# Your AI Feature Is Popular but You Cannot Explain Its Margin

> Join AI usage, cost, tenant, pricing plan and billing records; distinguish missing attribution from an economic problem before changing prices.

Source: https://grandrounds.tech/insights/ai-feature-margin · Published 2026-11-10 · CTO Grand Rounds

**Short answer.** When a paid AI feature's margin is unclear, check record joins before deciding the price is wrong: usage to workflow, cost to customer account (tenant), tenant to pricing plan, plan to billing record. Investigate the first link you cannot reproduce. Once all four hold, compare attributed revenue with full cost to serve for each tenant—not only the average—to see whether the plan covers it.

**Who this is for:** founders, CFOs, product leads and engineering leads of an AI SaaS product with real usage and billing. Product sees adoption, engineering sees inference charges, finance sees subscription revenue. This page shows how to make the three views describe the same work.

*Figure: A chain from usage through cost and tenant to plan has a highlighted missing join; the headline reads Find the missing link.*

## Three dashboards may be answering different questions

Product may count customer jobs, engineering execution attempts, and finance recorded usage or revenue; check what each counts in your company. Define events before comparing totals. Keep system retries, user-requested reruns and delayed human acceptance identifiable.

Customer action, execution attempt, accepted result and billable unit are separate definitions; their totals can differ.

## Locate the broken link

Pick a closed period and contrasting tenants. Follow usage to workflow, cost to tenant, tenant to the plan effective at authorization and that policy to the billing record. Keep unattributed spend explicit and assign its investigation.

| Link | What you need to explain | A gap worth investigating |
| --- | --- | --- |
| Usage → workflow | Which customer job produced these execution attempts and outcomes? | Background jobs, retries, or tools have a missing parent job identifier |
| Cost → tenant | Which tenant and workflow incurred the recorded cost? | Shared credentials or missing identifiers leave a pool of unattributed spend |
| Tenant → plan | Which plan, allowance, and policy applied when the job was authorized? | Today's plan is applied to yesterday's work; exceptions live only in messages |
| Plan → billing record | Why did this action consume this many units, including failures and corrections? | Billing counts attempts, duplicate messages, or late events differently from the agreed rule |

## What a margin explanation needs

For one workflow and one period, a finance-ready explanation has three parts.

- **Revenue treatment.** The part of subscription or usage revenue that finance assigns to this workflow. Finance chooses and records the portion of shared subscription revenue attributed to the feature.
- **Cost to serve.** Model and tool cost, retries, review and support time that you can trace to the tenant and workflow, shown next to an explicit unattributed amount.
- **Distribution.** Cost per tenant and per job as a spread. An average can look healthy while a few heavy tenants carry most of the spend.

Workflow margin is then attributed revenue minus attributed cost to serve, reported together with the unattributed amount. Until the four links hold, report "explanation unresolved" for that workflow, with the missing link named.

## Illustrative example: one unattributed spend pool, three tenants

Imagine an AI document-review feature sold on one flat plan. For the closed period, the provider invoice is the 100% total of AI spend. The team can tag it as follows:

| Spend line | Share of AI spend |
| --- | --- |
| Tagged to Tenant A | 30% |
| Tagged to Tenant B | 22% |
| Tagged to Tenant C | 10% |
| Tagged to the feature, tenant unresolved (a shared service credential used by a background job) | 25% |
| Untagged | 13% |

Tenant A looks like the heaviest user, and an equal split of the 25% pool agrees (A 38.3%, B 30.3%, C 18.3%). The ranking remains unresolved in the records. If the whole pool belongs to C, C reaches 35% and A stays at 30%. If it belongs to A, A reaches 55%. A's true share lies between 30% and 55%, B's between 22% and 47%, C's between 10% and 35%. These ranges ignore the 13% untagged amount, which could widen them further.

*Figure: Invented 100-percent spend bar: A 30, B 22, C 10, shared credential 25 and untagged 13. Range bars show A 30–55, B 22–47 and C 10–35 before assigning the untagged 13; equal-split ticks are 38.3, 30.3 and 18.3. Next check: trace each job or attempt to its tenant ID and reconcile the shared spend.*

The next step is therefore a measurement step: find which job uses the shared credential and attach a tenant identifier to it. A price change or a usage cap for Tenant A would be premature. Review time is a further cost line; keep its unmeasured value explicitly unknown.

## Build a usage–cost–tenant–plan record

Start with approved traces, job records, plan versions and billing exports. Use identifiers and permission-controlled evidence references. Finance chooses revenue allocation; product defines the customer-visible unit; engineering owns event identity; operations supplies review and support effort.

| Field group | Record | Question it resolves |
| --- | --- | --- |
| Scope and identity | Workflow/version, closed period, tenant ID, customer job ID | Are the records comparable, and whose work is this? |
| Revenue treatment | Revenue attributed to the workflow, the finance-approved allocation method and who approved it | What part of revenue does this workflow carry? |
| Cost allocation | Attributable runtime/tool cost, review/support reference, allocation method | Which costs are traced to the tenant and workflow? |
| Unattributed amount | Spend with no tenant or workflow, with the suspected cause | How large is the unresolved pool, and who investigates it? |
| Plan and billing link | Plan/version in force, billable-unit rule, units recorded for the job | Which join is the first one you cannot reproduce? |
| Evidence and owner | Source reference, confidence/unknown status, person responsible for each gap | What can be verified, and who accepts the explanation? |

Reconciling retries, replay and late billing events across systems is covered in the [companion guide](https://grandrounds.tech/insights/ai-usage-based-billing).

## Distinguish a measurement gap from an economic problem

A missing join needs instrumentation or reconciliation. Once the links hold, compare each tenant's attributed revenue and full cost to serve as a spread, reporting unattributed spend alongside it. If tenants cover those costs, the evidence may support keeping the plan. If a few heavy tenants do not, investigate a tier or cap for that pattern; if most do not, examine plan design and workflow cost. These are prompts, not pricing thresholds or a recommendation to change contracts. Use [cost per accepted result](https://grandrounds.tech/insights/ai-inference-economics) for the underlying unit-cost analysis.

## What shapes the scope of this measurement work

- How many workflows are in scope: start with one.
- How many systems hold the records (AI gateway or traces, job queue, billing system, finance exports).
- Whether event and tenant identifiers already exist, or must be added first.
- Who owns revenue treatment (finance) and the customer-visible unit (product).
- Who builds the change: your team or a provider you choose. Implementation is a separate scope.

## Questions and answers

### Does a growing AI bill mean the feature loses money?

Not by itself. Margin depends on the revenue assigned to the workflow and on the cost you can trace to it. A rising bill with unattributed spend leaves the economic result unresolved.

### What is the first thing to check?

Pick one workflow and two contrasting tenants for a closed period. Test whether usage joins to the workflow, cost to the tenant, the tenant to the plan in force at that time, and the plan to the billing record.

### When can we leave the setup as it is?

When the existing billing system and observability already explain the workflow and the reviewed records agree. Document the explanation and keep the current setup.

### Does Grand Rounds build the metering or billing change?

Grand Rounds does not build the change. The curator listens to the problem and can identify relevant experience in the CTO network. Your team or a provider you choose implements any agreed change.

## Boundaries

The figures are invented spend shares. Finance owns revenue treatment; product owns the billable unit. Report unattributed cost explicitly. This diagnostic record supports a measurement decision, rather than accounting, tax or pricing advice.

Grand Rounds is an initiative of Dinarys, and a conversation may lead to cooperation with Dinarys.

## Next step

Complete one closed-period record and name the missing link, owner and next check before changing a price or cap. Once attribution holds, compare [metering, entitlements and cost controls](https://grandrounds.tech/insights/ai-usage-based-billing). Bring a nonconfidential summary to the [Grand Rounds curator](https://grandrounds.tech/talk-with-the-curator) to discuss relevant CTO experience.
