# Forecasting and controlling enterprise AI/LLM costs

**Executive answer.** AI cost is moving from a simple “per-seat software subscription” budget line to a **hybrid of seats + metered credits + token consumption + agent runtime**, and the leadership implication is that CFOs should treat enterprise AI like cloud FinOps, not like a traditional SaaS renewal. Microsoft now describes usage-based AI billing as charges based on actual usage measured in Copilot Credits, complementing fixed subscription licensing; OpenAI’s Business and Enterprise/Edu plans now use flexible credits for advanced models and features; Anthropic’s Team plan combines fixed seats with usage credits and API-rate consumption paths for heavier work ([Microsoft usage-based billing](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-overview-copilot-credits), [OpenAI flexible pricing](https://help.openai.com/en/articles/11487671-flexible-pricing-for-the-enterprise-edu-and-business-plans), [Anthropic Team billing](https://support.anthropic.com/en/articles/9266767-claude-team-plan-set-up-and-billing)).

The most important board-level distinction is between **human-paced chat** and **machine-paced agents**: a chat user can only ask so many questions in a day, while a 24/7 agent can call models thousands to tens of thousands of times per month, and its bill scales with frequency, context size, tools, model choice, and retry loops. Microsoft’s Copilot Cowork pricing explicitly uses four task inputs—model use, context retrieval, tool calls, and runtime—while OpenAI says Codex task usage varies with size and complexity of the work, model, where the task runs, number of instances, automations, and fast mode ([Microsoft Copilot Cowork GA](https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/), [OpenAI Codex rate card](https://help.openai.com/en/articles/20001106-codex-rate-card)).

## 1. What leaders can understand now

### The four pricing models in plain English

| Model | Plain-English description | What to watch |
|---|---|---|
| Per-seat subscription | You pay a fixed monthly price per named user, usually with admin/security features and some included usage. | High adoption makes it predictable; low adoption creates shelfware. |
| Token/API pricing | You pay for input and output text chunks; output tokens are usually much more expensive than input tokens. | Long prompts, long context, verbose outputs, and premium models drive cost. |
| Credit systems | Vendor-specific “points” are deducted for messages, tasks, model calls, tool calls, or tokens. | Credits can hide unit cost unless finance converts every credit to dollars. |
| Subscription + credits | Seats include baseline access, while advanced usage draws from pooled credits or overages. | This is becoming the dominant enterprise pattern because it caps routine work but monetizes heavy or autonomous use. |

ChatGPT Business standard seats are listed at **$25 per user per month monthly** or **$20 per user per month billed annually**, with a minimum purchase of **2 standard ChatGPT seats**; OpenAI says the April 2, 2026 pricing change reduced standard ChatGPT seats by **$5 per month** ([OpenAI Business billing](https://help.openai.com/en/articles/8792536-managing-billing-and-seats-in-chatgpt-business)). Microsoft originally announced Microsoft 365 Copilot at **$30 per user per month** for eligible Microsoft 365 E3, E5, Business Standard, and Business Premium customers, and its 2026 partner guidance shows promotional standalone Microsoft 365 Copilot Business at **$18 per user per month after a 15% promo** through December 31, 2026 for 1–300 users ([Microsoft 2023 Copilot pricing announcement](https://www.microsoft.com/en-us/microsoft-365/blog/2023/07/18/introducing-bing-chat-enterprise-microsoft-365-copilot-pricing-and-microsoft-sales-copilot/), [Microsoft Partner Center June 2026 announcement](https://learn.microsoft.com/en-us/partner-center/announcements/2026-june)). Anthropic’s Claude Team plan has **Standard seats at $25 monthly or $20 annually per member** and **Premium seats at $125 monthly or $100 annually per member**, with a five-member minimum and up to 150 seats ([Anthropic Team billing](https://support.anthropic.com/en/articles/9266767-claude-team-plan-set-up-and-billing)).

API pricing shows why model choice matters: OpenAI lists GPT-5.5 at **$5 per 1M input tokens** and **$30 per 1M output tokens**, GPT-5.4-mini at **$0.75 input** and **$4.50 output**, and GPT-5.5-pro at **$30 input** and **$180 output** for short context ([OpenAI API pricing](https://platform.openai.com/docs/pricing)). Anthropic lists Claude Sonnet 5 introductory pricing at **$2 per 1M input tokens** and **$10 per 1M output tokens** through August 31, 2026, moving to **$3/$15** afterward; Opus 4.8 is **$5/$25**, and Fable 5 is **$10/$50** ([Anthropic pricing](https://www.anthropic.com/pricing), [Anthropic Sonnet 5 announcement](https://www.anthropic.com/news/claude-sonnet-5)).

Credit systems are not all equivalent. OpenAI’s ChatGPT Business/Enterprise rate card lists **GPT-5.6 Sol at 10 credits per message**, **GPT-5.6 Sol Pro at 50 credits per message**, **Agent mode at 30 credits per message**, **Deep research at 50 credits per task**, **Images at 5 credits per generation**, and **Voice at 5 credits per minute** ([OpenAI ChatGPT rate card](https://help.openai.com/en/articles/11481834-chatgpt-rate-card)). Microsoft Copilot Studio rates include **1 Copilot Credit for a classic answer**, **2 for a generative answer**, **5 for an agent action**, **10 for tenant graph grounding**, and **10 Copilot Credits per 1K tokens** for premium text/generative AI tools when using reasoning-capable language models ([Microsoft Copilot Studio billing rates](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management)).

### What actually drives cost

The first driver is **tokens**: language and vision models process inputs by breaking them into tokens, and token-based metering can vary by model, deployment type, and meter ([Microsoft Foundry cost management](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/manage-costs)). The second driver is **model tier**: OpenAI’s short-context GPT-5.5-pro output price is 40x GPT-5.4-mini output price, using $180/M versus $4.50/M output tokens ([OpenAI API pricing](https://platform.openai.com/docs/pricing)). The third driver is **context length**: Claude Sonnet 5 supports a 1M-token context window on paid chat plans, and Claude Code supports 1M-token windows for several paid-plan models, so users and agents can carry much larger context than earlier chat tools ([Anthropic context window help](https://support.anthropic.com/en/articles/8606394-how-large-is-claude-pro-s-context-window)). The fourth driver is **agent behavior**: Microsoft’s Cowork task price is calculated from model use, context retrieval, tool calls, and runtime, and OpenAI says Codex costs vary with model used, number of instances, automations, and fast mode ([Microsoft Copilot Cowork GA](https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/), [OpenAI Codex rate card](https://help.openai.com/en/articles/20001106-codex-rate-card)).

## 2. What to expect from Microsoft, OpenAI, and Anthropic

### Microsoft: from Copilot seats to a Copilot Credit economy

Microsoft is keeping the familiar Microsoft 365 Copilot seat while expanding usage-based billing for agents and data-grounded APIs. Microsoft says usage-based billing in Microsoft 365 charges actual usage in Copilot Credits, supports prepaid credits and pay-as-you-go, and includes controls for budgets, alerts, hard caps, users, groups, services, and agents ([Microsoft usage-based billing](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-overview-copilot-credits), [Microsoft Copilot credit management](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-manage-copilot-credits)).

Microsoft’s June 2026 Copilot Credit announcement says PayGo is available with no upfront purchase at **$0.01**, capacity packs cost **$200 per tenant per month for 25,000 Copilot Credits**, the Microsoft 365 Copilot User License is **$30 per user/month**, and the P3 Pre-Purchase Plan is a one-year pay-up-front pool usable across Copilot Studio, Dynamics 365 first-party agents, and Copilot Chat ([Microsoft Copilot Credit pre-purchase announcement](https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/scale-your-agent-rollout-with-confidence-introducing-copilot-credit-pre-purchase-plan/)). Microsoft’s Copilot Credits Guide lists PayGo at **$0.01 per Copilot Credit** and P3 tier discounts ranging from **5% at 300,000 credits** to **20% at 300,000,000 credits**, and it says unused P3 credits expire at the end of the annual term ([Microsoft Copilot Credits Guide](https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/bade/documents/products-and-services/en-us/ai/Microsoft-Copilot-Credits-Guide-June-16-2026-PUB.pdf)).

Copilot Cowork is the clearest signal of future agent billing: it requires a Microsoft 365 Copilot User Subscription License, then bills tasks on usage in Copilot Credits, with each task priced from model use, context retrieval, tool calls, and runtime ([Microsoft Copilot Cowork GA](https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/)). Microsoft says PayGo for Cowork is **$0.01 per Copilot Credit**, users see the credit price for each task as they run it, and admins can set spending limits at tenant, group, and user levels ([Microsoft Copilot Cowork GA](https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/)). Microsoft’s Copilot Credits Guide gives example Cowork ranges of **70–200 credits** for a light weekly status-update task, **400–600 credits** for a medium customer-meeting briefing, and **more than 1,500 credits** for a heavy six-month product-usage analysis ([Microsoft Copilot Credits Guide](https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/bade/documents/products-and-services/en-us/ai/Microsoft-Copilot-Credits-Guide-June-16-2026-PUB.pdf)).

### OpenAI: Business seats plus flexible credits for advanced and agentic use

OpenAI’s current Business plan remains a predictable per-seat product at **$20 annual / $25 monthly per user**, but OpenAI explicitly says credits can extend included limits for advanced models and features once included limits are exceeded ([OpenAI Business billing](https://help.openai.com/en/articles/8792536-managing-billing-and-seats-in-chatgpt-business), [OpenAI flexible pricing](https://help.openai.com/en/articles/11487671-flexible-pricing-for-the-enterprise-edu-and-business-plans)). OpenAI’s flexible pricing article says Business users receive per-seat limits for advanced features, Enterprise/Edu users have no per-seat caps by default, Enterprise/Edu credits are purchased and allocated at the contract level, Business credits are pooled after purchase by workspace owners, and Business credits are valid for 12 months after purchase ([OpenAI flexible pricing](https://help.openai.com/en/articles/11487671-flexible-pricing-for-the-enterprise-edu-and-business-plans)).

OpenAI is also moving coding and workspace-agent work toward token-based credits. OpenAI’s Codex rate card says Codex pricing shifted on April 2, 2026 to align with API token usage instead of per-message pricing for new and existing Plus, Pro, ChatGPT Business, and new Enterprise plans, and the change applied to all existing Enterprise plans on April 23, 2026 ([OpenAI Codex rate card](https://help.openai.com/en/articles/20001106-codex-rate-card)). OpenAI states that Codex costs average **about $100–$200 per developer per month**, with large variance depending on model, number of instances, automations, and fast mode ([OpenAI Codex rate card](https://help.openai.com/en/articles/20001106-codex-rate-card)).

OpenAI’s June 2026 spend-controls announcement is a governance signal: Enterprise admins can see ChatGPT and Codex credit usage in one Global Admin Console view, break down credit spend by user, product, and model, use a unified Cost API, set default workspace limits, configure group limits, and create individual overrides ([OpenAI Enterprise spend controls](https://openai.com/index/chatgpt-enterprise-spend-controls/)). OpenAI’s Enterprise/Edu usage-limit help says overages occur when eligible workspaces continue using credit-based features after the committed credit pool is exhausted, and overage amounts are based on credits used above purchased or granted credits multiplied by the contracted overage rate ([OpenAI Enterprise usage limits](https://help.openai.com/en/articles/20001001-setting-usage-limits-in-chatgpt-enterprise-and-edu)).

### Anthropic: fixed seats for teams, usage credits for high-end use, API economics for builders

Anthropic’s Team plan already has a two-tier seat design—Standard and Premium—and it supports spend controls and prepaid usage credits so team members can continue after reaching included usage limits ([Anthropic Team billing](https://support.anthropic.com/en/articles/9266767-claude-team-plan-set-up-and-billing)). Anthropic’s public API pricing is transparent by model, and its usage tiers cap monthly API spend at **$500 for Start**, **$1,000 for Build**, **$200,000 for Scale**, and **no monthly spend cap for Custom** ([Anthropic API rate limits](https://docs.anthropic.com/en/api/rate-limits)).

Anthropic’s direction is toward more agentic and workplace-integrated products that need token spend limits. Claude Tag is in beta for Claude Enterprise and Team customers, works with Opus 4.8, lets administrators set token spend limits for the organization and individual Slack channels, and provides logs of what @Claude did and who requested each task ([Anthropic Claude Tag](https://www.anthropic.com/news/introducing-claude-tag)). Anthropic’s Fable 5 announcement says Fable 5 is fully available on the Claude API and consumption-based Enterprise plans, while subscription-plan use after the included window requires usage credits ([Anthropic Fable 5 announcement](https://www.anthropic.com/research/claude-fable-5-mythos-5)).

## 3. Unit economics worked examples

### Example A — Seat economics for 100 employees

| Product | Unit price assumption | 100-user annualized cost | Interpretation |
|---|---:|---:|---|
| ChatGPT Business annual | $20/user/month | $24,000/year | Predictable if adoption is broad; excludes API billing and may need credits for advanced usage. |
| ChatGPT Business monthly | $25/user/month | $30,000/year | Higher flexibility, 25% more than annual annualized price. |
| Microsoft 365 Copilot list | $30/user/month | $36,000/year | Most relevant for Microsoft 365-centric workforces. |
| Microsoft 365 Copilot Business promo | $18/user/month | $21,600/year | Promotional and channel-dependent; confirm term and renewal pricing. |
| Claude Team Standard annual | $20/member/month | $24,000/year | Comparable to ChatGPT Business annual for standard seats. |
| Claude Team Premium annual | $100/member/month | $120,000/year | Use selectively for power users rather than everyone. |

The ChatGPT Business, Microsoft 365 Copilot, and Claude Team numbers above come from current vendor pages and support articles, while the 100-user totals are simple annualized calculations from those unit prices ([OpenAI Business billing](https://help.openai.com/en/articles/8792536-managing-billing-and-seats-in-chatgpt-business), [Microsoft 2023 Copilot pricing announcement](https://www.microsoft.com/en-us/microsoft-365/blog/2023/07/18/introducing-bing-chat-enterprise-microsoft-365-copilot-pricing-and-microsoft-sales-copilot/), [Microsoft Partner Center June 2026 announcement](https://learn.microsoft.com/en-us/partner-center/announcements/2026-june), [Anthropic Team billing](https://support.anthropic.com/en/articles/9266767-claude-team-plan-set-up-and-billing)).

### Example B — API cost per “large knowledge task”

Assume a task reads **1,000,000 input tokens** and produces **250,000 output tokens**. Using published short-context or standard API prices, that single task costs about **$1.88 on GPT-5.4-mini**, **$12.50 on GPT-5.5**, **$75.00 on GPT-5.5-pro**, **$4.50 on Claude Sonnet 5 introductory pricing**, **$11.25 on Claude Opus 4.8**, and **$22.50 on Claude Fable 5** ([OpenAI API pricing](https://platform.openai.com/docs/pricing), [Anthropic pricing](https://www.anthropic.com/pricing)).

| Model | Math | Approx. cost |
|---|---:|---:|
| GPT-5.4-mini | $0.75 + 0.25 × $4.50 | $1.88 |
| GPT-5.5 | $5.00 + 0.25 × $30.00 | $12.50 |
| GPT-5.5-pro | $30.00 + 0.25 × $180.00 | $75.00 |
| Claude Sonnet 5 intro | $2.00 + 0.25 × $10.00 | $4.50 |
| Claude Opus 4.8 | $5.00 + 0.25 × $25.00 | $11.25 |
| Claude Fable 5 | $10.00 + 0.25 × $50.00 | $22.50 |

The practical point is not that one model is always cheaper; it is that **routing a routine task to a premium model can be a 10x–40x decision** on that task’s token bill. OpenAI’s GPT-5.5-pro output price is **$180/M**, while GPT-5.4-mini output is **$4.50/M**, and Anthropic’s Fable 5 output is **$50/M** versus Sonnet 5 introductory output at **$10/M** ([OpenAI API pricing](https://platform.openai.com/docs/pricing), [Anthropic pricing](https://www.anthropic.com/pricing)).

### Example C — A 24/7 agent cost curve

Assume an agent checks a work queue every five minutes, or **8,640 runs per 30-day month**. If each run uses **5,000 input tokens** and **500 output tokens**, GPT-5.4-mini costs about **$52/month** and GPT-5.5 costs about **$346/month** before orchestration, storage, retrieval, and tool costs, using OpenAI’s published token rates ([OpenAI API pricing](https://platform.openai.com/docs/pricing)).

If the same agent carries **100,000 input tokens** of context and produces **2,000 output tokens** every five minutes, GPT-5.5 costs about **$4,838/month** before other costs. If that 100,000-token context is cacheable at GPT-5.5’s cached-input price of **$0.50/M**, the token cost falls to about **$950/month**, which illustrates why cache design can matter more than vendor negotiation for agentic workloads ([OpenAI API pricing](https://platform.openai.com/docs/pricing)). AWS says prompt caching can reduce inference latency and input token costs for repeated long contexts, and Google says context caching reduces cost and latency for repeated Gemini content with 90% discounts on cached tokens for relevant models ([AWS Bedrock prompt caching](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html), [Google Vertex AI context caching](https://cloud.google.com/vertex-ai/generative-ai/docs/context-cache/context-cache-overview)).

If that long-context GPT-5.5 agent runs every minute instead of every five minutes, the same uncached token math rises to about **$24,192/month**, because the monthly run count increases to **43,200**. This is the non-linear executive risk: once agents are autonomous, “one more workflow” can mean thousands of additional model calls, tool calls, retries, and context-retrieval events.

### Example D — Microsoft agent credits

A Copilot Studio interaction that uses a generative answer plus tenant graph grounding consumes **12 Copilot Credits** under the published rates: **2 credits** for a generative answer plus **10 credits** for tenant graph grounding ([Microsoft Copilot Studio billing rates](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management)). At **$0.01 per Copilot Credit**, that pattern is roughly **$0.12 per interaction**, or **$12,000 for 100,000 interactions**, before any additional AI-tool, voice, or action charges ([Microsoft Copilot Cowork GA](https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/)). With a **$200 capacity pack for 25,000 credits**, the effective capacity-pack rate is **$0.008 per credit**, so the same 1.2M-credit workload would be **$9,600** if entirely covered by capacity packs and if terms remain as published ([Microsoft Copilot Credit pre-purchase announcement](https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/scale-your-agent-rollout-with-confidence-introducing-copilot-credit-pre-purchase-plan/)).

## 4. Planning framework for CEOs, CHROs, CFOs, CTOs, CISOs, and CIOs

### Build the budget in five layers

1. **Seat layer:** committed ChatGPT Business/Enterprise, Microsoft 365 Copilot, Claude Team/Enterprise, and role-based premium seats.  
2. **Human overage layer:** credits for heavy users, advanced models, deep research, voice, images, and coding.  
3. **Workflow layer:** assistants and departmental agents with expected task volume, average tokens, tools, and retrieval calls.  
4. **Autonomous layer:** 24/7 agents with frequency, runtime, retries, monitoring, sandboxing, and kill-switch budgets.  
5. **Governance layer:** logging, data loss prevention, security review, evaluation, red teaming, training, and FinOps reporting.

This layering aligns with FinOps Foundation guidance to establish baselines from invoices and usage reports, track AI costs and usage, set quotas, tag resources, optimize model selection, implement showback, and use dashboards and alerts to make cost information actionable ([FinOps Foundation AI overview](https://www.finops.org/wg/finops-for-ai/)).

### Use a CFO-readable unit-economics model

For every AI use case, require a one-page unit model:

`Monthly cost = fixed seats + (tasks × model calls/task × tokens/call × price/token) + platform credits + tool/retrieval costs + cloud/orchestration + monitoring/security.`

The model should report **cost per employee**, **cost per active user**, **cost per task**, **cost per successful task**, **cost per avoided hour**, and **cost per business outcome**. Microsoft’s Cost Management dashboard supports drill-down by user, group, service, or agent and provides budgets, alerts, and hard caps, while OpenAI Enterprise provides breakdowns by user, product, and model plus a Cost API ([Microsoft usage-based billing](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-overview-copilot-credits), [OpenAI Enterprise spend controls](https://openai.com/index/chatgpt-enterprise-spend-controls/)).

### Pilot-to-scale cost curve

| Stage | Budget rule | What to measure | Decision gate |
|---|---|---|---|
| 0. Discovery | Time-box and cost-cap experiments. | Prompt volume, active users, task categories. | Does this replace, compress, or improve work? |
| 1. Pilot | Fund seats plus a small shared credit pool. | Cost/task, success rate, user adoption, risk incidents. | Is the unit cost below the value of the task? |
| 2. Department rollout | Add showback by department. | Power users, top workflows, credit burn, failed/retried tasks. | Should usage be seat-based, credit-based, or API-based? |
| 3. Enterprise scale | Negotiate committed pools, overage limits, and data terms. | Cost per outcome, utilization, model mix, vendor concentration. | Which workloads deserve premium models? |
| 4. Autonomous systems | Separate budget, SLOs, kill switches, and incident processes. | Run frequency, context size, tool calls, runtime, retry storms. | Can the system fail safely and stop spending? |

Microsoft recommends budgets, alerts, spending limits, user ceilings, and service/agent access policies for Copilot Credits, and OpenAI lets Enterprise/Edu admins configure workspace limits, group limits, user overrides, and overage limits ([Microsoft Copilot credit management](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-manage-copilot-credits), [OpenAI Enterprise usage limits](https://help.openai.com/en/articles/20001001-setting-usage-limits-in-chatgpt-enterprise-and-edu)).

### Procurement strategy

For **100–300 employee organizations**, start with limited Business seats and selective Microsoft 365 Copilot seats unless Microsoft 365 is the daily work hub, then add credits only after adoption and task economics are visible. ChatGPT Business has a two-seat minimum and self-serve card billing, while Microsoft’s 2026 partner guidance includes 1–300-user Copilot Business promotional terms and no license/user minimum for certain SMB offers ([OpenAI Business billing](https://help.openai.com/en/articles/8792536-managing-billing-and-seats-in-chatgpt-business), [Microsoft Partner Center June 2026 announcement](https://learn.microsoft.com/en-us/partner-center/announcements/2026-june)).

For **300–5,000 employee organizations**, negotiate enterprise terms around credit pools, overage rates, admin analytics, security reviews, and data residency rather than only per-seat discounts. OpenAI says Enterprise credit allocation, expiration, rates, eligible usage, and billing terms are defined in the agreement, and Microsoft says P3 pre-purchase credits provide discounted rates versus pay-as-you-go with automatic continuation into pay-as-you-go after prepaid credits are exhausted ([OpenAI Enterprise usage limits](https://help.openai.com/en/articles/20001001-setting-usage-limits-in-chatgpt-enterprise-and-edu), [Microsoft Copilot credit management](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-manage-copilot-credits)).

For **5,000–40,000 employee organizations**, create an AI consumption committee spanning finance, IT, security, legal, procurement, and business owners. FinOps Foundation recommends cross-functional collaboration across leadership, data science, engineering, finance, procurement, product management, and cloud architecture for AI cost governance ([FinOps Foundation AI overview](https://www.finops.org/wg/finops-for-ai/)).

## 5. Cost strategy across the four escalation tiers

### Tier A — Chats

Chats are human-paced and easiest to budget by seat. The control levers are adoption tracking, unused-seat reclamation, role-based rollout, basic model defaults, and training users to ask concise questions.

Best controls: reclaim inactive seats monthly; route routine chats to default/mini models; reserve premium models for executive, analytical, legal, coding, or research tasks; publish a “what model for what work” guide. OpenAI says ChatGPT may route lightweight tasks to mini models and use **0 credits** when a mini model handles the request, while FinOps Foundation warns against using the most complex and expensive models for every task ([OpenAI flexible pricing](https://help.openai.com/en/articles/11487671-flexible-pricing-for-the-enterprise-edu-and-business-plans), [FinOps Foundation AI overview](https://www.finops.org/wg/finops-for-ai/)).

### Tier B — Assistants

Assistants combine chat with files, connectors, retrieval, spreadsheets, presentations, and coding help. The cost shift is from seat count to usage intensity, because context, files, and connectors can increase tokens even when the user count is flat.

Best controls: set per-user credit budgets; use retrieval instead of pasting entire documents; cache repeated instructions and large reference material; maintain department-level showback; and approve connectors based on data sensitivity and cost. Google says context caching is well suited to repeated large context, recurring queries against large document sets, and frequent code repository analysis, while AWS says prompt caching helps with long repeated contexts and reduces input token costs ([Google Vertex AI context caching](https://cloud.google.com/vertex-ai/generative-ai/docs/context-cache/context-cache-overview), [AWS Bedrock prompt caching](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html)).

### Tier C — Agents

Agents execute multi-step tasks and call tools, so cost depends on steps, retries, tool calls, context retrieval, and model tier. Microsoft’s Copilot Studio rates separately charge features such as generative answers, agent actions, graph grounding, flow actions, and premium token-based tools, while OpenAI’s Workspace Agent rate card prices runs by input, cached input, and output credits rather than a fixed task fee ([Microsoft Copilot Studio billing rates](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management), [OpenAI ChatGPT rate card](https://help.openai.com/en/articles/11481834-chatgpt-rate-card)).

Best controls: require a cost estimate before publishing each agent; set maximum steps, maximum tool calls, maximum tokens, and maximum runtime; route simple steps to smaller models; use approval gates before external actions; log cost by task and task outcome; and give every agent an owner and monthly budget.

### Tier D — Agentic systems and 24/7 autonomy

Agentic systems operate continuously, watch queues, call other systems, maintain memory, and trigger workflows, so the cost curve is closer to cloud infrastructure than SaaS. Microsoft’s Cowork model explicitly includes runtime in task pricing, and Azure Foundry warns that hosted fine-tuned models can incur hourly hosting costs even if unused ([Microsoft Copilot Cowork GA](https://www.microsoft.com/en-us/microsoft-365/blog/2026/06/16/copilot-cowork-is-now-generally-available/), [Microsoft Foundry cost management](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/manage-costs)).

Best controls: ban unlimited loops; set hard monthly spend caps; require circuit breakers for error spikes; cap retries; separate dev/test/prod budgets; monitor tokens per successful task; page humans on anomalous spend; and use explicit “sleep schedules” for agents that do not need true 24/7 operation. Microsoft’s Copilot credit management supports monthly policy budgets, per-user ceilings, alerts, hard caps, and service/agent access selection, and Azure recommends budgets and alerts to prevent cost overruns ([Microsoft Copilot credit management](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-manage-copilot-credits), [Microsoft Foundry cost management](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/manage-costs)).

## 6. Control levers that materially reduce cost

| Lever | Why it works | Executive policy |
|---|---|---|
| Model routing | Smaller models are often 10x+ cheaper for routine work. | Default to low-cost models unless the task proves it needs premium reasoning. |
| Context discipline | Long context multiplies input cost and latency. | Use retrieval, summaries, and context budgets instead of loading everything. |
| Prompt/context caching | Repeated instructions and documents can be billed at discounted cached-input rates. | Require caching for repeated long-context agents. |
| Output limits | Output tokens often cost more than input tokens. | Set maximum answer length by task type. |
| Tool-call budgets | Agents can spend through repeated tool use. | Cap tool calls and require approval for expensive actions. |
| Rate limits and hard caps | They stop runaway loops and retry storms. | Every agent must have a monthly budget and a kill switch. |
| Showback before chargeback | Visibility changes behavior without immediate political friction. | Publish monthly AI cost by department, product, model, and agent. |
| Vendor tiering | Seats are good for humans; API/credits are better for systems. | Do not force all workloads through the same licensing model. |

OpenAI’s API pricing shows cached input discounts as large as 90% versus standard input on GPT-5.5, Anthropic’s pricing shows prompt caching read prices at 10% of input for several models, AWS says cache reads are charged at reduced rates, and Google says implicit caching can provide a 90% discount on cached tokens ([OpenAI API pricing](https://platform.openai.com/docs/pricing), [Anthropic pricing](https://www.anthropic.com/pricing), [AWS Bedrock prompt caching](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html), [Google Vertex AI context caching](https://cloud.google.com/vertex-ai/generative-ai/docs/context-cache/context-cache-overview)).

## 7. A sample 12-month AI budget model

For a 1,000-employee enterprise, a practical first-year model might allocate:

| Budget bucket | Example planning assumption | Why |
|---|---:|---|
| Broad chat seats | 300–600 seats | Start with roles where daily knowledge work is highest. |
| Premium/power users | 5–10% of seats | Coding, analytics, legal, finance, product, and operations power users consume disproportionately. |
| Shared credits/API | 20–40% of seat spend | Covers advanced models, research, coding, agents, and overflow. |
| Department agents | 5–20 funded workflows | Measure cost per successful workflow before scaling. |
| 24/7 systems | Separate capped budget | Prevent autonomous systems from hiding inside departmental software spend. |
| Governance and enablement | 10–20% of AI run-rate | Pays for training, security review, evaluations, monitoring, and FinOps reporting. |

The most CFO-friendly approach is to begin with **showback**, then move the highest-volume departments to **chargeback** after two or three billing cycles of clean data. FinOps Foundation specifically recommends implementing showback to break down AI costs by resource, project, or department and to build cost accountability before using chargeback models ([FinOps Foundation AI overview](https://www.finops.org/wg/finops-for-ai/)).

## 8. Decisions to make this quarter

1. **Name the AI cost owner.** Give finance, CIO/CTO, CISO, procurement, legal, and business-unit leaders a shared monthly review.  
2. **Create the cost taxonomy.** Track seats, credits, API tokens, agents, tool calls, retrieval, cloud hosting, and security/governance separately.  
3. **Set default limits before scale.** Configure user, group, workspace, tenant, and agent budgets before inviting thousands of employees.  
4. **Adopt model-routing rules.** Require justification for premium models and long-context runs.  
5. **Instrument every agent.** No production agent should run without owner, budget, rate limit, retry cap, logs, and kill switch.  
6. **Negotiate overage terms.** Ask vendors for credit price, overage rate, expiration, rollover, true-up timing, admin analytics, Cost API/export access, and data-retention terms.  
7. **Separate human AI from autonomous AI.** Seats are a workforce productivity budget; 24/7 agents are operational systems with cloud-like cost and risk.

## Bottom line

The next two years of enterprise AI cost management will reward organizations that learn to buy **seats for people, credits for variable human demand, and token/agent budgets for systems**. Microsoft’s Copilot Credit direction, OpenAI’s flexible credits and Enterprise spend controls, and Anthropic’s Team/Premium/usage-credit pattern all point to the same future: AI costs will be manageable, but only if executives require unit economics, showback, model routing, hard caps, and agent-level observability before autonomous usage scales ([Microsoft usage-based billing](https://learn.microsoft.com/en-us/microsoft-365/copilot/usage-based-billing-overview-copilot-credits), [OpenAI Enterprise spend controls](https://openai.com/index/chatgpt-enterprise-spend-controls/), [Anthropic Team billing](https://support.anthropic.com/en/articles/9266767-claude-team-plan-set-up-and-billing)).