---
title: "Track AI cost per finished task"
description: "Connect AI spending to accepted work by counting every attempt, review, and supporting resource, then comparing tasks with similar scope and quality."
slug: "track-ai-cost-per-finished-task"
url: "https://withlodge.com/writing/track-ai-cost-per-finished-task"
published: true
published_at: "2026-10-20T12:00:00.000Z"
updated_at: "2026-11-02T15:19:22.898Z"
image_url: null
image_alt: null
image_width: null
image_height: null
word_count: 776
---
A research brief can consume resources long before anyone accepts it. There may be abandoned searches, revised instructions, failed tool calls, a second draft, and a manager's review. Looking only at the final generation leaves much of the cost outside the account.

Consider a hypothetical furniture business asking an agent to prepare a research brief on a possible new customer segment. The business question is specific: what should the commercial team investigate before deciding whether to pursue the segment? Tracking cost per finished task begins by defining a useful answer to that question.

## Agree on what finishes the task

Set the acceptance conditions before the work starts. For the brief, these might include a defined segment, traceable sources, a distinction between evidence and assumptions, relevant uncertainties, and questions that a commercial manager can act on. The manager remains responsible for accepting the work.

A draft awaiting review is unfinished. A brief rejected for unsupported claims is unfinished. A document that meets the agreed conditions counts as accepted even if the manager ultimately decides the opportunity is unattractive. The business decision and the quality of the research are separate matters.

Give the task a stable identifier that follows its attempts and revisions. Keep its scope visible. If a request changes from an initial segment scan to a detailed investigation of several markets, record that change instead of quietly comparing it with the original assignment.

Decide how reopened work will be treated. A correction needed to meet the original conditions belongs with the original task. A new business question can become a new task. Consistent treatment makes the accounting useful.

## Build the complete cost record

Capture attributable model use, paid tools, computation, and other resources used across the task's attempts. Include failures and retries as well as the accepted draft. When work stops without completion, its consumed resources still belong in the period's cost record.

Account for human effort too. Record time spent preparing inputs, reviewing drafts, investigating errors, and correcting the result. Use an agreed internal method to value that effort, or report time separately where a monetary allocation would create false precision. Be clear about which method the comparison uses.

Shared resources need a stated allocation rule. A subscription or common runtime may support several workflows; assigning its full cost to each task would double count it. Allocating none of it would make the work appear cheaper than it is. Choose a defensible rule and show it alongside direct consumption.

Keep amounts traceable to their source records and avoid counting the same charge through both a provider record and an internal summary. A useful ledger should help someone explain a change in cost without reconstructing the entire workflow from memory.

## Compare similar work at the same quality

For a defined group of tasks, cost per finished task is the total attributed cost divided by the number of accepted completions. Show the attempts, unfinished cases, and time period beside that result. The ratio alone cannot explain the operation.

Use a consistent boundary for the reporting period. Work started in one period and accepted in another needs a clear treatment, particularly when comparing a new workflow with an established one. State the method so a change in accounting does not masquerade as an improvement.

If there are no accepted completions, the ratio is undefined. Report the cost incurred, the unfinished work, and the reason for the lack of completions. There is no useful average to manufacture.

Compare tasks with similar scope, source availability, urgency, and acceptance standards. In the furniture business, a brief on a well-documented segment differs from research into a poorly understood one. A shift toward harder assignments can raise cost even while the process improves.

Keep quality and timeliness visible. Lower cost achieved through thinner evidence or a growing pile of unreviewed drafts does not establish better performance.

## Use the account to govern the next task

Agree on resource limits before a run begins and define what happens when the work approaches them. A workflow may pause with its findings and unresolved questions so an authorized person can decide whether further investigation is justified. The agent cannot approve its own larger budget.

Review costly unfinished cases with the operating owner. Repeated searches may point to an unclear question; extensive corrections may reveal unsuitable source material. The next improvement should follow the evidence rather than an automatic demand to use a cheaper model.

For the next research request, attach a simple cost record to the acceptance record. Bring both to the review: what the team accepted, what it took to produce, and what remains unresolved. Use that pair to decide how much work to authorize next.