AI cost management: how to control AI spend at scale

Accenture now funds US$29 a month per person for Claude, after some staff ran bills of up to US$1,000 a month. That is AI cost management arriving the hard way: a cap set after the money is spent.

Good AI cost management works the other way round. You meter every model call, tie it to a business outcome, send routine work to cheaper models and put a hard stop on every agent before it can run up a bill. It matters now because token prices fall while agents use far more tokens per task. Gartner expects inference cost per agentic workflow to rise more than fivefold through 2028; inference is running a trained model to get an answer, and it is billed per token.

This guide explains how AI costs build up, the myths that inflate them, the levers that cut them, and a pilot you can run this quarter to price one workflow by its outcome.

What is AI cost management?

AI cost management is the practice of knowing what every use of AI costs, who owns that cost and what it returns. It has four parts:

• Visibility. See AI spending across models, applications, infrastructure, data pipelines, shared services, and indirect costs, not only the model vendor's invoice.
• Allocation. Charge every cost to an owner, a team and a use case, so each line has someone who answers for it.
• Optimisation. Cut the cost of each completed task with model routing, caching, batch processing and smaller models.
• Governance. Set the budgets, limits, review dates and decision rights that keep spend inside what the business agreed.

Cloud teams already run the first three for servers and storage, a practice often called FinOps (shared management of cloud spend by finance and engineering). AI adds new meters, which the next sections explain.

Why are AI costs rising when token prices fall?

Bills rise because use grows faster than prices fall, and agents multiply use.

An agent is software that plans a task and carries out several steps on its own: it reads a file, calls a system, checks the result and decides what to do next. Each step is another call to the model, and each call is billed. Gartner calls the result the inference paradox. It says routing a task to an agentic reasoning model raises provider inference costs by at least 5x compared with a basic chatbot interaction. Gartner analyst Will Sommer said that agents need 5 to 30 times more tokens per task than basic chatbots.

Model makers also spend their efficiency gains on bigger models. Older capabilities get much cheaper, while frontier models get more expensive; a frontier model is one of the newest, most capable and most expensive models. The newest top models can cost more per token than the ones they replace.

The effect now shows up in large firms. Uber burned through its entire 2026 AI coding tools budget in four months, after an internal leaderboard ranked teams by how much they used AI tools. Forrester lists Tesla limiting AI spending to US$200 a week and Microsoft ending Claude Code licences after it burned through its yearly AI budget. None of these firms lacked finance teams. They lacked meters that matched how AI is billed.

How AI costs build up, step by step

Every AI bill is the sum of a few meters. Once you know them, you can question any vendor quote.

A token is a small chunk of text, often part of a word. Model vendors charge per million tokens the model reads (input) and per million tokens it writes (output). Here is what drives the count:

1. Input. Everything the model reads on each call: your instructions, the documents you attach, the conversation so far and a description of every tool it may use. In a long task, this grows with every step.
2. Output. What the model writes back. It costs more per token. On Anthropic's published list, Claude Opus 5.5 costs US$4 per million input tokens and US$20 per million output tokens.
3. Loops. An agent re-sends its growing context at each step. A ten-step task pays for its opening instructions ten times over.
4. Model tier. The gap between tiers is wide. Anthropic lists Claude Fable 5.1 at US$10 per million input tokens and Claude Haiku 5.5 at US$0.10 for prompts up to 100,000 tokens, rising to US$0.50 above that. For shorter prompts, that is a hundredfold difference.
5. Waiting time. Agents often sit idle while they wait for a tool or a human approval, and the server under them can keep billing. Google has launched Agent Compute Engine to suspend inactive agents and restore them on demand for this reason.
6. Seats and platforms. Per-user licences, platform fees and the data pipelines that feed the model. Forrester names model training, inference, data pipelines, dynamic pricing and specialised infrastructure as the new cost drivers.
7. Hidden multipliers. A model upgrade can change the count itself. Anthropic says the tokeniser in its newer models (the part that cuts text into tokens) produces approximately 30% more tokens for the same text.

On Anthropic's list, a cache hit costs 10% of the standard input price, and writing the cache for five minutes costs 1.25x. Claude Opus 5.5 charges 5% for a cache hit, so the real saving would be larger. Caching pays only while the cache stays warm. Here it does, because calls arrive every few minutes, so the write charge falls on few calls.

What AI spend means for cost, people, risk and scale

AI is now a running cost that grows with use, like a utility bill, not a one-off project with a fixed budget. The comparison stops at one point: a single user or agent can multiply the bill overnight.

Cost. Estimates drift fast once use spreads. Gartner warned in 2024 that chief information officers could miscalculate AI costs by as much as 1,000% as they scale their AI plans.

People. Rollouts now cover whole workforces. Aon has introduced AI to roughly 60,000 employees, and says proposal work that took 40 to 50 hours now takes four or five. It has also strengthened its financial operations capabilities to monitor consumption and manage costs as use grows. Uber has just put all 30,000 employees on ChatGPT Work, an agent product from OpenAI. It did not disclose what it spends, and says it will track how staff use the tool.

Risk. The weak point is the link to value. In McKinsey's 2026 survey, as reported by CIO, only 37% of respondents report AI making a positive contribution to company EBIT, meaning earnings before interest and tax. The share of high performers stayed flat at 6% year over year. Uber's chief operating officer said of the line between AI spend and useful features shipped: "that link is not there yet".

Scale. Controls lag use. A VentureBeat survey found that 21% of enterprises using an orchestration platform have nothing that automatically stops agents' token spending. An orchestration platform is the software that runs and coordinates agents. Without a stop, one faulty loop can run all night.

What people get wrong about AI costs

Four beliefs inflate AI bills more than any vendor price.

Myth 1: cheaper tokens mean a smaller bill

They do not. Gartner expects inference on top models to cost AI firms 90% less by 2030 than in 2025, yet it also found that cheaper tokens will not translate into cheaper enterprise AI. Agents need far more tokens per task, and providers will not pass on every saving. Plan for cost per task, not price per token.

Myth 2: the most capable model is the safe default

For narrow, high-volume work, it is often the expensive default. NVIDIA reports that on Shopify's internal benchmark, a post-trained open model beat a leading frontier model in accuracy while cutting serving costs by 90%. Post-trained means trained further for one specific task after the model was built. An open model is one whose files you can download and run on infrastructure you choose. Amazon's chief technology officer put the test plainly: "do you really need to have the biggest, highest-end model to solve this? The answer is no, you don't."

Myth 3: a per-person cap controls cost

A cap limits spend. It says nothing about value. At Accenture, use above the cap is charged to a project or team budget, and premium plans come with US$250 of token usage a week. Some staff now ask for premium accounts so their projects avoid the extra charge. The cost has not gone away. It has moved into project codes, where it no longer shows up as one AI line.

Myth 4: open-source models are free

The download is free. The running is not. Open models can usually be downloaded for free, but users then pay for their own cloud computing, plus the engineers who keep them working and the security around them. They tend to pay back on one narrow task at large volume, and rarely on many small, varied ones.

How to reduce AI costs: the levers compared

Most savings come from a handful of levers, and each one has a case where it fails.

The discounts are large and published. Anthropic's batch interface gives a 50% discount on both input and output tokens, for work that does not need an instant answer. On the same list, a cache hit costs 10% of the standard input price. Routing is already common: in the VentureBeat survey, 31% send routine, high-volume agent tasks to cheaper models.

Start with routing and caching. They need no new vendor and no retraining, and they apply to most agents you already run.

AI cost management checklist: what to ask before you fund

Ask these questions before you approve an AI budget or renew a contract. A team that cannot answer them is not yet managing the cost.

• What does one completed task cost, all in? Include model calls, compute, seats, the platform and the people who review the output.
• Which model runs by default, and who may override it? Gartner advises teams to block default invocation of frontier models for routine tasks.
• What stops a runaway agent? You want a hard token budget per run and per day, with an alert before the limit.
• How much of each prompt is cached? Fixed instructions and reference documents should bill at the cache rate.
• Which jobs can wait? Reports, reconciliations and bulk classification can often run in batch overnight.
• What happens if the vendor changes its pricing? Anthropic, for example, moved from a flat fee to a usage-based model for autonomous agents. Test the budget against a price rise and a model change.
• Who owns the cost line? Forrester calls for costs fully allocated to owners, departments, business units and use cases.
• What is the ceiling? Agree the cost per task at which the use case stops paying back, before launch, not after the first bill.

How to run an AI cost pilot and monitor it

A pilot should prove cost per outcome on real volume, not show that the model works.

1. Pick one workflow with volume and a clear unit of work. An invoice, a ticket or a proposal. A CIO guide to scaling AI says pilots need sufficient workflow volume, integration feasibility and measurable results.
2. Measure the work before AI. Record time, cost and error rate per unit for a normal month.
3. Tag every call. Each model call carries the workflow, team and use case, so spend reports by outcome, not by vendor invoice.
4. Test two or three model tiers on the same sample of real cases. Keep the cheapest one that meets the quality bar.
5. Turn on caching and batch where they fit. Then set hard limits per run and per day.
6. Run it for four to six weeks at real volume. Report cost per completed unit, quality, rework and human review time each week.
7. Decide. Scale it, re-route it to a cheaper model, or stop it. Each is a good result if the numbers drove it.

The model bill is the small line. The clearance rate decides the case, so measure it before you scale.

AI cost control risks, and how to manage each

Each cost lever brings a risk. Name the control before you pull the lever.

• Shadow AI. Tight caps push heavy users to personal tools. A college CIO told a CIO 100 panel that widespread use of personal AI subscriptions complicates governance. Control: give proven heavy users an approved route with a higher limit.
• Quality loss from cheaper models. A cheaper model that gets more cases wrong costs more in rework. Control: measure the error rate per task, and keep a route back to a stronger model for hard cases.
• Lock-in. Google's new engine is closely integrated with Google Cloud, which one analyst says could make moving agent workloads difficult. Control: keep prompts, test sets and logs portable, and write exit terms into contracts.
• Retries that double the spend. If an agent pauses between an action and its confirmation, it may wake up unsure and try again. One AI engineering lead warns that every external action has to be safe to repeat, meaning a second attempt must not pay an invoice twice. Control: test pause and restart before production.
• Data exposure. Routing work to a cheaper outside model can send company data somewhere new. Control: route only to models inside environments you have approved, and check how Thriiv hosts and runs the systems it builds or ask the same of any provider.
• Price swings. Vendors change prices, plans and models often. Control: stress-test each business case against a price rise before you approve it.

Where Thriiv fits

Thriiv works on this problem from the cost side first. We find where operating cost sits, quantify the saving before budget is committed, and build the system that removes it. We are paid against value, not hours, so cost per outcome is the number we answer for. Every system runs in a private environment we host and manage.

For Anglo, the system Thriiv built cut design time from ten hours to ten minutes per design, a 98% cut, as the Anglo case study sets out. If you want a cost-per-outcome view of one workflow, tell us where the cost sits.

What to do this quarter

Five decisions turn AI spend from a surprise into a number you chose.

1. Name one owner for AI spend across technology, finance and the business.
2. Tag every model call and seat to a team and a use case. Allow no untagged spend.
3. Publish cost per completed task for your three largest AI uses, next to what the same work cost before.
4. Set the default model tier for routine work, with a named route to stronger models, and put hard limits on every agent that can call tools.
5. Review seats and contracts each quarter, and ask every AI vendor the checklist questions above before you renew.

The firms now capping spend after the fact show the cost of waiting. Meter first, then scale what pays.