If sales, content and customer support all use the same AI model, the model's monthly bill is not a departmental expense report. It describes a technical purchase shared by several kinds of work. Allocating the entire amount to the team that first requested the model can make later users disappear from the accounts without reducing their actual consumption.

AWS announced Bedrock product-attribute filtering and grouping on October 8. The dimensions include provider, model, inference type and feature, with related budget filtering. These are useful ways to inspect technical spending. They do not automatically contain a business's own distinction between enquiry handling, product-content preparation and after-sales support.

For an international operation sharing AI services with a China-based team, there are consequently two questions to answer: what technical service incurred the charge, and what work used it? A single table can show both, but the second answer needs information that a model label alone cannot provide.

Map the workflow instead of guessing from the model

One department can use several models, and one model can serve several departments. Looking by model helps investigate technical choices. Looking by workflow helps explain the business activity. Neither view should quietly substitute for the other when managers compare spending.

A practical starting point is a stable workflow category attached where the request is made. Enquiry classification, product-document preparation and support retrieval are different uses even if they share a runtime. Retain an appropriate basis for connecting that activity with the cost record, rather than asking an AI system to infer ownership from a product name.

The mapping does not require the finance report to contain buyer emails or full conversations. In many cases, a task category and responsible business area are enough for attribution. Collecting more sensitive commercial text merely because it is available makes the report harder to share without necessarily making the allocation more accurate.

Shared activities need an explicit treatment. An application operated by the technical team may primarily serve sales. A background knowledge refresh triggered through a sales interface may support several teams. Define which activities count as direct use and which belong in a shared pool before interpreting the numbers.

There is a reasonable tradeoff between detail and effort. For a small shared service, workflow-level attribution may be sufficient. Personal allocation can create an appearance of precision that the records cannot support. If management chooses a fixed ratio, headcount basis or usage basis for shared costs, label it as an allocation rule rather than a fact supplied by the cloud bill.

Keep missing attribution visible

The current AWS guide says costs without a value for a selected product attribute are grouped together. Product-attribute data starts on September 1, 2026. Keep that boundary visible in comparisons with earlier periods: absent attribute data is not evidence of absent spending, and different coverage can change the apparent breakdown.

Begin reconciliation with the total for a consistent period and cost basis. Then distinguish amounts mapped to workflows, amounts intentionally shared and amounts not yet attributed. If they do not reconcile, investigate coverage before spreading the difference across departments to make the report look complete.

An unallocated amount is informative. It may identify a workflow without a category or a cost that cannot be connected at the chosen level. Keep that uncertainty visible until there is evidence or an agreed policy for it. A neat departmental chart is less useful if it conceals the assumptions needed to produce the totals.

Cost ownership also differs from economic value. A workflow can be correctly assigned to sales while still producing outputs that require substantial rework. Our discussion of cost per accepted AI task addresses quality, retries and human effort. Link those questions where needed, but do not use the model bill as a complete measure of the task's value.

For cross-border reporting, agree on the period boundary as well. The day work occurs, the date a cost is aggregated and the internal reporting cutoff may serve different purposes. Teams can produce individually valid numbers that do not match because they selected different periods. Resolve that definition before arguing about departmental shares.

The first useful business view can remain modest: known workflows, the calls they account for and a clearly labelled shared balance. Add detail when it answers a real management question. If one small team uses the service and there is no allocation dispute, model analysis and a shared total may be all the accounting structure it needs.

Sources