How to Control AI Costs Before Token Bills and Idle Seats Compound

The short answer

Control AI costs with three moves: attribute every dollar to a workflow, kill idle seats on a 60-day usage rule, and put engineering controls (routing, caching, output caps) on anything running at volume. Only 26% of enterprises have real-time visibility into AI operating costs (KPMG Global AI Pulse, 2026). Visibility first, optimization second — you cannot cut what you cannot see.

AI spend is the fastest-growing line item most mid-market companies have never governed. It arrives through three doors at once: per-seat licenses bought by department heads, token usage from anything engineering shipped, and a shadow layer of tools nobody approved.

Each door has a different failure mode. Seats go idle. Tokens compound silently. Shadow tools never hit the ledger at all.

If you're the COO or VP Ops who just got asked "what are we actually spending on AI?" and had to go find out — this is the control system. It is not a procurement exercise. It's an operating discipline, and it takes about six weeks to install.

Why can't most companies see what their AI actually costs?

Because AI spend doesn't behave like software spend. It's usage-based, distributed across teams, and billed after the fact. Only 26% of enterprises have real-time visibility into AI operating costs; 42% have partial visibility and 22% see nothing until the invoice lands (KPMG, 2026).

Tracking is not forecasting. A Mavvrik and Benchmarkit survey found 98% of enterprises now track AI infrastructure costs, but only 11% can forecast AI spending within 10% accuracy (via ai2roi, 2026). Knowing last month's number tells you nothing about next quarter's.

The tooling gap is real. Flexera's 2026 State of ITAM report found only 31% of organizations have accurate visibility into AI software, and 59% say wasted AI software spend increased over the past year (Flexera, 2026). Your existing SaaS management stack was built for seats, not for tokens.

Nobody owns the number. In most mid-market companies AI spend is split across IT, engineering, and three department budgets. When ownership is diffuse, so is accountability — which is the same root cause behind why so many AI pilots stall before production.

Where does AI money actually leak?

Four places, in roughly this order of size for a mid-market company: idle seats, uncontrolled token usage, orchestration overhead, and shadow tools.

Idle seats are the biggest and most fixable. Enterprise Copilot adoption averages around 35.8%, leaving roughly 64% of licenses idle — about $115K a year wasted on a 500-seat rollout at $30/user/month (Peafowl IT analysis, 2026). IDC found 49% of organizations discovered at least 10% of their Copilot licenses unused in a routine audit (via CloudNuro, 2026). This is pure waste with no upside — and it's usually an adoption problem, not a licensing problem.

Token usage compounds where nobody is watching. Input and output lengths in production rarely match what testing predicted, and context-window creep in long-running conversations can dwarf the cost of the output itself. Deloitte's 2025 AI expenditure work put the average cloud-spend increase attributable to unmonitored token usage at 35% (via DigitalOcean, 2026).

Agents cost an order of magnitude more than chat. EY estimates a simple 2023 chatbot call at roughly $0.04 versus about $1.20 for a 2026 orchestrated agent interaction — nearly 30× (via Forbes, 2026). Multi-step agents re-read context on every hop. That's the math most teams discover in month three, not month one.

Shadow AI never hits the ledger. An estimated 25–35% of enterprise AI tool spending happens entirely outside IT visibility (Second Talent, 2026), and reporting from Okta found nearly all organizations have employees using unsanctioned AI apps (Cybersecurity Dive, 2026). The cost is the smaller problem here; the data exposure is the bigger one.

What does a real AI cost control system look like?

Three layers, installed in order. Attribution, then seat hygiene, then engineering controls. Skipping to layer three is the most common mistake — teams optimize token spend on a workflow that shouldn't exist.

Layer 1 — Attribution (weeks 1–2). Every AI dollar gets tagged to a workflow and an owner. One spreadsheet is fine to start: tool, monthly cost, billing model, workflow it supports, named owner, and whether it has a measured outcome. If a line has no named owner, it's a cancellation candidate by default.

Layer 2 — Seat hygiene (weeks 3–4). Set a standing rule: any seat with no meaningful use in 60 days gets reclaimed, no exceptions, no appeals to seniority. Buy seats in waves tied to enablement, not all at once at contract signing. Gartner notes only about 16% of Copilot pilots convert to production, so committing the full seat count upfront is a bet you'll usually lose.

Layer 3 — Engineering controls (weeks 5–6). For anything running at volume, three controls do most of the work: route simple requests to smaller models, cache repeated context, and cap output tokens. Teams applying all three report inference bill reductions in the 40–70% range without changing application logic (GMI Cloud, 2026). Prompt caching alone can cut input costs 80–98% on cached prefixes (Flexera, 2026).

Set alerts before you need them. Nearly three-quarters of enterprises exceeded AI budget last year, and 98% of FinOps practitioners now manage AI spend (FinOps Foundation, State of FinOps 2026). A spend alert at 70% of monthly budget costs nothing and prevents the quarter-end surprise that gets a program frozen.

Which AI costs are worth defending?

The ones attached to a measured outcome. Everything else is a hypothesis with a subscription.

This is where cost control stops being finance work and becomes strategy. McKinsey's State of AI research puts AI usage at 88% of organizations but EBIT impact at roughly 39% (McKinsey, 2026) — and MIT's Project NANDA found around 95% of enterprise GenAI pilots deliver no measurable P&L impact (via Fortune, 2025).

So the question isn't "is this cheap?" It's "does this line move a number we already track?" A $4,000/month workflow that removes 60 hours of manual reconciliation is cheap. A $400/month tool nobody can tie to an outcome is expensive. We wrote about which workflows actually pay back — the cost decision and the ROI decision are the same decision.

Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Gartner, 2025). Notice that cost sits alongside unclear value in that list. They're not separate failures — projects get expensive precisely because nobody defined what "working" meant.

The mistake we see most often

Companies treat AI cost control as a cleanup job that happens after the spend is already out of hand. By then the political cost is high: someone has to walk into a leadership meeting and explain a number they can't defend, and the usual outcome is a blanket freeze that kills the two workflows that were actually working alongside the eight that weren't.

When we run an automation build at Mesh Flow, cost attribution goes in at design time, not after launch — every workflow ships with a per-run cost estimate and an owner attached. It's unglamorous and it takes an extra afternoon. It's also the difference between a program that survives its first budget review and one that doesn't.

The other frequent mistake: optimizing the model bill before fixing the seats. Token optimization is the interesting engineering problem, so teams start there. But for most mid-market companies, idle licenses are the larger and far easier number — and you don't need an engineer to fix them.

Frequently asked questions

How much should a mid-market company budget for AI?

There's no useful benchmark percentage, because the range across companies is enormous. Budget per workflow instead: estimate cost-per-run, multiply by expected volume, add 40% for the overruns that always happen in month one. We broke down realistic figures in what AI automation actually costs.

Why did our AI bill jump without our usage changing?

Usually context growth, not user growth. Longer conversation histories, larger retrieved documents, or an agent added to an existing workflow all increase tokens per request while request counts stay flat. Agent orchestration is the most common culprit — EY puts an orchestrated agent interaction at roughly 30× the cost of a simple chatbot call.

Who should own AI spend in a mid-market company?

One named person with budget authority, sitting close to operations rather than in IT procurement. Diffuse ownership is why 42% of companies have only partial visibility into what they're spending. We covered the ownership question in detail in who should own AI automation.

Is it cheaper to build agents in-house or buy tools?

It depends on volume and how specific the workflow is, and the honest answer is that in-house builds carry ongoing maintenance costs most teams underestimate. We ran the numbers in the real cost of building AI agents in-house.

Do we need a FinOps tool for AI, or is a spreadsheet enough?

A spreadsheet is enough below roughly $10K/month in AI spend across fewer than ten tools. Above that, you want tagged usage data and automated alerts. Buy the tool when the manual tracking starts failing — not before, and not because a vendor told you AI FinOps is a category.

The bottom line

Visibility first. Only 26% of enterprises have real-time AI cost visibility. Attribution beats optimization.
Kill idle seats before optimizing tokens. Roughly 64% of Copilot licenses sit unused — that's the biggest, easiest number.
Every line needs an owner and an outcome. Spend with neither is a cancellation candidate.
Put controls in at design time. Routing, caching, and output caps cut inference bills 40–70%, but only if someone builds them in.

If you want a second set of eyes on where your AI spend is actually going — and which workflows are worth defending — that's the kind of work we do at Mesh Flow.