Agent Cost Per Validated Customer: A Real Walkthrough

The phrase "cost per acquired customer" is misleading for AI products. An acquired user who churns before paying anything has cost you tokens, infra, and time, but produced no revenue. The metric that matters is cost per validated customer: the user who activated and stuck around long enough to plausibly pay. This walkthrough runs a real scenario through the Agent Cost Per Validated Customer calculator and shows where every dollar lands.

1. The scenario: 4,200 MAU coding agent

The product is a single-founder coding agent SaaS at 4,200 monthly active users. The model layer is Claude Sonnet at the published Anthropic rate of $3 per million input tokens and $15 per million output tokens^[1]. Each active user consumes roughly 18,000 tokens per month, split evenly between input and output. Non-model infrastructure (Postgres, edge compute, vector store, observability) lands at $0.12 per user per month. List price is $19 per seat per month.

Activation rate runs at 32%. That is the share of monthly actives who complete the product's defined first-value action within seven days of sign-up. Mixpanel's product benchmarks place the median activation rate at roughly 34% for B2B SaaS, with the top quartile above 50%^[3]. A 32% activation rate puts this product slightly below median, in the band where most pre-product-market-fit AI tools live.

30-day retention sits at 55%. That figure comes from ChartMogul's cohort data, where median 30-day retention for B2B SaaS clusters between 50% and 65% depending on segment^[2]. A 55% retention rate is unremarkable but real. Combined with 32% activation, the product converts a monthly active user into a validated customer with probability 0.32 × 0.55 = 0.176.

Running the full scenario through the calculator returns:

# agent-cost-per-validated-customer (computed live from /engines/agent-cost-per-validated-customer.js)
Engine input
  monthly_active_users  = 4200
  activation_rate       = 32
  retention_30d         = 55
  tokens_per_user_per_month= 18000
  model_input_price_per_mtok= 3
  model_output_price_per_mtok= 15
  infra_cost_per_user_per_month= 0.12
  price_per_user_per_month= 19

Engine output
  totalMonthlyAiCost    = 680.4
  totalMonthlyInfraCost = 504
  totalMonthlyCost      = 1184.4
  validatedCustomers    = 739
  costPerValidatedCustomer= 1.6
  monthlyRevenueAtPrice = 79800
  grossMarginAtPrice    = 78615.6
  grossMarginPercentAtPrice= 98.52

2. Token math: where the $680 comes from

Token spend per user works out to 9,000 input tokens and 9,000 output tokens at the assumed even split. At $3 per million input and $15 per million output, that is $0.027 input plus $0.135 output, totalling $0.162 of model cost per user per month. Multiplied across 4,200 MAUs, model cost lands at $680.40.

Per user per month:
  9,000 input × $3 / 1,000,000      = $0.027
  9,000 output × $15 / 1,000,000    = $0.135
  Token cost per user               = $0.162

Across 4,200 MAU:
  4,200 × $0.162                    = $680.40
  Infra (4,200 × $0.12)             = $504.00
  Total monthly cost                = $1,184.40

Two observations matter for solo founders. First, output tokens are dominant. At a 5x output premium, output tokens drive 83% of model spend even at an even token split. Halving the output ratio (write more concise responses, summarize tool output, strip code blocks the user did not ask for) is the largest cost lever before negotiating pricing. Second, infra ($504) is close to model spend ($680.40) at this scale. The reflex to focus on token cost while ignoring infra is wrong below 20,000 MAU; both lines are roughly equal in this band.

3. Validated customers: 739 of 4,200

Validated customers come from the activation-then-retention funnel: 4,200 × 0.32 × 0.55 = 739.2, rounded to 739 in the engine output. That is the user count to which the cost gets allocated. Not the 4,200 MAUs the platform served. Not the 1,344 activated users the product showed first value to. The 739 who activated and stuck.

The funnel decay matters. If you allocate the $1,184.40 monthly cost across 4,200 MAUs, the headline number is $0.28 per user — a flattering misread that hides where customers actually come from. Allocate across 739 validated customers and the honest number is $1.60. The 4x gap is the activation-times-retention multiplier (0.176 in this case), and it dominates almost every other variable in the engine. The CAC Calculator handles the headline acquisition spend separately from product economics; this engine handles the post-signup economics.

If activation moved from 32% to 40%, validated customers would rise to 924, and cost per validated customer would fall to $1.28. If retention moved from 55% to 65%, validated customers would rise to 874, and cost per validated customer would fall to $1.35. Both are larger swings than dropping Sonnet's output price by 30%.

4. $1.60 cost per validated customer

The $1.60 figure is total monthly cost divided by validated customers: $1,184.40 / 739 = $1.6027, rounded to $1.60 in the engine. That is the monthly product cost of carrying one user who is plausibly worth paying for. It does not include acquisition spend (paid ads, content production, partnerships), it does not include support time, and it does not include payment-processor fees. It is purely the cost of running the AI product surface itself for a validated user.

Compare against price. At $19 per user per month and a $1.60 cost per validated customer, the gross margin per validated user is $17.40, or 91.6% on that cohort. The engine reports a 98.52% gross margin against priced revenue, which is the right number for unit economics: $79,800 of priced revenue (4,200 × $19) minus $1,184.40 of cost is $78,615.60. That is the line founders should reference when talking to themselves about whether the product is healthy.

Two caveats keep the number honest. First, $79,800 of priced revenue assumes 100% of MAUs are paying. In reality, the conversion rate from MAU to paid is the activation-retention compound, around 17.6% in this case. Real recoverable monthly revenue is closer to 739 × $19 = $14,041, not $79,800. Second, the cost figure excludes the founder's time, which the true-CAC framework insists belongs in every cost-per-customer number a solo founder uses.

5. What actually moves the number

Run the same scenario with five different single-variable changes and the rank order of impact tells the strategic story:

• Activation 32% → 42%: validated customers 739 → 970, cost per validated customer $1.60 → $1.22. A 24% improvement from a 10-point activation lift.
• Retention 55% → 65%: validated customers 739 → 873, cost per validated customer $1.60 → $1.36. A 15% improvement.
• Output ratio falls from 50% to 30%: token cost per user $0.162 → $0.108 (driven by lower output share), total cost ~$1,008/month, cost per validated customer $1.36. A 15% improvement.
• Sonnet output rate $15 → $7.50 (a hypothetical 50% price cut): token cost per user $0.162 → $0.095, total cost ~$903.40, cost per validated customer $1.22. A 24% improvement, equivalent to the activation lift.
• Infra cost $0.12 → $0.08: total cost $1,184.40 → $1,016.40, cost per validated customer $1.60 → $1.38. A 14% improvement.

Two takeaways. First, a 10-point activation lift is equivalent to a 50% model-output price cut at this scale. Activation work is in the founder's control; price cuts are not. Second, the combined gain from realistic improvements (activation up 5 points, retention up 5 points, output ratio down 10 points) compounds into a cost per validated customer near $1.10. That is roughly 30% lower than the baseline with no vendor-side change.

6. The pricing floor at 98.52% gross margin

The 98.52% gross margin against priced revenue looks unassailable, and at a per-validated-customer level (91.6%) it still is. Pricing strategy questions get harder once you account for the share of MAUs who never pay. The product has roughly 3,461 non-validated users (4,200 − 739) consuming tokens and infra without ever becoming paying customers. They cost $1,184.40 × (3,461 / 4,200) = $975.85 per month. If even half of those users could be converted to a $5 starter plan, that recovers $8,652 of annual revenue from cost the product is already absorbing.

The pricing floor question is not "can we afford $19" — at 91.6% margin on validated users, the answer is yes. The question is "what is the lowest price that still produces positive margin on the median non-validated user." If non-validated users consume the same $0.282 of cost per month ($1,184.40 / 4,200) and the conversion rate to $5/month sits at 10%, the math says monthly contribution per non-validated user is $0.50 − $0.282 = $0.218. Positive, but barely. The LTV calculator handles the lifetime extension; the pricing floor sits well below $19 for low-intent users.

For solo founders, the cleanest move is usually a metered free tier with a per-action quota, plus a $19 single tier for activated users. The free tier cost is absorbed as marketing spend; the paid tier carries the 91.6% margin. Avoid the trap of building three tiers before the activation funnel is debugged.

7. Mistakes that hide the real cost

Four patterns recur in solo-founder cost-per-customer numbers. Each one understates the real figure by 30% or more.

• Allocating across MAUs instead of validated customers. The $0.28 number is technically correct and operationally useless. It hides the funnel, makes the product look cheaper than it is, and obscures the activation lever. Always show cost against the validated denominator. The headline MAU-denominated number is a footnote.
• Ignoring infra below 10k MAU. At this scale, infra ($504) is 43% of total cost. Founders fixate on token spend because vendor invoices arrive monthly while infra is fixed-cost and easy to forget. Track them as a single line: product unit cost.
• Counting the cached input tokens at full rate. Anthropic's prompt-cache hit rate at $0.30 per million input tokens is 10x cheaper than uncached at $3^[1]. A product that does not use prompt caching is overpaying on input tokens. The model column in this scenario assumes uncached pricing; cached pricing would drop model cost roughly 35% before retention math.
• Reporting margin against priced revenue, not realized revenue. The 98.52% gross margin against priced revenue is theoretical. Against realized revenue (739 × $19 = $14,041), the margin is still 91.6%, but the absolute contribution shrinks to $12,857 — much smaller than the $78,615 the engine's gross margin line suggests. Solo founders who plan headcount or runway off the priced number end up short.

One operational rule survives every variation. Cost per validated customer is the metric to optimize directly; cost per MAU is the metric to monitor. The first is a business number. The second is a vanity number that occasionally reveals an outage. Run the engine quarterly, watch which lever moved the most, and spend the next quarter on whichever lever has the most slack. For this scenario, that lever is activation, and it will stay activation until the rate clears 45%. See the methodology for the full derivation^[5].