AI Cost Governance: Budgets and Spend Controls

Only 26 percent of large US enterprises can see their AI costs in real time, according to KPMG's Q2 2026 pulse of 204 C-suite leaders at firms with more than $1 billion in revenue, which plan to spend an average of $202 million on AI over the next 12 months. The analysis recommends five controls — attributing every API call to an owner, capping spend, routing work to the cheapest adequate model, measuring cost per outcome, and reviewing at every vendor change — and compares OpenAI, Anthropic, Azure, AWS, Google Cloud and GitHub Copilot on which settings stop spending versus only alerting after the money is gone. It cites a 28 September 2026 price-tracker entry showing Codex's frontier tier rising to $10 per million input tokens from $5 while its efficient tier fell from $1 to $0.10, and Pragmatic Engineer data across 27 workplaces showing AI coding spend from a $150 monthly cap to $3,000 per engineer.

← All posts https://reinvently.co.uk/blog/ Key Takeaways - Why AI Spend Escapes Normal Budget Controls why-ai-spend-escapes-normal-budget-controls : Prices move without anyone deciding, spend varies widely between teams, and buying is scattered across keys, seats and marketplaces. - Five Controls for AI Spend five-controls-for-ai-spend : Attribute every call, cap spend, route to the cheapest adequate model, measure cost per outcome, and review at every vendor change. - Which Provider Settings Stop Spend, and Which Only Warn which-provider-settings-stop-spend-and-which-only-warn : OpenAI, Anthropic, Azure, AWS, Google Cloud and GitHub Copilot compared on what stops spending and what only sends an alert after the money has gone. - What the Monthly Cost Report Should Show what-the-monthly-cost-report-should-show : The five lines a budget owner needs each month, and which of them billing data alone can provide. - Your First Month your-first-month : Four weeks, from finding every place AI is bought to reporting cost per outcome. Most large enterprises can watch their AI spend. Far fewer can stop it. Only 26 percent can see their AI costs in real time, according to KPMG’s Q2 2026 pulse of 204 C-suite leaders at US firms with more than $1 billion in revenue. Those firms plan to spend an average of $202 million on AI over the next 12 months. If your budget tool emails you after the money has gone, you have a record, not a control. Below are five controls that make AI spend governable, and a check of which provider settings actually stop spending. For the value side, see what companies spend per engineer https://reinvently.co.uk/blog/ai-coding-spend-roi-per-engineer/ . Why AI Spend Escapes Normal Budget Controls Most IT spend comes in units that stay put: a seat, a server, a contract. AI is priced per token, and three things make that hard to budget. First, the price of the same work moves without anyone deciding. On 28 September 2026 our AI coding price tracker https://reinvently.co.uk/tools/ai-coding-tool-prices/ recorded Codex’s frontier tier moving to a model at $10 per million input tokens, up from $5. Its efficient tier fell from $1 to $0.10. One team’s input price doubled overnight; another’s fell by 90 percent. Neither made a purchasing decision. Second, spend varies widely. Across 27 workplaces surveyed by The Pragmatic Engineer, AI coding spend ran from a $150 monthly cap to $3,000 per engineer, with no link to company size. Last quarter’s average tells you little about next quarter’s bill. Third, the buying is scattered. The FinOps Foundation’s practitioner survey ranks token costs in SaaS AI as its members’ top challenge, citing developer-led purchasing, opaque billing and missing allocation tools. AI arrives through API keys, seats, cloud marketplaces and features inside existing contracts. Gartner expects AI coding costs to exceed developer salaries by 2028. That is a forecast, not a measurement. But subsidised pricing is ending, and the organisations that cope best will already know where their tokens go. Five Controls for AI Spend They apply on every platform. The next section shows how each provider supports them. 1. Attribute every call to an owner. Give each AI workload its own API key, project, workspace or tag, so you can trace spend to a team and a use case. The FinOps Foundation calls API key governance “the foundational control layer for everything else”. Shared keys make every other control impossible. 2. Cap spend where the provider allows it. Use hard limits for anything that runs unattended, such as agents and batch jobs, and for every non-production environment. A cap also stops a runaway agent, so it is a safety control as well as a financial one. 3. Route work to the cheapest adequate model. The FinOps Foundation calls this the highest-impact optimisation. On our Ed-o-meter value board https://reinvently.co.uk/tools/ed-o-meter/value/ on 23 October 2026, the best-value model scored 9.2 out of 10 at $0.0004 per task. The top model scored 9.5 at $0.0991: about 250 times the cost for 0.3 points. Revisit the rules whenever vendors change their defaults. 4. Measure cost per outcome, not per token. Track cost per resolved ticket, per release or per document processed. Tokens and seats tell you what was consumed, not what it bought. Our ROI analysis https://reinvently.co.uk/blog/ai-coding-spend-roi-per-engineer/ shows why pull request counts overstate the gain. 5. Review prices at every vendor change. Treat a new model, price or default as a trigger for review, not just renewal. If your spend is large enough to negotiate, write notice periods for price and model changes into the contract. The first two are configuration, and you can have them in place within weeks. The other three need an owner who reads the numbers every month. Which Provider Settings Stop Spend, and Which Only Warn Every major platform offers a budget. What matters is whether reaching it stops the spending. The table reflects each provider’s documentation on 23 October 2026. | Platform | Attribute spend with | Settings that stop spend | Settings that only warn | |---|---|---|---| | OpenAI API https://developers.openai.com/api/docs/guides/spend-limits | Projects and API keys | Organisation and project hard limits; requests then fail with a 429 error | Spend alerts, which OpenAI says “do not enforce a cap” | | Anthropic Claude https://platform.claude.com/docs/en/api/beta/organization/spend limits | Console workspaces and API keys | Monthly organisation and workspace spend limits for Claude Console accounts; per-user limits in Claude Enterprise, set in claude.ai. The API for setting Console limits is in early access | Usage reports, which record spend but do not limit it | | Microsoft Azure | Subscriptions, resource groups, and API Management keys | The API Management token limit policy https://learn.microsoft.com/en-us/azure/api-management/llm-token-limit-policy : a token quota per key, hourly to yearly, after which calls get a 403 error | Cost Management budgets https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets : “consumption isn’t stopped”, and cost data can take 8 to 24 hours to arrive | | AWS Bedrock | Application inference profiles https://docs.aws.amazon.com/bedrock/latest/userguide/cost-mgmt-application-inference-profiles.html with cost allocation tags | AWS Budgets actions https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-controls.html that apply an IAM or service control policy, automatically or after approval | Budget notifications without an action attached | | Google Cloud https://docs.cloud.google.com/billing/docs/how-to/budgets-spend-caps | Projects and labels | Spend cap budgets for the Gemini API and the Gemini Enterprise Agent Platform: one project and one service, monthly, after which new usage is paused. First-party customers only, not reseller accounts | Standard budgets without a spend cap | | GitHub Copilot https://docs.github.com/en/billing/how-tos/set-up-budgets | Organisations, enterprises and users | AI credit budgets with “stop usage” selected; user-level budgets always stop | Any budget without “stop usage”, and threshold alerts at 75, 90 and 100 percent | Hard limits exist on every platform, but they are rarely the default; someone has to choose them. On Azure and AWS the strongest controls sit outside the model service, in a gateway or an IAM policy, so they work only if every call goes through them. Google’s spend caps sit closer to the service, but each covers one project and one service. Even hard limits are approximate. OpenAI says “enforcement is not instantaneous, so recorded spend can slightly exceed the configured amount”. Microsoft says concurrent requests “can temporarily exceed the configured token limit”. Set each cap below the amount you cannot exceed, and do not treat it as a substitute for attribution. What the Monthly Cost Report Should Show The board questions on AI risk https://reinvently.co.uk/blog/what-boards-should-ask-about-ai-risk/ ask what AI information reaches your board. For cost, the budget owner needs five lines a month: - Spend by owner against budget, with any spend that cannot be attributed shown as its own line. That line should shrink every month. - Cost per outcome for the largest use cases, and its trend. - Share of spend on frontier models, which shows whether routing is working. - Vendor price and default-model changes since the last report, with their estimated effect. - Forecast at the current run rate, and which hard limits would be reached first. Billing data gives you the first and last lines. The middle three need the attribution from control 1 and an agreed definition of each outcome. Your First Month 1. Week 1: find every place AI is bought. List API accounts, seats, marketplace subscriptions and AI features inside existing contracts. Deal with shared and personal API keys first. 2. Week 2: switch on hard limits for agents and non-production. Check each platform against the table, and replace alert-only budgets wherever a stopping control exists. 3. Week 3: model next year at today’s prices and at double. Start with the AI cost calculator https://reinvently.co.uk/tools/ai-cost-calculator/ for coding tools. If you cannot afford the doubled figure, the routing and contract controls are urgent. 4. Week 4: agree one cost-per-outcome measure for each of your three largest use cases, and report it from next month. Start with attribution. Every other control depends on knowing whose spend it is. Sources - KPMG, AI Quarterly Pulse Survey, Q2 2026, as analysed in our agent governance gap https://reinvently.co.uk/blog/ai-agent-governance-gap/ piece. - Reinvently AI coding price tracker https://reinvently.co.uk/tools/ai-coding-tool-prices/ , change log entries dated 28 September 2026. - The Pragmatic Engineer survey and Gartner forecast, as cited in our AI coding ROI analysis https://reinvently.co.uk/blog/ai-coding-spend-roi-per-engineer/ . - FinOps Foundation, Tokenomics: Managing AI Value in SaaS Model Token Costs https://www.finops.org/wg/token-economics-saas/ , last updated 3 June 2026. - OpenAI, Spend limits https://developers.openai.com/api/docs/guides/spend-limits ; Anthropic, Spend Limits API https://platform.claude.com/docs/en/api/beta/organization/spend limits ; Microsoft, llm-token-limit policy https://learn.microsoft.com/en-us/azure/api-management/llm-token-limit-policy ; Microsoft, Create and manage budgets https://learn.microsoft.com/en-us/azure/cost-management-billing/costs/tutorial-acm-create-budgets ; AWS, Application inference profiles https://docs.aws.amazon.com/bedrock/latest/userguide/cost-mgmt-application-inference-profiles.html ; AWS, Configuring budget actions https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-controls.html ; Google Cloud, Manage spend cap budgets https://docs.cloud.google.com/billing/docs/how-to/budgets-spend-caps ; GitHub, Setting up budgets https://docs.github.com/en/billing/how-tos/set-up-budgets . All checked 23 October 2026. This is the second of five decision pieces for technology leaders. Get the next ones by email https://reinvently.co.uk/contact/ . ← All posts https://reinvently.co.uk/blog/ Editorial note, 23 October 2026: Provider controls change often. The table reflects each provider’s public documentation on the date checked; confirm current behaviour in your own billing settings before relying on a limit.