# The AI MVP brief: one page that decides what you build first

*Problem, user, core job, what the model does and does not do, data, metrics, cut list, risks, budget. With a worked example.*

This is the document I ask every founder for before the first call. Fourteen fields, one page, plain sentences. If you can fill it in, a good team can estimate and start your AI MVP in days. If you can't fill a field, that field is your first piece of work — and it is far cheaper to discover that now than in week five of a build.

**Time:** 45 minutes · **Format:** 14 fields + worked example · **Online version:** https://stepikin.com/templates/ai-mvp-brief/

---

## How to use it

1. Write in plain sentences, not bullet fragments. "Users want AI" is not an answer; "Sales reps lose 40 minutes a day re-typing call notes into HubSpot" is.
2. Use a number wherever you can: minutes, money, volume, error rate. Guesses are fine — mark them with (est.).
3. Keep it to one page. If a field needs more than 3–4 sentences, you are describing the roadmap, not the MVP.
4. Share it with whoever builds the product. A good partner will push back on the cut list and the AI role — that is the point.

## Part A — Product

### Product in one sentence

*For [primary user] who [situation], [product] [does the core job], unlike [current alternative].*

> Your answer: 

### Problem and evidence

*What is painful today, for whom, how often, and what does it cost? What have you seen or heard that proves it?*

> Your answer: 

### Primary user (one) and other people involved

*Name one person you design for first. List others who touch the product (buyer, admin, reviewer) but do not design for them yet.*

> Your answer: 

### Core job

*When [situation], I want to [motivation], so I can [outcome]. Exactly one job for the MVP.*

> Your answer: 

### Current alternative and cost of the status quo

*What do people do today instead? Why is it not good enough? What does switching cost them?*

> Your answer: 

## Part B — The AI

### What the model does

*The specific transformation: input → output. Be concrete about format.*

> Your answer: 

### What the model does NOT do

*Draw the line explicitly. What stays deterministic code or a human decision?*

> Your answer: 

### Human in the loop and acceptable failure

*Where does a person review, edit or approve? What happens when the AI is wrong — and how wrong is acceptable?*

> Your answer: 

### Data sources and access

*Where does the input come from? Format, volume, permissions, personal data, retention.*

> Your answer: 

## Part C — Success and scope

### Success metrics

*One north-star for the MVP plus: activation, quality (eval), cost per task, and one guardrail that must not get worse.*

> Your answer: 

### Scope: must-haves (max 5) and the cut list

*Five must-haves for v1. Then an explicit list of what is NOT in v1 — this list is more important than the first one.*

> Your answer: 

### Riskiest assumption and how to test it in week 1

*What, if false, kills the product? How can you test it before building most of it?*

> Your answer: 

### Constraints

*Platforms, integrations, compliance (GDPR, HIPAA, SOC 2), languages, accessibility, existing tech you must reuse.*

> Your answer: 

### Budget, timeline and decision-maker

*Budget range, hard deadline and why it exists, who signs off, and how much of your time is available per week.*

> Your answer: 

---

## Worked example: CallNote

*CallNote is a fictional product used to show what good answers look like. Numbers are illustrative.*

**Product in one sentence.** For B2B sales reps who run 5–8 discovery calls a day, CallNote turns each call recording into a ready-to-approve CRM update in under a minute, unlike manual note-taking or generic transcription tools that still leave the CRM empty.

**Problem and evidence.** Reps spend 30–45 minutes a day writing call notes and updating deal fields. In 14 interviews, 11 reps admitted they skip updates on busy days; sales managers then forecast on stale data. One design-partner team (12 reps) measured 6.5 hours per rep per week on CRM admin.

**Primary user (one) and other people involved.** Primary: an account executive at a 10–50 person B2B SaaS sales team, uses HubSpot, on calls via Zoom or Google Meet. Also involved: the sales manager (buyer, reads the pipeline), RevOps admin (installs the integration).

**Core job.** When I finish a discovery call, I want the CRM to already reflect what was said, so I can move to the next call without admin work and my manager trusts the pipeline.

**Current alternative and cost of the status quo.** Manual notes in a doc, then copy into HubSpot; or Gong/Otter transcripts that nobody reads. Transcription exists — the gap is structured, field-level CRM updates the rep trusts enough to approve in one click.

**What the model does.** Input: call transcript + existing deal record. Output: a JSON update with 6 fields (next step, close date change, budget mentioned, decision makers, objections, competitor mentioned), each with the transcript quote it came from, plus a 5-line summary.

**What the model does NOT do.** Never writes to the CRM without the rep's approval. Does not change deal stage or amount (rep-only). Does not send emails to prospects. Does not score or rank reps. Pipeline math stays in code, not in the model.

**Human in the loop and acceptable failure.** Rep reviews a diff of proposed field changes and approves, edits or rejects each one. A wrong suggestion costs 5 seconds to reject; a wrong silent write would cost trust in the whole product — so there are no silent writes. Target: 80% of suggested fields accepted without edits.

**Data sources and access.** Zoom and Google Meet recordings via their APIs (OAuth per rep); HubSpot deals via private app token. ~6 calls/rep/day, 30–45 min each. Contains prospect names and emails (personal data) → EU data region, 30-day transcript retention, provider with zero data retention.

**Success metrics.** North-star: % of calls that end with an approved CRM update within 1 hour (target 60% by week 4 of pilot). Activation: rep approves first update on day 1. Quality: ≥ 80% of fields accepted unedited on a 50-call golden set. Cost: ≤ $0.15 per call. Guardrail: zero writes without approval.

**Scope: must-haves (max 5) and the cut list.** Must: Zoom + Meet import; HubSpot deal matching; 6-field suggestion with source quotes; approve/edit/reject UI; Slack ping when ready. Cut: Salesforce, Teams, mobile app, manager dashboards, coaching/scoring, multi-language, custom fields, email drafting. All are "later" — none blocks learning whether reps trust the updates.

**Riskiest assumption and how to test it in week 1.** Riskiest: reps will trust AI-written CRM fields enough to approve them without re-listening. Week-1 test: run 30 real transcripts through a prompt in a notebook, show the diffs to 5 reps in a Figma prototype, measure accept-without-edit rate. Below 60% → rethink the output before building integrations.

**Constraints.** Web app (desktop-first), Chrome. HubSpot only for v1. GDPR: EU hosting, DPA with model provider. English only. Team already uses Next.js and Postgres — reuse.

**Budget, timeline and decision-maker.** Budget $25–40k for MVP. Deadline: pilot with 3 design-partner teams by 15 January (their Q1 planning). Decision: CEO (me), weekly 60-min review + async in Slack within 24h. Success at the pilot unlocks the seed round conversation.

---

Made by [Stepikin Studio](https://stepikin.com/) — design and engineering for AI products, MVPs shipped in weeks.
Want this done for you? [Get an estimate](https://stepikin.com/estimate/) · [Book a call](https://stepikin.com/book/) · More templates: https://stepikin.com/templates/