AI budget planning for 2027 cannot work by extending the current bill. User adoption, use cases, model mix, price versions and architecture often change at once, and the invoice records none of those drivers. The useful output is not a falsely precise annual figure but a driver-based model: it shows which assumption moves the number, who evidences it and when it will be reviewed.

Why last year's spend is not a baseline

For mature infrastructure, last year's run rate plus known changes gets you a long way. AI spend lacks the conditions that make that work. The history is short and often dominated by experiments that never reach production. Consumption per transaction depends on engineering choices such as context length, response length, caching and routing, which move with every release. And providers change models and price tiers often enough that the same transaction may land on a different rate next year.

The FinOps Foundation™'s State of FinOps 2026 survey reflects this: nearly all respondents to the relevant question say they now manage AI spend, and the report names visibility, allocation and proving value as the main challenges. Its sample leans towards large budgets, but the conclusion for a small team is the same: plan from the driver, not from the invoice.

Set the budget boundary first

Separate three pools before estimating volume:

  • Experiments receive a spending ceiling and an end date. They fund learning, not indefinite operations.
  • Production applications are driven by demand and consumption: users, transactions, input, output and model choice.
  • Shared platform covers costs that do not belong neatly to one request, such as evaluation, vector search, guardrails, logging, network or provisioned capacity.

Record where each charge appears. Amazon Bedrock, Azure OpenAI in Microsoft Foundry, and Google Cloud's Gemini Enterprise Agent Platform (formerly Vertex AI) bill through their cloud accounts. On AWS, third-party models on Bedrock appear under the AWS Marketplace billing entity, so a budget filtered to the AWS entity alone will miss them. Models bought directly from a model provider sit on separate invoices. A budget built only from cloud exports would then miss part of AI spend.

This split avoids two common mistakes: a successful pilot does not quietly become a production baseline, and shared platform work does not disappear inside an attractive cost-per-request figure.

The 2027 AI budgetExperiments: a spendingceiling and an end dateProduction: users, transactions,tokens and model choiceShared platform: evaluation,search, guardrails, loggingEach charge recorded where it lands: the cloud account,the Marketplace billing entity or a modelprovider's own invoice
Three pools before any volume estimate: experiments with a ceiling and an end date, production driven by demand and consumption, shared platform cost that belongs to no single request

Inputs to gather before the first number

Collect four kinds of evidence before you calculate:

  1. Demand from the product plan: planned users, enabled seats, transactions per month and rollout dates. The invoice shows consumption, not the rollout schedule.
  2. Measured token shape: input, output, cache-read and cache-write tokens per transaction from response usage fields or provider metrics. Bedrock publishes InputTokenCount, OutputTokenCount and cache counters in CloudWatch; Azure OpenAI reports cached_tokens in its usage details.
  3. Dated prices: model, region, purchase model, currency and review date for every rate. Output tokens cost more than input on many models, and cache writes can be priced above standard input depending on the model.
  4. Contract terms: credits, negotiated discounts, existing reservations and their end dates.

If product telemetry is immature, choose a named proxy such as processed cases and set a date for replacing it.

The worksheet: turn demand into cost

The sheet holds inputs and owners, not just monthly totals. Use one row per use case with at least these fields:

FieldWhat to recordEvidence owner
Purpose and phaseexperiment, rollout or steady productionproduct owner
Demandactive users, transactions or requests per periodproduct or business team
Model calls per transactionincluding retries, tool steps and fallbacksengineering
Input per calltext, image, audio or video units, including repeated contextengineering
Output per callexpected response length, reasoning share and modalityengineering
Model mixshare by model and the fallback or routing ruleengineering
Reusemeasured cache-read and cache-write sharesengineering
Purchase modelpay-as-you-go, batch, flex or provisioned capacity and termplatform and procurement
Supporting costretrieval, storage, network, evaluation, logging and safeguardsplatform
Billing locationcloud account, Marketplace entity or separate invoicefinance
Price sourceprovider, region, currency, price page and review dateFinOps or finance
Uncertaintyrange, reforecast trigger and ownerjoint decision

For usage-priced models, the structure per model is:

transactions × calls per transaction × (uncached input × input price + cache reads × cache-read price + cache writes × cache-write price + output × output price)

Add batch and multimodal rates only where the architecture will use them. A single average token rate hides that input and output are priced differently and that a model change alters both. The calls-per-transaction column is where agent and retry workloads are most often underestimated.

Use three scenarios, not one precise guess

Build base, lower-usage and higher-usage cases with the same formulas, changing only the assumptions that are genuinely uncertain. That shows whether the range comes from demand, response length, model choice or the capacity decision.

Each scenario carries five things into approval:

  1. the business demand driver;
  2. the technical consumption assumption;
  3. the current price source and currency;
  4. non-usage platform cost;
  5. the decision that follows if actuals leave the range.

Keep credits and negotiated discounts below gross cost: they change the invoice, not the demand, and they expire. Treat exchange rates the same way. For UK companies reporting in sterling, or groups consolidating in euros, record the conversion policy and review date and keep the local-currency view beside the group view, so a currency movement never reads as an engineering variance.

Treat provisioned capacity as a separate decision

Pay-as-you-go follows consumption and suits variable or unproven demand. Provisioned capacity exchanges that flexibility for a fixed cost block that accrues whether or not you use it. Terms differ materially:

OfferBillingTermWhat it means for the budget
Bedrock Provisioned Throughputhourly per model unit until deletedno commitment, 1 or 6 monthsfixed line from purchase date
Bedrock Reserved tierfixed price per 1,000 tokens per minute, billed monthly1 or 3 months; minimum 100,000 input and 10,000 output TPMfixed line plus overflow at Standard rates
Azure provisioned (PTU)hourly per PTU regardless of tokenshourly, or a 1-month or 1-year reservationa reservation fixes the rate, not the capacity
Google Provisioned Throughput (GSU)fixed price per term1 week (Google models only), 1, 3 or 12 monthsfees apply regardless of usage; unused throughput does not carry over; overage is billed pay-as-you-go by default

Do not convert a pay-as-you-go estimate directly into a capacity order. Measure input, output, concurrency, peak shape and cache share for the actual workload; Microsoft notes, for example, that cached tokens do not consume PTU capacity. A commitment belongs in the baseline only when demand and technical shape are stable. Until then, keep it as a dated decision with an owner and an entry condition.

Plan non-interactive work such as evaluations and backfills separately. Bedrock offers batch on select models at 50 percent below on-demand, Azure Global Batch at 50 percent below Global Standard, and Google batch at 50 percent below real-time inference. Those rates apply only to the share that can genuinely run asynchronously.

How budget alerts apply to AI

Budget alerts report a variance; they do not control consumption. AWS updates budget data up to three times a day, Azure evaluates budgets every 24 hours with cost data typically available within 8 to 24 hours, and Google documents a delay between usage and cost reporting. That is adequate for an annual budget, not for a fault that burns tokens within hours.

Set two kinds of threshold: budget alerts per pool and scenario for financial steering, and technical limits inside the application for operations. Since 27 July 2026 Google has offered spend cap budgets in Preview, which pause new usage of a single service, such as the Gemini API or Agent Platform, in a single project. They trigger on estimated gross cost, enforcement takes effect with a delay, requests already in flight still complete and accrue charges, and overages are billed as normal. Treat a cap as an emergency brake, not a budgeting method.

A fault burns tokens within hoursBudget alert per pool and scenario: a variancereport, refreshed up to three times a day in AWS,every 24 hours in AzureTechnical limit inside the application:controls consumption as it happensGoogle spend cap, Preview: pauses one servicein one project after a delay; an emergencybrake, not a budgeting method
Two kinds of threshold: a budget alert reports a variance a day later and steers the annual budget; a technical limit in the application stops a fault within the hour

Failure cases

  • An average token price replaces separate input, output and cache rates.
  • An experiment runs past its end date and becomes silent production load.
  • Credits are netted against gross cost and hide consumption until they expire.
  • A capacity order is sized from a pay-as-you-go extrapolation without load measurement.
  • Marketplace charges or direct API invoices are left out of scope.
  • A new model version goes live without re-measuring price and token shape.

Reforecast when the workload changes

Avoid an arbitrary meeting cadence. Define the events that justify a new estimate:

  • a model or price version enters production;
  • users, transactions or token shape leave the approved range;
  • caching, routing or output limits change consumption per transaction;
  • an experiment closes or becomes a production service;
  • provisioned capacity is purchased, renewed or left underused.

Compare actuals and forecast on the same cost basis. The accountable owner needs the driver, model, phase and latest assumption next to the amount. Unit economics in the glossary explains how to choose the denominator for cost per unit of value. To read the individual meters, see Azure OpenAI and Bedrock charges; for attribution, allocate AI cost to a product feature; for the rest of the cloud plan, build a 2027 cloud budget from usage and planned changes.

What Costfluent covers

Costfluent reads AWS, Azure and Google Cloud billing data through read-only connectors, including the AI services billed there, and shows cost by account, service, resource, tag and segment. You set a budget for the workspace in the product, or for one cost segment through the Public API and the Terraform provider, and compare actual spend with it for its period. The steps are in the docs under Budgets. A saved cost report can be emailed on the schedule your review follows. A cost report can show a spend forecast, and a budget shows where its current period is heading. Budget alert notifications are Planned. API spend with OpenAI and Anthropic joins once you connect those providers directly (Track OpenAI and Anthropic API spend). Costfluent does not read product metrics such as transactions, so the driver half of the worksheet stays with you.

Sources and method

Checked on 2026-10-10. FinOps Foundation: State of FinOps 2026, Amazon Bedrock: Provisioned Throughput, Amazon Bedrock: service tiers, Amazon Bedrock: how tokens are counted, Amazon Bedrock: prompt caching, Amazon Bedrock pricing, AWS Cost Anomaly Detection (Marketplace billing of third-party Bedrock models), AWS Budgets, Microsoft Foundry: provisioned throughput, Azure OpenAI: prompt caching, Azure OpenAI: global batch, Microsoft Cost Management: budgets, Google Cloud: purchase Provisioned Throughput, Google Cloud: calculate Provisioned Throughput, Google Cloud: use Provisioned Throughput, Google Cloud: batch inference, Google Cloud: budgets, Google Cloud: spend cap budgets, Cloud Billing release notes and Google Cloud: Agent Platform name changes. The worksheet, formula, capacity table and review triggers are Costfluent recommendations, not accounting advice. Record prices with their region, currency and review date in your own budget; this guide deliberately uses no example amounts.