Why a printable AI spend tracker is the cheapest cost control you can ship in 2026
The average mid-market company in April 2026 is paying for AI in five to twelve places — OpenAI direct, Anthropic direct, Google Vertex, Cursor seats, GitHub Copilot seats, Cursor enterprise plan, Otter or Fathom, an embedding-API line item, a vector database, plus whatever Marketing slipped onto a corporate card. Every one of those vendors has a dashboard. None of those dashboards talk to each other. The result is that nobody on the finance side knows what the total monthly AI spend actually is — until somebody asks, at which point three days of accounting work surface a number that is usually 30-60% larger than anyone expected.
The cheapest fix for this is the oldest one: a single printable monthly sheet with one row per provider and one column per workload. Spend two minutes filling it from the previous-month invoices. Circle anything that grew >25% month-over-month. Cancel anything that has been <$50 with zero clear owner for three consecutive months. The review is the value; the sheet is the forcing function.
What goes on the tracker
The minimum-viable monthly tracker has eight columns. The calculator above defaults to that shape:
| Column | What it captures |
|---|---|
| Workload | The thing the AI is doing — e.g. 'support chatbot', 'engineering copilot' |
| Provider | OpenAI, Anthropic, Google, Cursor, GitHub, etc. |
| Model / plan | Sonnet 4.5, GPT-5, Cursor Pro, Copilot Business |
| Volume this month | Calls, tokens, or seats — whatever the invoice unit is |
| $ this month | Total invoice amount |
| $ last month | For trend |
| Owner | Single person responsible — never 'team' or 'tbd' |
| Status | Active / under review / sunset |
Everything beyond those eight columns is decoration that gets in the way of the review. Resist the urge to add per-user breakdowns to a printable sheet — that data belongs in whatever observability tool you are running. The printable exists to force monthly attention, not to be the source of truth.
The monthly review ritual that makes the tracker work
- First Friday of every month: download the previous month's invoices for every provider on the tracker. Sum into the sheet. Whoever owns the workload owns the spend column for that row.
- Circle every row where the month-over-month change is >25%. Most of these are fine (volume grew because the product grew); a quarter of them are not (silent failure retries, a deploy that bloated prompts, an integration that lost its cache hits). The circles are the only way you find the not-fine ones in time.
- Tag every row that has been <$50 for three consecutive months and ask the owner whether to cancel or consolidate. The answer is usually cancel, and cancelling those rows typically frees 8-15% of total AI spend across the year.
- At quarter end, sort the sheet by spend descending. The top three rows should be 60-80% of total spend. If they are not, you have over-fragmented and should look for consolidation opportunities.
The five spend leaks the printable surfaces fastest
1. The orphan subscription
Almost every company we audited in 2025-2026 has 2-4 AI subscriptions on a corporate card that nobody on the leadership team owns. ChatGPT Plus or Claude Pro, sometimes Midjourney or Runway, occasionally a marketing-AI suite somebody trialed and never cancelled. Total leak is usually $400-2,000/month. The printable surfaces these in month two — they are always the rows with a blank "owner" cell.
2. The dev-environment-API-key bill
An engineer wires an API key into a dev environment. The key gets committed to a notebook somewhere. Six months later a test loop costs $4,000/month and nobody noticed because the API key is in an account labeled "ML experiments." The printable surfaces this as a provider row with no owner and growing spend.
3. The retried-into-oblivion workload
A workload with an 18% retry rate looks fine on the dashboard until somebody asks why the spend grew 22% in a month with flat usage. The printable surfaces it as a month-over-month circle that does not match any usage change. Catching this once usually pays for the review process for the rest of the year.
4. The over-provisioned model
A classifier workload running on Sonnet 4.5 at $0.012/call when Haiku 4 would do it at $0.001/call. The printable does not catch this directly — what it catches is the spend line that is 8-12× larger than peer companies' equivalent workload. The model-choice review then surfaces the over-provisioning.
5. The duplicate-vendor situation
Marketing pays Jasper for content; Sales pays Copy.ai; Operations pays Writer. All three teams are GPT-4o wrappers with different UI. The printable, sorted by workload not vendor, surfaces three rows doing the same thing — and is the only artifact senior leadership reads quickly enough to make the consolidation decision.
How much does typical 2026 AI spend look like by company size
Aggregating across the 60+ AI cost audits behind this calculator (small disclosed sample, April 2025 - April 2026):
| Company size | Total AI spend / month | % of OpEx | Top spend line |
|---|---|---|---|
| Solo / 1-5 person | $50-300 | 1-3% | ChatGPT Plus + Cursor seats |
| 10-50 employees | $800-4,500 | 2-5% | Cursor seats + ChatGPT Enterprise |
| 50-200 employees | $4,500-22,000 | 1.5-4% | Engineering copilots + chatbot |
| 200-1,000 employees | $22,000-180,000 | 1-3% | Multi-vendor + custom RAG |
| 1,000+ employees | $180,000-2.5M | 0.8-2.5% | Custom LLM apps + enterprise SKUs |
The variance inside each bucket is wider than the spread between buckets — a 100-person AI-native startup can spend more than a 1,000-person legacy services firm. Use these numbers as sanity checks, not as targets.
- Live AI Spend Tracker (interactive) — Use this if you want a working spreadsheet instead of a printable
- AI Tool Stack Cost — Total monthly cost across every subscription
- LLM vs LLM TCO 2026 — Right-size model choice for each workload
- LLM API Cost Calculator — Forecast individual workload spend
The Google-Sheets layout if you would rather not print
Same eight columns, with three formulas:
=SUM(E:E)— total spend this month (column E).=(E2-F2)/F2— month-over-month percent change per row.=COUNTIF(H:H, "Active")— number of active workloads.
Conditional formatting: red on any row where MoM change >25%. That is it. Anything richer becomes a sheet nobody opens.
FAQ
How often should this be reviewed?
Monthly. Anything less frequent and the spend leaks get expensive. Anything more frequent and the review fatigues and gets skipped.
Should developer API keys be on the tracker?
Yes. Aggregate to one row per environment (dev, staging, prod). Without it, the committed-key-in-a-notebook leak class never surfaces.
How granular should "workload" be?
One row per business-readable workload. "Support chatbot," "engineering copilot," "AI meeting notes." Not "OpenAI calls" or "all SaaS spend." The point of the tracker is to answer "what is this spend doing for the business" at one glance.
Who should own the review?
Whoever owns OpEx. Usually CFO or Head of Operations. The CTO can run the model-choice review on top, but the spend governance belongs to finance.
How do I get the printable PDF?
Click "Export PDF" on the calculator result card above, or sign up for the 2026 AI Pricing Cheat Sheet via the box below and we will email both the printable tracker and the Google Sheets template together.
Updated April 2026. The tracker layout itself is provider-agnostic and stable — the reference numbers per company size will continue to drift as 2026 pricing falls. Re-pull every quarter.