AI cost management
AI Cost Management
Visibility and control over unpredictable AI costs. Human-led, tech-enabled FinOps, now pointed at AI, from a team that has delivered £50M+ of cloud savings across 20+ enterprises.
Overview
What AI Cost Management Covers
AI costs follow the whole lifecycle of AI, from provider contracts and infrastructure to engineering decisions, token use and culture. Good governance is as much a part of the discipline as the numbers.
What it is
The spend tied to every stage of AI: provider contracts, seats, tokens, the infrastructure that feeds the models and the people who build and run them.
Why it matters now
AI is heading towards a third to a half of IT spend, and providers are moving from fixed per-seat fees to usage-based pricing that can swing month to month.
How we help
We find the drivers of your AI spend, track them, optimise each one on a fit-for-purpose basis and give every line of spend a named owner.
The industry challenge
AI Spend Is Growing Faster Than the Return
Enterprises are spending more on AI every quarter, while only a small share can point to a clear financial result. Cost discipline matters more as buyers move closer to paying true cost.
12%
of 4,701 CEOs saw both cost and revenue benefit from AI, while 56% saw no significant financial impact.
6%
of organisations attribute over 5% of operating profit to AI, even though 85% use it in at least one function.
$37B
enterprise generative AI spend in 2025, up from $1.7B in 2023.
24×
forecast growth in global token usage between 2026 and 2030.
The problem
Why AI Costs Are Hard to Control
AI lands on the same invoices as cloud and SaaS but behaves nothing like them. Most organisations only find out what it costs after the bill arrives.
Spend is non-linear
Token use can jump by orders of magnitude after one product change, model update or shift in user behaviour, so a forecast built on last month's run rate does not hold.
Little visibility
A growing share of AI cost sits inside SaaS products where the token meter is never shown to the buyer, and it can rise at every renewal.
Spread across the P&L
AI cost turns up in cloud bills, SaaS renewals, professional services, capital spend on self-hosting and shadow spend that nobody signed off.
From seats to usage
A fixed fee per user was budgeted once a year. Providers now charge by the token and add extra charges for reasoning time and long context.
Case study
60% Lower AI Cost per Document, With No Drop in Quality
A UK energy supplier needed AI to read business documents and pull out key details, at a cost it could control. We tested two models side by side on the same documents and found one that matched the incumbent's quality for far less.
- Lower cost per document
- 60%
- Reference values found by both models
- 8 of 9
- Spend ceiling, never breached
- £50
The economic model
From Cost to Value
Every AI cost question, from a seat licence renewal to a sudden spike in model spend, is answered by moving through the same four stages.
Cost
What you are billed
Vendor invoices, cloud billing exports, seat licences and contract terms.
Usage
What you consume
Tokens, requests, agent runs, active seats, credits and GPU hours.
Unit economics
What an outcome costs
Cost per token, per document, per case and per seat.
Outcome
What value it delivers
Tickets resolved, code shipped, documents processed and customer contacts handled.
A cost-only view tells you model spend rose 20% last month. The economic model tells you why, what each outcome now costs, and whether that is good value.
The AI estate
Beyond Tokens: The Whole AI Estate
A cost model built only around model APIs misses most of the estate. AI spend sits across seven categories, each with its own commercial model and cost driver.
| Category | Typical examples | Commercial model | Main cost driver |
|---|---|---|---|
| Category: Model APIs | Typical examplesAzure OpenAI, AI Foundry, Bedrock, Vertex AI | Commercial modelPer token, request or cached token | Main cost driverInput, output and reasoning tokens |
| Category: AI seats | Typical examplesGitHub Copilot, Microsoft 365 Copilot | Commercial modelPer seat, tier or add-on | Main cost driverLicensed versus active users |
| Category: Embedded AI | Typical examplesAI features inside ITSM, CRM, asset and analytics tools | Commercial modelBundled, credits or actions | Main cost driverActions, credits or contract allocation |
| Category: Agent platforms | Typical examplesCopilot Studio and agent builders | Commercial modelPer run, task or execution | Main cost driverAgent runs, tool calls, tokens and retries |
| Category: AI infrastructure | Typical examplesGPUs, AI search and vector stores | Commercial modelCloud consumption | Main cost driverHours, requests, storage and throughput |
| Category: Development and evaluation | Typical examplesEvaluation runs, fine-tuning and experiments | Commercial modelConsumption and compute | Main cost driverRuns, datasets, GPU hours and tokens |
| Category: AI data services | Typical examplesRetrieval pipelines, embeddings and document processing | Commercial modelConsumption | Main cost driverRecords, embeddings, queries and storage |
Production and non-production
Three Kinds of AI Spend, Managed Separately
Live services, pilots and experiments behave differently, so each is forecast, governed and optimised on its own terms. Experimentation is treated as investment, not automatically as waste.
Production
Customer contact AI, operations copilots and live document processing.
Forecast drivers
- Active users and transactions
- Tokens per transaction and model mix
- Cache hit rate and business growth
Forecast against business volume and optimised under a quality gate.
Pilot
Workloads moving from proof of concept into a live service.
Forecast drivers
- Adoption curve
- Planned go-live volumes
- Promotion criteria
Explicit promotion gates before a workload is treated as production.
Non-production
Developer copilots, prompt experiments, model evaluations and proofs of concept.
Forecast drivers
- Developer population
- Release and sprint cadence
- Evaluation and test runs
Budgets, alerts, auto-expiry, auto-shutdown and lower-cost default models.
In practice
What AI Cost Management Looks Like
Three typical AI workflows, what drives their cost, and the levers that bring the monthly bill down while holding quality. Switch a lever off to see what it is worth.
- The setting
- A mid-size insurer checks 150,000 documents a month, end to end, with no person in the loop.
- What drives the cost
- Each document passes through six model calls, and every call is sent the whole document again.
- The common issue
- All six calls run on the most expensive model, so the cost sits in the repeats rather than the answer.
The workflow today
- Intake
- Read the document
- Classify
- Check identity
- Check figures
- Check compliance
- Score confidence
- Approve or reject
The levers
- The setting
- A finance team processes 12,000 invoices a month, and a person approves every one before it posts.
- What drives the cost
- The AI draft is cheap. The real cost is the time each reviewer spends on every item.
- The common issue
- A frontier model drafts work that a person checks anyway, and it runs live when most of it could wait overnight.
The workflow today
- Intake
- Draft a recommendation
- Approve or edit
- Post
The levers
- The setting
- A consumer business answers 800,000 support questions a month, using its own knowledge base.
- What drives the cost
- Volume is very high, so the cost depends on how much of it still reaches the frontier model.
- The common issue
- Every question goes to the frontier model, when most could be answered just as well by a smaller one.
The workflow today
- Customer question
- Route by difficulty
- Answer from the knowledge base
The levers
Case studies
Related Case Studies
AI cost work we have delivered for regulated businesses, and the AI and FinOps work behind our approach to AI spend.
FAQs
Common Questions,
Answered For You
Here are common questions people have about AI cost management, answered for your convenience.
Have more questions? Email us
Contact usWhat counts as AI spend?
More than the API bill. We look at five families: direct consumption (model tokens and self-hosted inference), supporting infrastructure (vector databases, storage, pipelines, networking), build and adaptation (training, fine-tuning, data licensing), people and governance (engineering, evaluation and observability tools), and AI that is embedded or hidden in SaaS products and shadow use.
How is this different from FinOps for cloud?
It follows the same inform, optimise and operate cycle, in line with the FinOps Foundation's FinOps for AI scope, but AI behaves differently.Models are replaced within months, so long commitments carry more risk.Pricing is not linear: tokenisers, caching, context length and reasoning time all move the effective price.Demand is harder to forecast, and cheaper units often lead to more use, so total spend can still rise.
Will cutting AI costs reduce quality?
Not if it is done properly. Every change is tested against a quality bar agreed with your subject matter experts. Old and new setups run side by side on real volumes, hard cases still reach the most capable model, and we only switch once the quality bar holds.
Do we need to install new software?
No. Where a read-only integration can be approved, our tooling speeds things up. Where it cannot, we work from your own usage exports, billing data and admin console reports, and build the views you need inside your environment so you keep them when we leave.
What do you need from us?
An executive sponsor with authority over AI budgets, access to provider usage and billing data, permission to instrument a few key workflows, and a few hours a week from a small working group across finance, IT, security and engineering.
How long does it take to see results?
The Proof of Value runs for four to six weeks. Quick wins such as idle seats, orphaned API keys and frontier models on routine work often surface in the first fortnight, and you finish with a ranked, costed roadmap for the rest.
Can you help with AI we are building with you?
Yes. When we deliver AI for clients, we build cost tracking in from the start: tokens and model tier recorded on each run, and cost per outcome reported next to the quality score, so scaling decisions rest on evidence rather than estimates.