Case study · AI Startup (usage-based)
Promptly AI has the growth investors love and the margins they question
Promptly AI sells an API that drafts, checks and routes customer messages for software companies. Customers pay a monthly platform fee plus a price for every thousand requests. Investors loved the growth. Their first question was always the same: what happens to gross margin when usage grows?
- Starting point
- 25 paying accounts, about $300 platform fee plus usage, $300k in the bank, 3 founders and 2 engineers
- Seed round
- $2M
- at $12M pre-money
- Series A (plan)
- $7M
- at $40M pre-money
What this case teaches
- Gross margin for AI products must include inference costs, and both prices and costs change over time.
- Usage growth per account is the biggest lever, so it needs a decay curve, not a constant.
- How to present a margin that improves from 56% to 66% without overpromising.
- How a three-founder cap table dilutes through SAFEs, a seed and a Series A.
The starting point
Promptly had strong early demand, a usage-based price list and costs that rose with every request.
Each account paid a $300 platform fee and about $4 for every thousand requests, using about 200 thousand requests a month at the start and more each month. Every thousand requests cost Promptly about $1.80 in model and GPU charges.
The founders' first deck said '85% gross margins, like SaaS'. Their first serious investor meeting ended on that slide.
What the founders brought to the model
| Input | Value | Source |
|---|---|---|
| Paying accounts | 25 | Billing system |
| Platform fee | $300 / month | Price list |
| Usage per account | 200k requests / month | Usage logs, trailing 3 months |
| Price per 1k requests | $4.00 | Price list |
| Inference cost per 1k requests | $1.80 | Model provider invoices |
| Monthly account churn | 3.5% | Billing cohorts |
| Paid CAC | $3,500 | Paid campaigns + founder-led sales time |
| Cash in bank | $300k | Bank statement |
The first draft, and what the Self-Audit caught
The founders' first version looked great and was wrong in three ways. Each one would have been found in diligence.
| Mistake | First draft | Fixed version | Impact |
|---|---|---|---|
| SaaS gross margin | 85% gross margin, ignoring compute. | Inference cost per unit modelled as a variable cost: 56% in year 1, 66% in year 5. | Gross profit fell by about a third in year 1 and the pitch became credible. |
| Usage growth forever | 6% monthly usage growth per account for five years. | A decay from 2.5% to 0.6% a month. | Year-5 revenue roughly halved; the plan stopped looking like a hockey stick drawn by hand. |
| Flat unit price | $4.00 per thousand requests for five years. | Price falling to $2.20 while cost falls to $0.80. | Margin still improves, but for a reason investors can test: cost falling faster than price. |
The lesson
AI investors have seen many models that ignore compute. Showing the inference cost line, and the assumptions behind its decline, is what earns trust.
The plan, after the fixes
ARR grows from $1.2M at the end of year 1 to $5.5M at the end of year 3. Gross margin improves every year as inference cost falls faster than price. EBITDA turns positive in month 47: later than a SaaS company, because the plan keeps investing in ML engineering.
Gross margin moves from 56% in year 1 to 66% in year 5, and the team grows from 9 to 30 people.
Sizing the raise
The founders sized the round from the plan, then priced it against AI market comparables. The Raise Planner says about $1.3M is the minimum to month 21. Promptly raised $2.0M at a $12M pre-money valuation. The extra covered two things the minimum does not: prepaid compute commitments that lower unit cost, and SOC 2 certification that enterprise buyers require.
Where the money goes
| Category | Share | Buys | Unlocks |
|---|---|---|---|
| Product & ML engineering | 42% | 2 ML engineers, 1 product designer | Self-serve API, eval suite |
| Compute & inference credits | 10% | Model API and GPU costs ahead of revenue | Gross margin above 60% |
| Sales | 14% | Founding AE + solutions engineer | 40 paying accounts |
| Marketing & developer relations | 12% | Docs, community, launch campaigns | 2,000 developer sign-ups |
| G&A & Operations | 8% | Finance, legal, data-protection reviews | Clean diligence |
| Security & compliance | 7% | SOC 2 Type I, penetration test | Enterprise-ready |
| Contingency buffer | 5% | Compute price shocks, slower sales | 2+ months extra runway |
| Recruiting & hiring | 2% | Recruiter fees for ML hires | Team of 14 by month 18 |
Cap table and ownership
Three founders hold 9M shares. Before the seed, Promptly had two post-money SAFEs ($400k at an $8M cap and $250k at a $10M cap) and a 1M-option pool. With a 12% target pool, the seed adds about 980k options before the new money comes in. That option pool shuffle is the biggest single source of founder dilution in this round, and it was the founders' main negotiation point.
What the founders would tell you
- Put compute in cost of revenue from day one, and show how it changes.
- Use a decay curve for usage growth.
- Explain margin improvement with a mechanism investors can test.
- Know which assumption moves your outcome most, and make it your first KPI.
Not covered on this page
The full case also covers the Health Score review, the downside plan, the financial slides they built, the investor questions they rehearsed and their monthly routine after closing.