Financial Projections for AI-First Startups (The Three Load-Bearing Assumptions)

The problem

An AI-first SaaS forecast at the seed stage has three load-bearing assumptions: CAC, gross margin, and the ratio between paid acquisition and product-led growth. Get those three within ±20% of reality and the model will hold up under investor scrutiny. Get any one of them wrong by 2× and the rest of the spreadsheet is theater.

Most templates founders inherit from accelerators were built between 2018 and 2022. They assume $50–$500 CAC, 70–85% gross margins, 2–4% conversion, and a 12-to-18-month CAC payback. Those numbers were tight against reality for traditional SaaS. They are not tight against reality for AI-first distribution — they are wrong in the same direction, by an order of magnitude, on each metric. A model built on the old assumptions tells you the wrong story about runway, the wrong story about when to raise, and the wrong story about what a healthy month looks like.

This walks through the structural inversions, the unit-economics rebuild, and what the load-bearing assumptions should look like for a product whose primary distribution channel is an MCP server rather than a Google Ads account.

What most people get wrong

Mistake one: applying a traditional SaaS template to AI-first economics. The headline metrics invert. Traditional SaaS CAC ranges $50–$500 per acquired customer. AI-first CAC, when distribution runs through MCP and A2A protocols, ranges $0.10–$1.00 per user — three orders of magnitude lower. Conversion in the traditional funnel is 2–4% from page-visit to paid; conversion inside an AI assistant (where the model has already pre-qualified intent before invoking your tool) runs 15–25%. CAC payback in traditional SaaS is 12–18 months; for an AI-first platform business with edge-deployed infrastructure, it can be 1–2 months.

The cost structure inverts to match. Traditional SaaS spends 30–50% of revenue on sales and marketing. AI-first platform businesses spend almost nothing on customer acquisition, because protocol integration is acquisition. The same engineering investment that builds your MCP server, your AP2 mandate handler, and your ACP checkout endpoint is what makes you discoverable to every AI assistant in the world. There is no separate marketing line item — building product and acquiring customers are the same activity.

A template that doesn’t account for this inversion will tell you to budget six months of seed runway for paid acquisition that you do not need, then tell you your gross margins are unacceptable because it expects the wrong cost ratio, then tell you your forecast looks “too optimistic” because the benchmark it compares against is fifteen years old.

Mistake two: conflating “AI startup” with “AI-first platform.” The economics differ by an order of magnitude between two business shapes that get conflated in pitch decks. An AI model provider — the team running its own LLM inference — faces 25–65% gross margins because GPU compute eats the line. An AI application built as a thin wrapper around someone else’s model faces 40–60% gross margins because API calls to the model provider eat the line. An AI-first platform that orchestrates agents rather than running models faces 85–95% gross margins because the heavy compute happens on the agent’s infrastructure, not yours.

The book’s three-bucket table is reproduced in the working-approach section below. Your forecast lives or dies on which bucket you are in — and many founders quote “85% gross margins” while building bucket two, which means the model is wrong from the first row.

A working approach

The unit-economics calculator structure the book uses:

const unitEconomics = {
  // Customer Acquisition (AI-first)
  cac: {
    protocol_integration_cost: 1800, // one-time build
    users_from_integration: 10000,   // Year 1 projection
    cac_per_user: 0.18,              // $1,800 / 10,000
  },

  // Lifetime Value
  ltv: {
    arpu_monthly: 12.50,         // blended across tiers
    avg_lifetime_months: 24,     // 97% annual retention
    ltv_per_user: 300,           // $12.50 × 24
  },

  // Efficiency
  ltv_to_cac: 1666,              // 300 / 0.18 — three orders above SaaS norm
  payback_months: 0.014,         // immediate
};

The 1,666:1 LTV:CAC ratio looks absurd against the traditional 3:1 SaaS benchmark. That is the point. The number is not aspirational — it is what happens structurally when CAC drops from $50 to $0.18 and lifetime value holds.

The three-bucket gross-margin table that decides which row of your P&L tells the truth:

Business Type	Cost Driver	Gross Margin
AI Model Provider	GPU compute, training data	25–65%
AI Application (wrapper)	API calls to model providers	40–60%
AI Platform / Orchestrator	Edge compute, protocol fees	85–95%

The book is built around the platform/orchestrator bucket. Every projection assumes that the heavy AI compute happens elsewhere — inside Claude, ChatGPT, Gemini — and that your platform’s job is to expose tools, accept AP2 mandates, complete ACP checkouts, and route the result. On Cloudflare Workers at the edge, compute per request is fractions of a cent.

The three load-bearing assumptions to get within ±20%:

1. CAC — total cost of protocol integration, divided by users acquired through those integrations in Year 1. If you cannot defend the denominator, you cannot defend CAC. The book covers how to attribute AI-discovered users using MCP server logs and AP2 mandate IDs.
2. Gross margin — calculated from edge-compute cost per request plus payment-processor fees plus database egress. If you are paying GPU costs, you are in bucket one, and the spreadsheet template the book provides will tell you which line items to revise.
3. The paid vs. PLG ratio — what percentage of users arrive through paid channels (which still applies to some AI-first models) versus AI-channel organic discovery. The mix determines the shape of your S&M line and the realism of your CAC payback.

The book walks the month-by-month projection from $0 to $100K ARR (foundation phase), $100K to $500K (growth-lever phase), and $500K to $1M (scale-economics phase) — with the unit economics, the cost structure, and the milestone definitions for each band. The financial-model spreadsheet in the bonus folder is the same model you can open and customize.

Where this scales

The article above is the unit-economics frame. The full book extends in four directions:

• The 12-metric investor pack — the metrics VCs actually evaluate (not the 50-metric template; the 12 that decide the conversation), with definitions and per-metric calculation formulas.
• The five revenue streams — transaction commissions, subscriptions, API/developer access, B2B trust services, digital products — each modeled with the price tiers and conversion benchmarks the book uses for its own projections.
• The 12-month milestone framework — month-by-month revenue targets, user counts, and the trigger events that say “the model is on track” versus “the assumption needs revisiting.”
• The fundraising readiness package — pitch-deck financial slides, valuation framework for AI-first startups, the red flags that kill AI deals before they reach term-sheet, and a due-diligence checklist.

The bonus financial-model CSV is the spreadsheet behind every chart in the book. Open it, replace the assumptions with yours, and the formulas recalculate the full $0-to-$1M arc.

Included with the book

• README.md — bonus-folder index explaining what each artifact is and how to apply it.
• README.pdf — same index, rendered for offline reading.
• financial-model.csv — the full unit-economics and revenue projection model. Drop into a spreadsheet, edit the assumption cells, watch the rest recalculate.

Get the full picture

Readers of this also chose

Questions readers ask

Is this only for AI-first businesses, or does it apply to traditional SaaS too?

The book is explicitly for AI-first models. The unit-economics inversions described here do not apply to a traditional SaaS that runs a Google Ads campaign and a sales team. The forecasting structure (CAC, LTV, gross margin, growth-lever sequencing) is universal; the calibrated values are AI-first.

What if I'm an AI application (wrapper), not a platform?

The book includes the bucket breakdown explicitly so you can locate yourself. The unit-economics calculator and revenue projection framework apply across all three buckets — the gross-margin row and the CAC ratio are what change. The book covers what the model looks like for each bucket so you don't apply the wrong benchmarks.

Will the spreadsheet work in Google Sheets / Excel / Numbers?

The model is a plain CSV with formulas in the columns. It opens cleanly in Google Sheets and Excel. Numbers reads it but recalculates some formula syntax — minor manual fix for two cells.

How current are the SaaS benchmarks?

The benchmarks are calibrated to 2025–2026 data: median SaaS growth ~26% annually, CAC inflation +14% in 2024, AI-first CAC at $0.10–$1.00. The book is explicit about which numbers are observed-now versus forward-projection.

What's the refund policy?

Lemon Squeezy's standard refund window applies. If the model doesn't fit your business shape, the refund link is in the receipt email.