototyper for the build.
5. Trust system for a zero-reputation stranger — interactive demo, free trial OR money-back guarantee, transparent pricing, founder attention for the first customers, turn STAGE 4.5 pilots into the first testimonials/case studies.
6. Pricing — anchor to the human/incumbent cost you replace, not to a race-to-$9:
   • Price against what they pay today (the $500/mo freelancer, the agency retainer, the salary). Salesforce's Agentforce prices at ~$2/conversation explicitly vs the $30–50 human cost — copy that logic.
   • Prefer outcome- or usage-based, or hybrid (a predictable base + per-outcome/credit charge). Hybrid is the dominant 2026 structure; pure per-seat is dying for AI products because it charges most exactly when the product deflects the most work.
   • Keep it recurring and high-margin.
7. Go-to-market, honestly sequenced (first customers almost never come from ads/SEO):
   • Hours 0–72: hand-to-hand. Direct outreach to 50–100 exact-ICP people with a 3-sentence message naming their specific pain (expect ~15–25% reply when the pain is real). Community selling (give value ~2 weeks before pitching). A launch-platform / directory listing (Product Hunt). This is where the first rupee comes from.
   • After first signal: Meta/Google ads to scale what already converts. Warn the user: ad accounts, pixels, and review cycles rarely produce paying customers inside 72h from a cold start — treat paid as the week-2 lever, not the day-1 one.
8. Automation architecture — what runs without the founder (onboarding, fulfillment, billing, support, follow-up). Map each to a tool. The goal is ≥95%-automated over time.
9. Metrics & kill-criteria for the business itself — define up front the signal that says "double down" vs "kill it" (e.g. "0 paying customers after 30 qualified conversations = wrong wedge, restart at STAGE 1").
10. Scaling roadmap — first paying customer → $1k MRR → $10k MRR → $100k MRR, naming the constraint that changes at each step (and where the data flywheel / switching-cost moat starts to compound).

Operating rules

• Evidence over eloquence. Every demand claim traces to a STAGE 1 verbatim, a paid gig, or a STAGE 4.5 dollar. No unsupported "people would love this."
• Cash > opinions. A landing-page signup is interest; a deposit is demand. Weight them accordingly.
• Reject freely. Killing 7 candidates and shipping 1 paid-validated wedge is success. Rationalizing the first idea is failure.
• Sell the outcome, not the seat. If it can't be framed as completed work, rethink it.
• Honest timelines. Don't promise ads-from-zero revenue in 72h. Name the real cold-start channel (hand-to-hand).
• Productize or don't recommend. If it can't become ≥95%-automated software, it fails unless the user explicitly wants a service — and even then, concierge is a bridge to software, not the destination.
• One artifact. Keep everything in docs/idea-hunt-<slug>.md so the hunt is resumable and auditable.

Companion skills

• last30days — STAGE 1 pain mining. Real complaints, last 30 days, with engagement signal.
• office-hours — STAGE 4 demand-reality gate. Run survivors through the six questions.
• After a wedge clears STAGE 4.5, hand off to codex / rapid-prototyper for the STAGE 6 build.