The AI Startup Playbook

Most AI startup advice is either too vague to be actionable or too specific to generalize. “Find product-market fit” is not advice. “Here’s exactly what to do in week two” is. This guide is the latter.

Before the breakdown, the conclusion: there are six principles that govern every AI startup that makes it past the first year. Read these first. Everything else in this article is in service of them.

The Six Principles

1. The problem comes first, the AI comes second. The startups that survive are the ones that started with a problem so painful that users would accept an ugly, unreliable, half-broken solution — and then used AI to make the solution dramatically better. The startups that die are the ones that started with AI and went looking for a problem.

2. Speed of learning beats speed of building. In the first six months, your only job is to learn what your users actually want. Building fast matters. But building in the wrong direction at high speed is just a faster way to fail.

3. Charge from day one. Not “plan to charge eventually.” Not “free during beta.” Charge money from the first user. Willingness to pay is the only real signal that you’ve solved a real problem.

4. Your moat is not the model. GPT-5.5, Claude Sonnet, Gemini 3.5 — these are commodities. Any moat built entirely on model capability will erode. The durable advantages are proprietary data, workflow lock-in, network effects, and operational complexity that is expensive to replicate.

5. Most AI startups undercharge. The natural instinct is to price low to drive adoption. This is almost always wrong. AI products that provide genuine value should command premium prices. If you’re worried about price, your real problem is that you haven’t demonstrated enough value yet — and lowering the price won’t fix that.

6. Defensibility is a deadline, not a milestone. The window to build a moat is short. The incumbents are watching. The copycats are fast. You need to be moving toward a defensible position from day one, not waiting until you have revenue to think about it.

Stage 1: Ideation — Finding a Real Problem

The standard advice is “find a painful problem.” That’s true but useless without a method.

The best AI startup ideas come from one of three sources: your own deep professional experience in a domain, direct conversations with people in a specific industry who are doing work that is obviously miserable, or pattern-matching on repetitive high-stakes work that currently requires expensive human judgment.

What to look for: Tasks where the output is valuable, the input is variable, and the current solution is either impossibly expensive (human experts), unreliably slow (manual processes), or completely inaccessible to people who need it. Healthcare documentation. Legal contract review. Financial analysis for mid-market companies. Customer support at scale. Security monitoring. These categories have been picked over, but most industries have equivalent problems that haven’t been touched yet.

What to avoid: Problems where the pain is mild, the existing solution is “good enough,” or where the decision-maker is not the one experiencing the pain. Anything where your customer is a large enterprise but your champion is a middle manager with no budget — that sales cycle will kill you before you find product-market fit.

The right question to ask: “Who would pay $500/month right now for a solution to this problem, and do I have access to them?” If you can’t answer that concretely, keep looking.

Do not start building. Not yet.

Stage 2: Validation — The Two-Week Test

Before you write a line of code, you have two weeks to validate the idea. This is not optional.

Week one: talk to ten potential customers. Not “would you use this?” conversations — those are worthless. Structured discovery calls where you understand how they currently solve the problem, what it costs them in time and money, what’s broken about the current approach, and what a solved version of this would be worth to them. If you can’t get ten people to take a thirty-minute call, you don’t have a market. Stop.

What you’re looking for: Pain that’s specific, measurable, and current. “We spend fifteen hours a week on this” is a signal. “Yeah it’s annoying sometimes” is not. You want to hear the same specific problem described the same specific way by multiple people who don’t know each other.

Week two: build the cheapest possible demonstration of the solution. This does not have to be software. It can be a Notion doc, a Google Form that feeds into a spreadsheet, you manually doing the AI-assisted work and emailing results back. The goal is to show someone the output and get them to commit money to receive more of it. An LOI, a prepaid invoice, a signed contract — anything that has real cost to them.

If you get through two weeks and have at least three people who have expressed willingness to pay in a concrete way, you have enough signal to build. If you don’t, pivot the framing or the target customer — not the core idea yet — and run the test again.

Stage 3: MVP — The Smallest Thing That Proves the Core Value

The mistake founders make is building an MVP that showcases AI. The right MVP showcases the value delivered to the user — and it often doesn’t need AI to do that.

If your product is “AI-powered contract review,” your MVP might be: you manually reviewing contracts using a combination of AI tools and human judgment, delivering a structured report within 24 hours, and charging $200 per contract. No product. No automation pipeline. No dashboard. Just the output, delivered reliably.

This accomplishes several things: you learn what the output actually needs to look like, you learn what questions customers care about, you identify the parts of the workflow that are genuinely hard, and you start generating revenue immediately.

What the MVP must do: Deliver the core value. Demonstrate that users will pay for it. Teach you what to build next.

What the MVP does not need: A polished UI. Automation. Scale. Multi-user support. Integrations. Analytics. Billing infrastructure beyond a Stripe link.

Build the least impressive thing that still delivers the value. Then start replacing the manual parts with software incrementally, prioritizing whatever is taking you the most time.

Stage 4: Stack Selection — Speed vs. Cost vs. Lock-in

By the time you’re building the real product, you need to make stack decisions. Here’s how to think about them without overthinking them.

Model selection: Start with the best model available (currently GPT-5.5 or Claude Sonnet 4.6) for your main value-generating workflow. Do not optimize for cost at this stage. Inference cost as a percentage of revenue will look enormous at small scale and trivial at scale. Get the product right first, then optimize the cost.

Frameworks: For most AI startups, the fastest path is: Next.js or a Python FastAPI backend, Vercel AI SDK or LangChain for orchestration (with the understanding that you may replace LangChain as you mature), Supabase or Postgres for persistence, and whatever cloud provider your team knows. Do not build a custom LLM orchestration framework. Do not use five different vector databases. Do not chase the new framework that launched last week.

Hosting: Vercel or Railway for fast iteration. AWS or GCP when you have specific needs that justify the operational complexity — which is not when you’re at $0 MRR. You will not need Kubernetes until you have a real infrastructure problem, and you won’t have that problem before you have revenue.

Lock-in tradeoffs: Model lock-in is real but overstated. Abstract your model calls behind a thin interface layer from the start (it takes two hours) and you can swap providers. Vendor lock-in on infrastructure is more dangerous — avoid building on proprietary managed services for anything business-critical until you understand the pricing at scale.

The actual advice: Pick boring, well-documented technology. Your competitive advantage is your product, not your infrastructure choices. The stack doesn’t matter nearly as much as your speed of iteration.

Stage 5: First Users — Getting Them and Learning From Them

Your first ten customers are not a distribution challenge. They’re a sales challenge that you solve manually.

How to get them: Go direct. LinkedIn messages to the specific job titles who have the problem you’re solving. Cold email with a specific, personalized pitch. Warm intros from your network. Posting in communities where your target users are already discussing the problem. You are not running paid ads. You are not doing content marketing. You are doing manual, one-to-one outreach until you have enough signal to know what works.

What good onboarding looks like: You personally onboard every early customer. You watch them use the product over video call if at all possible. You are available for questions in Slack. You follow up within 24 hours after their first use to ask what confused them and what they wanted to do that they couldn’t.

What to learn: Three things matter at this stage. First: do they actually use it after the first session? Second: can they articulate the value they’re getting in concrete terms? Third: would they be upset if it went away? If the answer to all three is yes, you have something. If the answer to any of them is no, that’s your next problem to solve.

Do not: Hire a salesperson yet. Do not build features based on what early users ask for without understanding why they’re asking. “Can you add an export to Excel” might mean “the output format isn’t working for my workflow” — solve the workflow problem, not the feature request.

Stage 6: Pricing — Charge Early, Charge More

The default instinct of technical founders is to underprice. This is almost universally a mistake.

Why to charge from day one: Free users are not customers. They generate noise, not signal. When you start charging, users who pay tell you something real: they believe the product is worth money. Users who don’t pay tell you something real too: the value isn’t clear enough or compelling enough yet. Both signals are more useful than free users who kind of like the product.

How to set the price: Start at whatever number makes you slightly uncomfortable. Then raise it. Most AI products that deliver real value to business users should be priced at $100-500/month at minimum. Products that replace significant human labor should be priced at 10-30% of what that labor would cost. Do not price based on your inference costs. Price based on the value you deliver.

What not to do: Don’t offer a free tier until you understand your activation and retention mechanics well enough to know it won’t cannibalize your paid conversion. Don’t offer per-seat pricing that makes your revenue grow slower than your customer’s value capture. Don’t let early customers negotiate your price down because you’re afraid of losing them — a customer who won’t pay your actual price is not a real customer.

Annual plans: Offer annual pricing at a 15-20% discount from day one. Cash flow is real. Retention from annual commitments is real. Founders consistently leave money on the table by not offering this early.

Stage 7: Product-Market Fit — What It Looks Like for AI Products

PMF for AI products has a different signature than traditional SaaS. Here’s what to look for.

The standard indicators apply: MRR growing month-over-month without unusual promotional effort. Churn below 5% monthly for self-serve, below 1% monthly for enterprise. Users who actively evangelize the product to colleagues. NPS above 40.

The AI-specific indicators: Users who adjust their workflow around your product rather than using it occasionally. Users who notice and complain when the AI output quality drops even slightly. Users who share outputs from your product externally, making it part of their professional work product. Users who ask for more AI-generated content rather than asking to turn AI features off.

What PMF is not: A moment. It’s not a point in time when everything clicks. It’s a zone you enter when the core value hypothesis is correct, which you then spend months reinforcing with product work. You will feel it as a shift in the quality of the problems you’re solving: you stop asking “will anyone pay for this” and start asking “how do we keep up with demand.”

For AI products specifically: Watch for the moment when users stop being impressed by the AI and start being frustrated when it’s wrong. That’s the transition from novelty to utility. That’s when you have something.

Stage 8: Scaling — $0 to $10k, $10k to $100k MRR

These are two different companies with two different problems.

$0 to $10k MRR

Your problems are: finding and closing customers manually, keeping early customers happy enough to stay, and learning what the actual product needs to be. Your operational structure should be: founders doing sales, founders doing customer success, one or two engineers building what customers need next.

At this stage, you are not optimizing infrastructure. You are not building internal tooling. You are not hiring specialists. Every hour goes to talking to customers and shipping product. The goal is to reach $10k MRR with a cohort of customers who are genuinely successful with the product.

The technical changes that matter: Observability into AI outputs so you know when the model is failing. A way to collect and act on user feedback quickly. Reliability — the product needs to work consistently, because at this stage a single bad experience with a customer is a significant percentage of your revenue at risk.

$10k to $100k MRR

Now you have different problems: you can’t close every deal yourself, you can’t personally onboard every customer, and the product needs to work without you watching it.

Sales: Hire your first sales hire at around $15-20k MRR, but only if you have a repeatable sales process that you’ve run yourself enough times to document. If you can’t write down what you say to convert a customer, you’re not ready to hand that off.

Customer success: At $30k+ MRR with enterprise customers, you need someone accountable for retention. This is often a founder’s role until you can hire specifically for it — don’t hire a generic “CS manager” too early.

Technical: At $50k+ MRR, inference costs become a real budget line. Invest in prompt optimization, caching, model routing (use cheaper models for simpler tasks), and output quality monitoring. At $100k MRR, you should have a clear picture of your gross margin per customer and a roadmap for improving it.

Stage 9: Defensibility — Building a Moat Before the Window Closes

Every successful AI startup has a window of roughly 12-18 months from their first real traction to the point where the incumbents or well-funded competitors notice and start competing directly. You need to build something defensible in that window.

Proprietary data: If your product generates valuable data about user behavior, outcomes, or domain-specific patterns, that data becomes an asset. Fine-tuned models trained on that data can outperform general-purpose models on your specific task. Build the data collection infrastructure early, even if you’re not using it yet.

Workflow lock-in: Integrations that touch users’ existing tools (Salesforce, Slack, GitHub, EHR systems) create switching costs that pure-play AI cannot replicate quickly. The deeper you are in someone’s workflow, the harder you are to displace.

Network effects: These are rare in B2B AI but worth pursuing when possible. If your product gets better as more users use it (through collective learning, shared benchmarks, community-generated content), that’s compounding value that competitors can’t replicate by copying your product.

Operational complexity: Some of the most defensible AI startups are ones that combine AI with genuine operational expertise — companies that don’t just provide software but provide a service layer that requires deep domain knowledge to execute. This isn’t glamorous, but it works.

The wrong moats: Being first. Having better prompts. Using a more advanced model. Moving fast. None of these are durable.

Stage 10: Team — Who to Hire First, What to Outsource

Most AI startups hire wrong. They hire for scale before they have scale, for specialization before they have clear needs, and for perceived legitimacy before they have product-market fit.

The founding team: You need at least one person who can build fast and one person who can sell. These can be the same person in exceptional cases. You do not need a COO, a VP of Marketing, or a Chief AI Officer as your first hires.

First hire: At $10k+ MRR, hire someone who is better than you at the thing you’re worst at. If you’re a technical founder who can’t sell, hire a revenue-focused generalist. If you’re a sales-focused founder with a technical co-founder who’s underwater, hire an engineer who owns infrastructure. Make them the second-best person at their function in the company, not the first — you should still be able to evaluate their work.

What to outsource: Design (use contract designers until you have a product that justifies a full-time designer), legal (use a startup-focused law firm on retainer), accounting (use a fractional CFO until you have revenue complexity that justifies full-time), and anything that is not directly related to building or selling the product.

What never to outsource: Customer discovery. Sales until you have a documented, repeatable process. Product decisions. These are the core competencies of a startup, and no contractor or agency will care about them as much as you do.

What to avoid: Hiring ahead of revenue. Hiring for title rather than output. Hiring friends and former colleagues out of comfort rather than fit. Every bad hire at this stage is a 6-12 month setback.

The Through-Line

Read back through everything above and you’ll see the same few ideas repeating: talk to customers before building, charge real money early, build something that can’t be replicated by copying your model calls, learn faster than you build, and stay paranoid about defensibility.

These are not novel ideas. They’re not AI-specific ideas. But AI makes it easier than ever to build impressive demos that fool you into thinking you’ve solved a problem before you’ve actually validated that anyone cares enough to pay for it.

The AI startup graveyard is full of companies with great demos, solid technology, and zero paying customers. Don’t be one of them. Find the problem, validate the pain, charge from day one, and build the moat while you still have time.

The playbook is simple. The execution is what separates the companies that make it from the ones that don’t.