How We Scope an AI Product in a Discovery Call

A serious AI discovery call is not a sales pitch — it is a structured audit of problem, data, risk, and measurable outcomes. Here is what we cover in sixty minutes and what founders should bring. Sixty minutes is enough if founders arrive prepared — unprepared discovery becomes sales theater that wastes both sides' time. We end discovery with open questions list — honest unknowns beat false certainty in timeline promises.

What founders should bring

Problem statement, example users, and any data samples or API docs. Known compliance constraints: HIPAA, PCI, DPDP. Competitive alternatives users employ today. Bring anonymized data samples — even ten documents reveal OCR needs, PII density, and format chaos. List systems integrations must touch: CRM, EHR, WhatsApp, payment gateway — each adds weeks. Ask who can kill the project internally — champion without executive sponsor stalls on data access.

A demo or wireframe helps — even rough. Budget and timeline bands, not false precision. Competitive alternatives users employ today define the improvement bar — not your feature wishlist. Budget bands honest about uncertainty beat false precision that forces vendors to pad quotes. Record discovery call with permission — quotes become acceptance criteria engineers reference later.

What we explore together

Job-to-be-done, success metric, and v1 anti-goals. Data availability, PII, and ingestion complexity. Architecture sketch: RAG vs agents vs fine-tune vs deterministic core. We sketch architecture live: where deterministic code ends, where RAG begins, where humans approve. Hallucination impact scoring — low, medium, catastrophic — drives eval investment level. Integration count drives timeline more than model choice — each OAuth and webhook adds test surface.

Risk register: hallucination impact, cost ceilings, human-in-the-loop needs. Ballpark phases: discovery, MVP, hardening. Cost ceilings and latency targets constrain model choices early — GPT-4 class models may be ruled out for high-volume classification. Phase ballparks include discovery, MVP, hardening — not one lump sum. Refusal and safety requirements for consumer-facing bots need explicit policy discussion, not assumptions.

Outputs and red flags

Written scope summary with assumptions and open questions. No fixed bid without data review — honest partners say so. Written scope lists assumptions explicitly: data available by week two, legal approval by week three, one design partner cohort. Fixed bids without data review are red flags from any vendor. Follow-up email within twenty-four hours confirms scope and next steps — async founders appreciate written truth.

Red flags on our side: vague metrics, unavailable data owners, expecting magic without eval. See pricing guide for how we translate scope into engagements. We say no to features that fail eval feasibility — shipping magic harms both sides. Pricing guide translates scope outputs into engagement structures. Compare our MVP playbook timeline against discovery outputs — flag if scope exceeds six-week thin slice.

What to do next

Start with one measurable outcome — latency, accuracy, adoption, or revenue — and design your stack around that constraint. GenOrbit ships production AI systems across healthcare, astrology, fintech, and SaaS; if you want a second pair of eyes on architecture or delivery, we are happy to talk.