How to Audit Your AI Spend: A Step-by-Step Guide

AI spend accumulates invisibly. A developer adds an API key for a side project. A marketing manager subscribes to Jasper at $49. A product lead subscribes to ChatGPT Plus at $20. Engineering routes all inference to GPT-5.4 by default at $2.50 / $15 per 1M.

Within six months, a ten-person company can be spending $500–$1,000/month on AI with no single person who sees the whole picture. The audit is the act of making that picture visible.

This guide gives you a four-step process to find every line of AI spend, categorise it and eliminate waste without disrupting workflows that deliver value.

Start by listing every AI product your organization is paying for. Common places to find them:

• Credit card statements: search for OpenAI, Anthropic, Google, Mistral, Cursor, Perplexity, Jasper, Writesonic, Copy.ai, Grammarly, GitHub Copilot
• Expense reports and reimbursements: team members often expense AI tools individually
• Engineering environment variables: audit .env files and secret managers for AI API keys
• SaaS billing dashboards: check AWS Marketplace, Azure, and GCP for AI service billing

Create a spreadsheet with: tool name, monthly cost, who uses it, what for, and when it was last used. This inventory is the foundation of the audit.

Once you have your inventory, categorise each item:

• Bucket 1 — Team subscriptions: flat-fee tools used by individuals (ChatGPT Plus, Claude Pro, Cursor Pro, Perplexity Pro at $20; ChatGPT Business at $25/seat; Claude Max from $100). Easy to count, easy to right-size.
• Bucket 2 — Internal automation: API usage for internal tools, workflows, agents and integrations. Usually the fastest-growing and least-monitored category.
• Bucket 3 — Customer-facing inference: API usage that powers product features. The most business-critical and the most expensive at scale.

Separating these buckets matters because each has different levers. Team subscriptions are easy wins. Internal automation usually has routing and caching opportunities. Customer-facing inference needs the most careful cost engineering.

Look for these common patterns:

• Unused seats: people who subscribed to ChatGPT Plus or Jasper and never used it after the first week. Check login history or ask directly.
• Duplicate capabilities: teams paying for Jasper and Writesonic and Claude Pro — all doing roughly the same writing tasks. Pick one.
• Overprovisioned models: APIs using GPT-5.4 ($2.50 / $15) or Claude Sonnet 5 ($2 / $10) for classification, summarisation and extraction that GPT-5.4 mini ($0.75 / $4.50), Gemini 3.8 Flash ($0.75 / $3.75) or DeepSeek V4 Flash ($0.05 / $0.16) handle identically at 70–98% lower cost.
• Uncached prompts: a long system prompt re-sent on every request when cached input reads cost roughly 10% of list on OpenAI and Anthropic.
• Uncapped API usage and agent loops: developer experiments or internal agents with no spending limits. These are the most dangerous — a single runaway tool-call loop can generate a bill in hours.

Mark each waste item with an estimated monthly saving. Prioritise the highest numbers.

The audit is only valuable if it creates lasting change. After cutting obvious waste:

• Set hard spend alerts in all API dashboards at 50% and 90% of your monthly budget
• Require explicit approval for new AI subscriptions above $20/month
• Add AI cost tracking to your monthly team metrics review
• Schedule a quarterly audit — pricing changes, new tools launch, and usage patterns drift

The goal is a lightweight governance process, not bureaucracy. Most teams need only a shared spreadsheet and a monthly five-minute check-in to maintain control after the initial audit.