Finding shadow AI spend: where AI cost hides outside the IT budget
In brief
Unbudgeted AI spend sits on cards, in price uplifts, in separate API accounts and in metered add-ons. How to find each, and whether to consolidate or block.
Steve Elliott · · 7 min read
On this page
In our view, most unbudgeted AI spend sits in five places: expense reports and corporate cards, price uplifts on software you already own, API accounts opened outside central billing, metered AI inside existing contracts, and AI services in cloud bills coded to someone else's category. Each leaves a different record, so no single report finds all of it.
Once found, we think most of it should be consolidated or approved rather than blocked, because spend that nobody budgeted is also evidence of demand. The security case for controlling shadow AI is real, but it is a separate exercise.
Why does AI spend hide outside the IT budget?
Because people buy AI faster than budgets are set. In Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 people in 31 countries, 78% of those using AI at work said they brought their own AI tools. A Censuswide survey of 2,003 UK employees, commissioned by Microsoft in October 2025, found 71% had used unapproved consumer AI tools at work, and 51% said they still did so every week. Explaining why, 41% said it was what they were used to in their personal life, and 28% said their company offered no work-approved option. A Gartner survey of 302 cybersecurity leaders, run from March to May 2025, found 69% of organizations suspect or have evidence that employees use prohibited public generative AI.
Ramp's AI Index, built from corporate card, invoice and ACH transactions at more than 85,000 US businesses, put paid AI adoption at 56.8% of those businesses in October 2026. Ramp says its figure likely understates adoption, partly because employees use personal accounts with AI companies for work.
The FinOps Foundation's State of FinOps 2026, with 1,192 respondents, found 98% now manage AI spend, up from 63% in 2025. Many report difficulty getting clear visibility into AI usage and costs.
Where does shadow AI spend show up?
| Where | What it looks like | Record to search | Who usually holds the record |
|---|---|---|---|
| Expense reports and corporate cards | Small recurring charges to AI vendors | Card feeds, expense lines, merchant names | Finance, card program owner |
| Price uplifts for bundled AI | A higher per-seat price at renewal | Renewal quotes, vendor price notices | Procurement, software asset management |
| API accounts outside central billing | A separate developer account on a team card | Payments to model providers, API key inventories | Engineering teams, accounts payable |
| Metered AI in existing contracts | Credits or usage meters drawn down by agents | Admin billing pages, credit balances | Application owners |
| AI services in cloud bills | Model and inference charges inside an engineering account | Cloud billing data by service category | FinOps |
How do you find each one?
We suggest running all five searches once, then monthly.
- Cards and expenses. Pull 12 months of card and expense lines and search merchant names for the model providers and the AI apps your teams mention. This misses free tools, which leave no payment. For those, network discovery helps: Microsoft says its Defender for Cloud Apps catalog includes more than a thousand generative AI apps, each with a risk assessment of 90+ factors, and lets admins sanction or block each one.
- Price uplifts. Read every renewal quote for a line about AI. Microsoft raised list prices for most commercial Microsoft 365 plans on July 1, 2026; those plans gained Copilot Chat enhancements and Copilot Chat Analytics alongside security and management features, and existing customers stay on current pricing until renewal. Copilot Chat enhancements also went to plans whose price didn't change, such as Office 365 E1. Slack raised its annual Business+ price from $12.50 to $15 per user per month when it added AI features, with existing web customers moving at their first renewal after August 17, 2025. Google said in January 2025 that Workspace Business Standard with Gemini would cost $14 per user per month, "only $2 more than they were paying for Workspace without Gemini."
- API accounts. A team can open its own developer account and pay by card. OpenAI's help center says each organization, including a sub-organization, is treated separately, and its usage dashboard doesn't combine data across organizations. Anthropic's Admin API, which reports usage and cost, isn't available to individual Claude Console accounts or to organizations on the claude.ai Team plan. A central admin view can't see an account it doesn't know about, so start from accounts payable and card payments to model providers, then match each one to a known organization ID.
- Metered AI in contracts. Some AI inside existing software is billed by use, not by seat. Microsoft 365 Copilot pay-as-you-go, which covers Copilot Chat, SharePoint agents and the Copilot Retrieval API (preview), is billed through the Azure subscription named at setup. It is off by default, and admins can set a budget limit on each billing policy. As of November 2026, Salesforce prices Agentforce Flex Credits at $500 per 100,000 credits; its May 2025 announcement put a standard action at 20 credits, or $0.10. Check admin billing pages and credit balances for each.
- Cloud bills. AI services bought through a cloud provider land in engineering accounts. If your cloud data follows the FOCUS specification, filter on the ServiceCategory value AI and Machine Learning; its ServiceSubcategory values include Generative AI. That gives one view across providers, though provider mappings vary.
Metered Copilot charges raise a classification question. In our view, when billed to Azure they can easily be reported as cloud spend when they belong with the productivity software budget.
What can renewals and meters add in a year?
The numbers below are illustrative, for a made-up company with 10,000 Microsoft 365 E3 seats, 2,000 Slack Business+ seats on annual billing, and a customer service team running 50,000 Agentforce actions a month. Prices are published list prices as of November 2026; your negotiated rates may differ.
| Item | Before | After | Added annual cost |
|---|---|---|---|
| Microsoft 365 E3 (with Teams; the no-Teams version rises by the same $3), 10,000 seats | $36 per user per month | $39 per user per month | +$360,000 |
| Slack Business+ annual, 2,000 seats | $12.50 per user per month | $15 per user per month | +$60,000 |
| Agentforce actions, 50,000 a month | None | $0.10 per action | +$60,000 |
| Total | +$480,000 |
The arithmetic: $3 × 10,000 × 12 = $360,000; $2.50 × 2,000 × 12 = $60,000; 50,000 × $0.10 × 12 = $60,000.
Two cautions. First, Microsoft ties its increase to AI, security and management capabilities together, so not all of the $360,000 is an AI cost. We think it still belongs on the AI budget's watch list. Second, none of these lines is wrong spending. Each arrives through a renewal or a meter, not through a business case, so it can grow without anyone deciding it should.
What should you do once you find it: consolidate, approve or block?
We suggest one question per item: is the use valuable, and is it running on terms you would accept?
| Decision | When | What it looks like |
|---|---|---|
| Consolidate | The use is valuable, but it runs on many personal or team accounts | Move it to an approved enterprise plan with central billing |
| Approve | One team has a valid use on an account you can see | Register it, name an owner, put it in the budget |
| Block | The data risk is unacceptable or an approved alternative exists | Unsanction the app and point people to the approved one |
Consolidate. OpenAI, for example, says that when a ChatGPT Enterprise or Edu invitation uses a verified company domain, an employee's existing personal account may need to be migrated, and an eligible web-billed Plus or Pro subscription may be canceled and refunded in that flow. For smaller teams, ChatGPT Business Standard seats list at $20 per user per month on annual billing, as of November 2026, though Business admins can't require employees to merge personal accounts.
Approve. A team API account with a real workload should be brought in, not shut down. Give it an owner, a budget line and a place in your cost allocation. In our view an AI gateway is a sensible later step, routing those calls through one billed account.
Block. Keep it for data risk or clear duplication. Gartner's advice is to set enterprise-wide policies for AI tool use, run regular audits for shadow AI activity, and add generative AI risk to SaaS assessments. We think blocking works best where an approved option exists; in Microsoft's UK survey, 28% said their company doesn't provide a work-approved option.
How do you keep useful experimentation alive?
In our view the aim is to make approved paths faster than unapproved ones, not to stop people trying things.
- A small, open sandbox. A team experiment budget with a card limit and an allow-list of providers, reviewed monthly.
- A fast route to approval. If approval takes a quarter, people will keep expensing subscriptions.
- An inventory that includes spend. NIST's AI Risk Management Framework says organizations should have mechanisms "to inventory AI systems." We think each entry should carry its cost, owner and budget line, not just its risk rating.
- Clear ownership. Card spend, SaaS renewals and API accounts often sit with different teams. Our post on who owns SaaS spend sets out how FinOps, software asset management and procurement split that work, and SaaS FinOps practices apply to AI subscriptions too.
Finding shadow AI spend is the first step; deciding who owns each line, against which budget and outcome, is what keeps it found. That is where we think an operating model helps: it connects each AI cost to the owner and the decision behind it. Card and API charges are also only part of an AI initiative's cost; labor is usually larger, as our post on AI initiative costs shows.
Sources
- Microsoft, Microsoft and LinkedIn release the 2024 Work Trend Index on the state of AI at work, May 8, 2024
- Microsoft UK, Rise in shadow AI tools raising security concerns for UK, October 13, 2025
- Gartner, Gartner Identifies Critical GenAI Blind Spots That CIOs Must Urgently Address, November 19, 2025
- Ramp, Ramp AI Index, accessed November 22, 2026
- FinOps Foundation, State of FinOps 2026, released February 19, 2026
- Microsoft Learn, Discover AI apps and data, accessed November 22, 2026
- Microsoft 365 blog, Advancing Microsoft 365: new capabilities and pricing update, December 4, 2025
- Microsoft Licensing, Microsoft 365 pricing and packaging updates, February 16, 2026
- Slack, June 2025 pricing and packaging announcement, June 17, 2025
- Slack Help Center, Updates to feature availability and pricing for Slack plans, accessed November 22, 2026
- Google Workspace blog, Empowering businesses with AI, January 15, 2025
- OpenAI Help Center, Reviewing API usage and costs, accessed November 22, 2026
- Anthropic, Usage and Cost Analytics API (beta), accessed November 22, 2026
- Microsoft Learn, Pay-as-you-go for Microsoft 365 Copilot overview, updated August 18, 2026
- Salesforce, Agentforce pricing, accessed November 22, 2026
- Salesforce, Agentforce flexible pricing announcement, May 15, 2025
- FinOps Foundation, FOCUS 1.4: ServiceCategory and ServiceSubcategory
- OpenAI Help Center, Onboarding employees with existing ChatGPT accounts, accessed November 22, 2026
- OpenAI Help Center, ChatGPT Business overview, accessed November 22, 2026
- NIST, AI Risk Management Framework 1.0: Core, 2023
Frequently asked questions
Where does most shadow AI spend show up?
In our view, in five places: expense reports and corporate cards, price uplifts on software you already own, API accounts opened outside central billing, metered AI inside existing contracts, and AI services in cloud bills. Each leaves a different record, so finance, procurement, application owners and FinOps each hold part of the picture. We suggest running all five searches once, then monthly.
Does bundled AI count as shadow AI spend?
It counts as unbudgeted AI spend if nobody planned for it. Microsoft raised most commercial Microsoft 365 plan prices on July 1, 2026, citing AI, security and management features, and Slack raised Business+ when it added AI. In our view these uplifts belong on the AI budget's watch list, even though not every dollar of them pays for AI.
How do you find AI API accounts that IT doesn't know about?
Start from payments, not admin consoles. OpenAI treats each organization separately and its usage dashboard doesn't combine organizations, while Anthropic's Admin API isn't available to individual Console accounts or claude.ai Team plan organizations. Search accounts payable and card data for payments to model providers, then match each to a known organization. Any payment without a match is an account to bring in.
Should companies block unapproved AI tools?
Only some, in our view. Block where the data risk is unacceptable or an approved alternative exists. Otherwise consolidate valuable use onto an enterprise plan or approve it with an owner and a budget. In Microsoft's 2025 UK survey, 28% said their company doesn't provide a work-approved option. In our view, blocking without one tends to push use elsewhere.
Steve Elliott
Published November 22, 2026
More on AI Economics
Tokens, inference cost, cost per task.
Keep reading
Cost center or product P&L: where technology cost should land
Should technology run as a cost center, a shared-services recharge or inside product P&Ls? What each model does to behavior, plus a worked margin example.
Cost per token is the wrong unit: how to measure AI cost per task
A cheaper per-token model can cost more per finished task. How to define a task, instrument it and compute cost per successful task, with worked math.
Self-hosting an open-weight model vs paying per token: where the break-even sits
At Google Cloud prices as of November 2026, self-hosting gpt-oss-120b struggles to beat the same model sold per token. The break-even math, and when it wins.
Forecasting cloud spend when usage won't sit still
Split committed, usage-driven and AI cloud spend, tie usage to business drivers, forecast in ranges and measure variance by line, with a worked example.