This question-and-answer guide is for AI SaaS founders, finance teams, and engineering leaders reviewing existing customer accounts. You will work through a monthly example, document allocation rules, handle missing data, and leave with a customer-level action list. All numbers below are illustrative USD amounts, not customer results or industry benchmarks.
Download the customer-margin review worksheet (HTML) · Open it in a browser to complete locally, calculate contributions, and print or save as PDF. No signup required.
How do you identify AI customers whose costs exceed revenue?
Build one row per paying customer, using the same period, currency, and cost boundary. Subtract included costs from revenue and sort by contribution in dollars, most negative first. A customer-level view reveals losses that a healthy portfolio average can hide.
Contribution = customer revenue − included customer costs
Contribution margin % = contribution / customer revenue × 100Use the percentage only when revenue is positive. Confirmed zero revenue produces a dollar loss but an undefined margin percentage. Missing revenue produces an incomplete result, not a confirmed loss.
Which records belong in the review?
- Customer mapping: a stable paying-account ID joined to users, workspaces, and billing records.
- Usage: model, timestamp, priced units, execution ID, and real charged attempts, including retries.
- Other delivery costs: tools, retrieval, telephony, speech, compute, and delivery review where included.
- Revenue: the amount assigned to that period and the basis used, with adjustments explained.
- Evidence: source totals, coverage gaps, allocation rule, reviewer, and review date.
For this operational review, document whether revenue is reported, calculated from a plan, or allocated. Do not compare a full annual cash receipt with only one month of costs. If an annual $6,000 contract is allocated evenly across twelve months for this review, use $500 per month and label the allocation. Accounting recognition may differ; use your approved revenue policy for accounting reports. See the separate AI gross margin per customer guide for that reporting boundary.
What does a worked customer-margin review look like?
Beacon costs $150 to serve against $100 of revenue, producing a $50 loss after included costs. Acme and Cedar remain contribution-positive, but Cedar falls below the illustrative 60% target.
Assume one complete monthly period, with revenue allocated consistently, all listed costs priced, and no unresolved amounts in this example. “Review” means included human delivery-review cost. Other company expenses are excluded, so these results are not net profit or full accounting gross margin.
| Customer | Revenue | Models | Tools | Shared | Review | Total cost | Contribution | Margin |
|---|---|---|---|---|---|---|---|---|
| Acme | $500 | $100 | $20 | $30 | $0 | $150 | $350 | 70% |
| Beacon | $100 | $110 | $15 | $20 | $5 | $150 | −$50 | -50% |
| Cedar | $250 | $60 | $15 | $40 | $10 | $125 | $125 | 50% |
| Portfolio | $850 | $270 | $50 | $90 | $15 | $425 | $425 | 50% |
The portfolio earns $425 after included costs while Beacon loses $50. Calculate portfolio margin from totals: $425 ÷ $850 = 50%. Do not average the three customer percentages.
What should the reviewer conclude?
- Acme: Keep plan; monitor usage.
- Beacon: Investigate retries; review allowance.
- Cedar: Review cost drivers against target.
The table identifies where to investigate; it does not prove why costs rose. For Beacon, inspect the executions behind its $110 model spend. Check whether repeated validation, larger context, or a changed model explains the increase before attributing it to customer behavior.
How do you allocate shared AI delivery costs fairly?
Use a documented consumption driver and allocate the pool once. In the example, a $90 shared serving pool is allocated by measured serving time: Acme uses 30 units, Beacon 20, and Cedar 40 out of 90 total. Their shares are $30, $20, and $40.
Customer shared cost = shared pool × customer driver units / total driver units
Beacon shared cost = $90 × 20 / 90 = $20Define what the pool includes and exclude any charge already counted under model or tool costs. When requests vary greatly in size, simple request counts may not reflect consumption. Keep idle capacity, trials, internal work, and unknown customer IDs in explicit categories according to your policy.
Check that customer allocations plus explicit remaining buckets equal the original pool. If a provider invoice confirms the model costs already counted from usage, reconcile it against those costs; adding it again would double-count the same spend.
What should you do when costs or revenue are missing?
Separate “loss-making,” “below target,” and “incomplete.” A customer cannot pass a margin review just because unpriced usage disappears from a total.
| Gap | How to show it | Next action |
|---|---|---|
| Unknown model price | Unpriced usage; costs incomplete | Add the applicable dated or contract rate and recalculate. |
| Missing customer ID | Unassigned cost bucket | Resolve the mapping from supported evidence. |
| Missing revenue | Revenue unknown; margin unavailable | Join the billing or revenue source for the same period. |
| Late tool or infrastructure bill | Provisional cost boundary | Record an explicit estimate if used, then replace it with actuals. |
| Usage differs from invoice | Unresolved reconciliation difference | Review period boundaries, discounts, credits, rates, and missing events. |
For example, a fourth customer, Delta, has $200 revenue and $80 known costs but 200 unpriced calls. Its known-cost contribution is $120, or 60%, but the review status is incomplete. It should not be approved at the 60% target until the missing costs are resolved.
Keep a gap owner and follow-up date in the worksheet. Use the AI invoice reconciliation tool to investigate supported usage-versus-invoice differences. Reconciliation supports cost completeness; it does not by itself assign every cost to a customer.
Should you raise the price or reduce AI delivery cost?
Choose an action after verifying the cost driver and testing its effect on outcomes. A cheaper request can be more expensive overall if it causes additional retries or human review.
Suppose Beacon’s review verifies $40 of avoidable retries. Removing them reduces included monthly cost from $150 to $110. At the same $100 revenue, contribution is still −$10. The improvement helps, but it does not fix the pricing gap.
At an illustrative 60% target margin and unchanged $110 monthly cost, required revenue is $110 ÷ (1 − 0.60) = $275. Alternatively, a $100 plan allows only $100 × (1 − 0.60) = $40 of included cost at that target. Neither calculation proves what the customer will buy; model the expected usage and contract terms before proposing a change.
- Engineering: test retry fixes or model routing against the same quality and completion criteria.
- Product: review included allowances, expensive features, and upgrade paths.
- Finance: verify revenue basis, rate changes, and complete cost coverage.
- Account owner: discuss any plan change with the customer under the applicable contract.
Record the chosen action, owner, expected impact, and review date. Repeat the calculation next period using the same boundary. See how to review AI subscription pricing for the broader plan analysis.
How do you use the free customer-margin worksheet?
Download the HTML worksheet, enter one period’s data, and calculate the review locally. It includes the three example customers, editable costs, allocation notes, completeness status, and action owners. It flags incomplete and zero-revenue rows and calculates portfolio margin from totals.
Replace the sample values with your own. Use the evidence notes to record any shared-cost remainder and reconcile it before approving the review. Print or save as PDF to retain your inputs; the worksheet does not upload or persist them automatically.
Choosing software for this process? Compare AI cost-per-customer tools by attribution, revenue matching and reconciliation.
Need to verify the source totals first? Follow the monthly AI reconciliation walkthrough and sample close pack.
How can Ganivra help with an ongoing customer-margin review?
Ganivra connects recorded AI costs with customer and revenue context so you can investigate contribution by customer, plan, and workflow. Instrument the relevant calls, use stable customer IDs, and supply the revenue and additional delivery costs your review requires.
Start with one workflow and reconcile its evidence before expanding coverage. The customer-margin review workflow explains the review process, and the integration guide covers capturing usage. Reported contribution reflects included costs; missing inputs cannot be repaired by a dashboard alone.
Frequently asked questions about AI customer profitability
How can I tell which customers cost more in AI usage than they pay me?
Group recorded model, tool, and delivery costs by paying customer for one period, then subtract those costs from that customer’s revenue on a stated basis. A negative contribution identifies a customer whose recorded costs exceed revenue. Check missing usage, revenue, and shared-cost allocations before making a pricing decision.
How do I calculate AI cost per customer?
Sum priced model calls, charged retries, paid tools, and other included delivery costs tagged with a stable customer ID. Add the customer’s documented share of shared delivery costs. Reconcile the total with source records, keeping unassigned and unpriced usage visible instead of treating it as zero.
What software connects LLM usage costs to customer revenue?
Ganivra links recorded AI usage and delivery costs to customer context and reported or calculated revenue. It shows contribution after included costs by customer, plan, and workflow. Complete results depend on instrumenting the relevant usage and supplying revenue and additional delivery costs; Ganivra does not replace your accounting ledger.
Should failed requests and retries count toward customer cost?
Yes, include genuine charged attempts incurred while delivering that customer’s service, even when the final result fails. Deduplicate duplicate telemetry records, not real charged retries. Keep internal experiments separate from customer delivery.
How do I allocate shared AI costs across customers?
Define a shared cost pool and a consumption driver, such as measured serving time. Multiply the pool by each customer’s share of the total driver. Allocate each cost once, document the period and rule, and keep idle or unassigned costs in an explicit bucket under your policy.
What should I do when AI cost data or customer revenue is missing?
Mark the result incomplete and list the missing records. An unknown price is not a zero cost, and missing revenue is not confirmed zero revenue. Assign an owner to resolve each gap before approving a customer’s margin or changing their price.
Is margin after AI cost the same as gross margin?
No. Margin after recorded AI and delivery costs describes only the included cost boundary. Accounting gross margin requires the appropriate revenue basis and all applicable cost of revenue. Neither measure is net profit, which also reflects other business expenses.
How should I price an AI customer who loses money?
First verify the loss and its driver. Evaluate reducing avoidable retries, changing model routing after quality checks, revising usage allowances, or discussing a new plan. For a fixed cost estimate and target contribution margin, required revenue equals cost divided by one minus the target margin. Recalculate if usage or quality changes.
Ready to review your own customer margins?
Use the worksheet for a first review, then connect a real workflow to make the next review repeatable.
Start a Ganivra trial →