How to read a week of an agent’s spending

A useful weekly statement connects every paid service call to the job it served. It should show what the agent bought, what it returned, and whether the output was useful. Grouping charges by sales, marketing, content, or support helps you decide what to fund next week. A list of provider names and prices cannot answer that on its own.

Build the report so a reviewer can trace a purchase back to a requested job and forward to a result. Use the structure below as a template for reviewing your own jobs and receipts.

Give the week a boundary

Choose a seven-day window and a timezone. Keep both the time a job began and the time its usage was recorded. A render submitted late on Sunday may finish on Monday; a model call’s small accrued charge may settle with other usage later.

Do not silently combine job-start totals with settlement totals. If the report groups by job start, disclose that choice and identify pending work. If it groups by recorded usage, keep enough identifiers to trace work started in an earlier period.

Keep funding and repayment in a separate section. Adding prepaid credit changes the balance, but it does not mean the agent completed that much work. Settling a credit balance pays an existing obligation; it should not be counted again as new service spend.

Group by the job before the provider

An Instinct assistant might use search, a page scrape, and a video model for one marketing task. It might use the same search provider for a support question later. Grouping only by provider merges those decisions and makes it hard to judge the original requests.

Use a parent job ID in your application and map each returned transaction or usage ID to it. A statement can group that work as follows:

Job	Purchased work	Result to inspect	Review decision
Sales research	Company lookup and verification	Qualified account record with sources	Keep only if it fits the target criteria
Marketing creative	Reference search and generation	Approved image or video	Record discarded attempts too
Content research	Scraping and model usage	Draft with working source links	Flag unsupported claims before reuse
Support repair	Replacement lookup or render	Delivered result tied to the ticket	Check whether the ticket reopened
Local planning	Location or route lookup	The route or place result	Check freshness and relevance

Beside each row, store the exact charged amount, its currency or unit, status, and the underlying receipt identifiers. Keep original precision for usage-priced models. Round only the presentation total after summing comparable amounts.

Read three lines closely

Start with the cheapest useful call. A small lookup can remove a real blocker, such as finding the correct official source for a draft. “Useful” needs an output or a decision attached to it. Low price alone does not establish value.

Then inspect the most expensive job. Include all its attempts, supporting calls, and review work. A final video may be worth the spend even after revisions, but the final successful render should not hide the earlier costs.

Finally, choose one purchase that was wasted. Perhaps an enrichment happened before the lead was checked for fit. Perhaps a second render began while the first was still running. Record the specific change that would prevent the repeat: move qualification earlier, preserve the existing job ID, or require review before another iteration.

Record the decision beside each purchase so the next run can use it.

Reconcile receipts without counting them twice

Keep request IDs, asynchronous job IDs, and billing IDs in distinct fields. One job may have several calls; one accrued settlement may cover several small usage events. Do not sum both the accrued usage and the later settlement debit as though they were separate services.

A successful API response also needs to be distinguished from a useful outcome. A generated image with unreadable copy may be a billable success. A failed later step does not erase successful earlier calls. Vaaya’s errors and billing documentation describes the difference between refusals, failures, and charges.

For usage-priced LLM calls, inspect the endpoint’s usage fields rather than assuming charged_cents: 0 proves the request was free. Preserve finer-grained amounts and their units until reconciliation. Vaaya’s LLM reference explains usage-based model access.

Separate spending from settlement

Prepaid usage consumes money already added to the account. Eligible credit-line usage creates an amount to repay under that account’s terms. Those are different funding paths, even when the purchases appear in one history.

Vaaya’s public pricing page, checked September 23, 2026, describes credit-line collection weekly or when drawn credit exceeds $50. Your reporting window and your account’s collection schedule serve different purposes. Check the current account terms before describing a specific balance as due.

Merchant purchases, where enabled, have their own approval and payment flow. Keep a merchant’s charge separate from metered-tool usage and any service fee. A report should make those amounts distinguishable rather than giving them one ambiguous “cost” column.

Turn the review into next week’s policy

For every proposed cut, name the purchase you intend to prevent. Cache an unchanged company record for an appropriate period. Check a lead’s fit before paying to enrich it. Stop a sandbox when its job ends. Require a visual review before buying another creative pass.

Keep the output links beside the costs, set the next budget, and assign an owner to unresolved rows. The next review should show whether those changes prevented unnecessary work while preserving the results people needed.

Questions

Is an agent’s spending statement a bank statement?

It is a record of purchased services and account activity. Calling it a bank statement is an analogy, not a claim that the agent has a deposit account.

Does a zero-cent response mean the model call was free?

Not necessarily. Some usage is accrued at finer precision and settled in aggregate. Check the endpoint’s usage and billing fields before treating zero charged cents as zero cost.