Pre-launch preview · Stepwell Technologies Pvt Ltd (incorporation in progress) · details marked [ ] are being finalised
Home / Insights / FinOps
FinOps

FinOps for AI: unit economics before scale

Almost every FinOps team now manages AI spend. Few can say what one resolved ticket or one processed document costs. Here is how to get there on AWS.

Key takeaways

  • 98% of FinOps practitioners now manage AI spend (State of FinOps 2026), but few track cost per unit of business value.
  • Define the unit first, with a quality condition: cost per resolved ticket, not cost per token.
  • On AWS, Bedrock application inference profiles plus tagging make AI cost attributable by use case.
  • Optimise in order: route by task, cache, batch, trim context, cap agent steps.

AI spend has become a core FinOps responsibility almost overnight. In the FinOps Foundation's State of FinOps 2026 survey (1,192 respondents representing more than $83 billion in annual cloud spend), 98% of respondents said they manage AI spend, up from 31% two years earlier. FinOps for AI was the top forward-looking priority.[1]

Managing AI spend isn't the same as understanding it. Most teams can see a monthly model bill. Far fewer can say what one resolved support ticket, one processed invoice or one generated report costs. Without that number you can't tell whether an AI use case is worth scaling, or which optimisation to do first. It's also how projects die: Gartner lists escalating costs among the main reasons it expects many agentic AI projects to be cancelled.[2]

Step 1: Define the unit before you optimise anything

The FinOps Framework calls this unit economics: relating cloud cost to a measure of business value.[3] For AI, choose a unit the business already counts:

Use caseGood unitWeak unit
Support assistantCost per resolved ticketCost per 1,000 tokens
Document processingCost per document accepted without reworkCost per API call
Code assistantCost per merged changeCost per seat
Incident triageCost per triage note accepted by the engineerMonthly model invoice

Note that the good units include a quality condition. A cheap answer that has to be redone isn't cheap.

Step 2: Make AI cost attributable

On AWS, model invocation costs can be allocated by application. Amazon Bedrock supports application inference profiles, which can carry cost-allocation tags so usage shows up by team, product or use case in your billing data.[4] Combine that with:

  • a tagging policy covering every AI-related resource: vector stores, GPU instances, storage and logs, not only model calls;
  • per-request logging of model, input tokens, output tokens, cache hits and latency, joined to the business event (ticket ID, document ID);
  • a weekly unit-cost report per use case, owned by a named product owner.

Step 3: Optimise in the order that pays

  1. Route by task. Not every step needs the largest model. Classification, extraction and routing often run well on smaller, cheaper models; reserve the most capable model for reasoning-heavy steps. Prove each switch on your evaluation set.
  2. Cache what repeats. Long system prompts and shared context are re-sent on every call. Prompt caching, where your provider supports it, can reduce the cost and latency of those repeated tokens.[5]
  3. Batch what can wait. Overnight document runs and evaluation jobs rarely need real-time responses. Bedrock batch inference is offered at a lower price than on-demand for supported models.[6]
  4. Trim the context. Retrieval that sends twenty documents when three would do is the most common hidden cost. Measure retrieval precision, not just answer quality.
  5. Cap the agent. Set a maximum number of steps and tool calls per task, and alert on outliers. Runaway loops are a cost incident.

Step 4: Put guardrails on spend, not just on content

  • Budgets and anomaly alerts per use case, not one account-wide alarm.
  • A cost-per-unit threshold in the release checklist: a change that raises unit cost by more than an agreed percentage needs sign-off.
  • A monthly review of model routing, because prices and model options change often.
If you can't state the cost of one unit of value, you aren't ready to scale the use case.

A two-week AI cost review

For teams already running AI in production, a short review usually answers three questions: what does each use case cost per unit today, where does that cost go (model, retrieval, infrastructure, retries), and which three changes would cut it most without hurting quality on the evaluation set. Savings are then measured against the actual invoice, not a projection.

This article is general information, not legal, regulatory or financial advice. Figures are quoted from the public sources listed below as at the publication date; check the source for updates.

Sources

  1. FinOps Foundation, State of FinOps 2026, data.finops.org, accessed November 2026. data.finops.org
  2. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", press release, 25 June 2025. gartner.com
  3. FinOps Foundation, FinOps Framework, "Unit Economics" capability. finops.org/framework
  4. AWS, Amazon Bedrock User Guide, "Set up an application inference profile" and cost allocation tagging. docs.aws.amazon.com
  5. AWS, Amazon Bedrock User Guide, "Prompt caching for faster model inference". docs.aws.amazon.com
  6. AWS, Amazon Bedrock pricing, batch inference. aws.amazon.com/bedrock/pricing