What does one completed AI task really cost?

Cost per accepted outcome

To calculate the cost of an AI-assisted business task, divide the cost of operating the whole workflow by the number of distinct outcomes that meet its acceptance criteria. Include model usage, connected services, human review and operating support. Keep the expense of unsuccessful attempts in the calculation, and report unfinished work separately.

A model's price per token answers a narrower question. A token is a unit used to process text. Completing a business task may require several model calls, a document lookup, a correction and a colleague's decision.

For GCC organisations preparing their next AI investment, this gives finance and operations a way to compare the proposed service with how the work is done today.

Connect AI spending to accepted work

The FinOps Foundation's State of FinOps 2026 identifies managing AI spend as a leading forward-looking priority among its respondents. It is a survey of the global FinOps community, not a measurement of GCC adoption or evidence that any particular investment will pay off. State of FinOps 2026.

• State of FinOps 2026

The Foundation's Unit Economics guidance distinguishes resource measures, such as cost per token, from business measures, such as cost per case resolved. Both are useful. Engineering needs to understand resource use; the business needs to understand the cost of work it can rely on. FinOps Foundation: Unit Economics.

• FinOps Foundation: Unit Economics

Bridges proposes the worksheet in this article as a management tool. It is not a vendor benchmark, an accounting standard or a client result.

Define what counts as completed

Start with a task small enough for its owner to judge. In an illustrative business-services team, an AI assistant prepares responses to requests for account statements. It locates the correct customer and entity, retrieves the permitted records and prepares a response for approval.

The process owner defines completion: the response uses the right account and reporting period, an authorised colleague approves it, and delivery is recorded. An attractive draft with the wrong entity is not an accepted outcome. Neither is a response that the system says it sent when delivery failed.

Anthropic's January 2026 evaluation guidance makes a useful distinction between the agent's transcript and the actual state left in the environment. Check the resulting business record to confirm that the task finished. The guidance concerns evaluating agents; it does not validate the cost assumptions in this article. Demystifying evals for AI agents.

• Demystifying evals for AI agents

Count each request once, even if three attempts were needed. Where a colleague repairs or takes over the work, a completed request can count towards the combined service's output. Record that human contribution separately so the completion figure cannot be mistaken for an autonomous success rate.

Agree the period and cohort before looking at the result. For a pilot, a fixed group of requests is often easier to follow than an open queue containing work from several months. Record the cutoff, outstanding cases and any later corrections. The eventual cost of that cohort will change if work remains unfinished.

Include review and operating costs

The model invoice is one input. A practical worksheet also needs the cost of retrieval, document processing, connected tools, storage and the people who operate the service. Use the actual configuration being evaluated and the commercial terms that would apply after the pilot.

Start with the categories below. Existing records may be enough; additional monitoring software is not a prerequisite.

For an early pilot, a case identifier, billing export and short review-time log may be enough. Link the request, attempts, final status and review effort. Restrict access to these records and avoid copying customer documents into a cost worksheet merely to make it more detailed.

A shared platform needs an allocation rule. Explain whether its cost is assigned using usage, capacity or another defensible basis. If some costs cannot yet be attributed, show the gap; do not silently record them as zero. A supplier invoice in another currency also needs a documented conversion basis before it is combined with AED or SAR operating costs.

An illustrative AED worksheet

Consider 1,000 requests entering the example service. At the review cutoff, 650 meet the acceptance criteria without correction and 250 are accepted after correction or manual completion. Another 100 remain unresolved. Routine approval can still be required for the first 650; "without correction" does not mean "without a person".

All figures below are fictional. AED is used to make the example familiar to UAE readers; the amounts are not market rates, salary benchmarks or Bridges results. Costs cover this cohort through the same cutoff, including effort spent on its unfinished requests.

There are 900 accepted outcomes. The combined service's operating cost per accepted outcome is AED 5,400 ÷ 900 = AED 6.00. Its model spend per incoming request is AED 600 ÷ 1,000 = AED 0.60. These measures answer different questions; the second does not represent the cost of delivering an accepted response.

The 100 unresolved requests remain visible beside the AED 6.00 figure. Later work on them will add cost and may add accepted outcomes. Do not call this the final cost of processing all 1,000 requests, or add their eventual completions without also adding the later effort.

If implementation cost AED 24,000, show it separately. A planning scenario that spreads that amount across six equal workload periods adds AED 4,000 per period. With the same assumed operating cost and 900 accepted outcomes, the combined planning figure becomes approximately AED 10.44 per outcome. This is an investment scenario, not a prescribed amortisation treatment; actual volume, useful life and accounting policy may differ.

Check the effect of review time and model pricing

Do not stop at one average. Hold the acceptance criteria constant and examine changes in review effort, volume and the mix of requests. The following calculations vary one assumption at a time from the illustrative operating case; they exclude initial delivery cost.

Review effort has a large effect in this fictional example. Labour will not necessarily dominate another service with a different workload or architecture.

Keep separate results where they could change a rollout decision. For example, Arabic and English requests, different legal entities, or routine and disputed cases may require different effort. Use those groups only where they reflect actual operations. Small samples do not establish that a model performs well or badly for an entire language or market.

Also inspect the expensive cases. An acceptable average can hide requests that repeatedly call tools, return to review or wait for an owner. Define a stop point and a handover route for those cases, proportionate to the task. A spending limit must not leave a customer request silently abandoned or bypass a required approval.

Use a comparable baseline

Estimate or measure the current process using the same task boundary, request mix, acceptance rules and time window. Include its checking and exception effort too. Comparing a reviewed AI workflow with an unchecked manual process will not support a useful decision.

Keep an ordinary automation option in the comparison. Anthropic's December 2024 guidance recommends starting with the simplest suitable approach and describes cost and latency trade-offs as agentic systems become more complex. Its tooling has evolved since publication; the architectural advice is the relevant point here. Building effective agents.

• Building effective agents

In the account-statement example, a fixed lookup and approval workflow may be sufficient for standard requests. An agent may be worth testing for exceptions that require choosing among several sources. Use observed performance and total effort to decide which approach fits each part of the work.

Cost per accepted outcome is also incomplete on its own. Keep service time, material errors, rework and unresolved cases alongside it. A cheap service that sends an incorrect statement to a customer fails the acceptance criteria, whatever its unit cost.

Separate released capacity from cash savings

A reduction in handling time can be useful without reducing payroll. It may let the same team serve more customers, reduce a backlog or spend more time investigating difficult cases. Report the hours released and what the team did with them.

Claim a cash saving only when an expense actually falls, such as a lower overtime payment or a reduced external processing bill. Avoid valuing released hours as labour savings and then counting the same hours again as extra revenue. Keep forecast benefits separate from realised results, with a named owner for each assumption.

For a GCC group, agree the entity, currency and cost basis with the finance owner. Local contracts and data-handling requirements still apply to the workflow. The method here creates no UAE, Saudi or GCC-wide reporting obligation, and cannot establish regulatory compliance or investment returns.

Prepare the funding decision

Prepare one worksheet for a bounded service. It should show the task definition, accepted and unfinished volumes, recurring cost, delivery investment, current-process comparison and the assumptions most likely to change the result. Identify which figures are measured and which remain estimates.

The process owner confirms acceptance and service quality; finance checks the cost basis; engineering explains the usage and difficult cases. Use that evidence to decide whether to expand the pilot, simplify it, repair a bottleneck or stop. There is no universal sample size: the next review needs enough completed work to test the decision.

Bridges can help define the investment case through Business Advisory, assess scope through AI Strategy, Readiness & Enablement, and connect measurement to delivery through Intelligent Automation & Agentic AI. A useful first brief names one recurring task and the evidence its owner needs before funding more of it.

• Business Advisory
• AI Strategy, Readiness & Enablement
• Intelligent Automation & Agentic AI

Common questions