Unit economics dashboards for AI SaaS finance teams

Finance teams at AI SaaS companies are asked questions their old dashboards cannot answer. Which customers are unprofitable? What does a completed task cost us? Is the new model cheaper per outcome or only per token? This guide sets out a dashboard that answers those questions with definitions precise enough to survive an audit.

Start with definitions

Every metric on the dashboard needs a written definition. These are the ones we recommend, with the values from the worked example in the snapshot.

Layout: four rows

• Row 1, the headline: revenue, model COGS, COGS per dollar of revenue, and gross margin with model tokens included, each against the prior month.
• Row 2, the distribution: a histogram of tenant model margin and a list of tenants below zero, with their plan and their cost trend.
• Row 3, the drivers: cost per task by feature, tokens per task, and cost per 1M tokens by model, which separates usage changes from price changes.
• Row 4, the controls: tenants at soft cap, fallbacks served, breaker trips and the spend they prevented.

Reading the example

In the snapshot, model COGS of $11,520 against $64,000 of revenue means every dollar of revenue carries $0.18 of model cost, a model-cost margin of 82.0%. Spread over 640 tenants that is $18.00 per tenant, and over 96,000 tasks it is $0.120 per task.

Those averages are a starting point, not a conclusion. The distribution row usually shows that most tenants sit well above the average margin and a small group sits far below it. The action list for the month comes from that group: which plan they are on, which feature drives their cost, and whether a cap, a fallback or an overage is the right response.

Cost per task beats cost per token

Token metrics tell engineering whether prompts got longer. Finance needs cost per outcome: a resolved ticket, a reviewed contract, a merged pull request. Define a task per feature, emit an event when one completes, and divide. When the team migrates to a cheaper model that needs more retries, cost per token falls and cost per task may rise, and only the second number matters to margin.

Keep it consistent with the books

The dashboard and the general ledger will disagree unless you design for it. Use the same period boundaries in UTC, the same classification of internal traffic as research and development, and the same price source. Reconcile model COGS on the dashboard against provider invoices monthly and display the reconciliation gap on the dashboard itself. A visible gap of a fraction of a percent builds more trust than a hidden one.

Alerts finance should own

• A tenant whose model margin turns negative for the month to date.
• COGS per dollar of revenue rising more than a set amount week over week.
• Tail share rising, which signals concentration risk before it shows in the total.
• Reconciliation gap above threshold at month end.

Where the data comes from

Everything above can be computed from a per-request ledger with tenant, feature, model, token classes and cost, joined to revenue per tenant from your billing system. PALM SOUQ provides the ledger, the margin views and an export to your warehouse, so finance can build on the same numbers engineering uses for quotas and breakers.

Review the dashboard with engineering and product leads monthly. Most margin improvements come from decisions they make: model choice, prompt length, caching and limits.