Remember shadow IT? Developers spinning up EC2 instances on personal credit cards, marketing teams buying SaaS tools nobody in finance knew about, that one intern who provisioned a GPU cluster for a hackathon and forgot to shut it down?

We spent a decade building FinOps practices to tame that chaos. Tagging policies, showback dashboards, commitment discounts, anomaly alerts — the whole discipline exists because cloud spending is invisible until it isn't.

Now we have a new version of the same problem. It's faster, harder to see, and it doesn't even provision infrastructure you can find in a console.

AI agents are the next shadow IT — and most FinOps teams aren't ready.

The Problem: Agents Spend Money Differently

Traditional cloud cost management tracks compute, storage, and network. The unit of spend is a resource: a VM, a container, a database. You can tag it, right-size it, schedule it, reserve it.

AI agents don't work that way. Their cost signature is:

  • Token-based — billed per input/output token, not per hour or per GB
  • Compounding — a single agent task can chain 10-50 model calls, each with full context re-injection
  • Autonomous — the agent decides how many calls to make, which tools to invoke, and whether to retry
  • Invisible — no infrastructure shows up in your cloud bill; it's API calls to OpenAI, Anthropic, or your internal model endpoints

A basic chatbot interaction costs ~1,000 tokens. A moderately complex agent workflow — research, reasoning, tool calls, verification — routinely consumes 20,000 to 100,000+ tokens. That's a 20-100x multiplier that scales with every employee who deploys one.

The Numbers Are Already Ugly

The data from early 2026 deployments is sobering:

  • One 35-engineer SaaS company hit $87,000/month in agent token costs — most of it from autonomous coding assistants running unchecked overnight.
  • A single developer generated $4,200 in a weekend when a refactoring agent entered a recursive loop nobody was monitoring.
  • POC-to-production scaling has seen costs jump from ~$50/month to $2.5M/month when agents go from 5 users to 5,000.
  • Agentic workflows consume 5-30x more tokens per task than conversational AI — and enterprises are consuming 13x more tokens year-over-year.

The pattern is consistent: teams prototype an agent cheaply, get excited about the results, roll it out broadly, and discover the bill 30 days later.

Why Your Current FinOps Practice Misses This

If you're running a mature FinOps organization, you probably have:

  • Cloud cost dashboards (AWS Cost Explorer, Azure Cost Management)
  • Tagging enforcement and allocation rules
  • Reserved instance / savings plan coverage optimization
  • Anomaly detection on infrastructure spend

None of that catches agent costs because:

  1. Token spend often hits a different GL code — it's an API subscription or a software license, not infrastructure. Finance categorizes it as "software" and nobody in FinOps sees it.

  2. There's no resource to tag — you can't tag an API call the way you tag an EC2 instance. The agent runs on someone's laptop or in a serverless function, calls an external API, and the cost shows up as a line item on a vendor invoice 30 days later.

  3. The spender isn't a team — it's an agent — traditional showback assumes a human team made a provisioning decision. With agents, the "decision" to make 47 API calls to complete a task was made by software, autonomously, with no approval workflow.

  4. Agents proliferate like shadow IT — by mid-2026, over 3 million AI agents operate in enterprises, with only ~47% actively monitored. 29% of employees admit to using unsanctioned agents. Departments spin them up with low-code tools, no procurement process required.

What FinOps Teams Should Do Now

This isn't a future problem. It's a right-now problem. Here's what I'd recommend:

1. Get Visibility Into Token Spend — Today

If you can't answer "how much are we spending on AI API calls, by team, by use case, per month?" then you're flying blind. Start here:

  • Audit all API keys and model endpoints across the organization
  • Centralize AI/LLM access through a gateway or proxy layer (this is the equivalent of requiring all cloud access through your organization's AWS accounts)
  • Build a token spend dashboard alongside your cloud cost dashboard

2. Treat Agent Deployments Like Infrastructure Deployments

An agent that autonomously makes decisions and spends money should go through the same governance as provisioning a production service:

  • Require registration of all agents (even internal/experimental ones)
  • Set token budgets per agent, per workflow, per department
  • Implement hard caps with human-in-the-loop approval to exceed them
  • Alert on anomalies: if an agent suddenly consumes 10x its baseline, something is wrong

3. Establish Unit Economics

The FinOps principle of "unit economics" matters even more here. Don't just track total token spend — track cost per business outcome:

  • What does it cost to resolve one customer support ticket with an agent?
  • What's the cost per code review, per research task, per generated report?
  • Is the agent actually cheaper than the human process it replaced, or are we spending more for the convenience of automation?

If you can't answer these questions, you can't make rational build/buy/automate decisions.

4. Build a Model Routing Strategy

Not every agent task needs GPT-4 or Claude Opus. Most organizations are running premium models for tasks that a cheaper model handles just fine:

  • Route simple classification/extraction to smaller models
  • Reserve expensive reasoning models for tasks that actually need them
  • Cache common queries and results to avoid redundant calls
  • Implement tiered model selection based on task complexity

This is the AI equivalent of right-sizing your instances — and it can cut token spend by 40-60%.

5. Add AI Spend to Your FinOps Reporting

Your monthly cost review should now include:

  • Total token spend by department and use case
  • Cost per agent workflow vs. value delivered
  • Trend analysis: are we scaling spend linearly with value, or is it compounding?
  • Shadow AI audit: what unsanctioned tools and agents are we finding?

The Bigger Picture

Gartner projects 40% of enterprise applications will include task-specific AI agents by end of 2026, up from less than 5% in 2025. PwC found 88% of executives plan to increase AI budgets specifically due to agentic AI.

This isn't optional anymore. AI agents are going to be everywhere. The question is whether your FinOps practice evolves to govern them — or whether you're the last to know when your AI bill hits seven figures.

The organizations that figure this out will compound their advantage. They'll deploy agents confidently, scale them efficiently, and measure their ROI clearly. Everyone else will be writing incident reports about why Q3 costs were 300% over forecast.

Trey Morgan is a cloud FinOps leader based in Austin, Texas. He previously led FinOps product strategy at Microsoft and Walmart, and currently drives Cloud FinOps as a Service delivery. Connect with him at treymorgan.com.