The 3-Step AI Cost Audit (Before You Blow The Budget)

In June, the Wall Street Journal reported a KPMG survey where only 26% of companies had full visibility into their AI costs, and KPMG’s global head of AI said some clients burned through annual budgets in months. (Implicator, citing the WSJ)

KPMG’s September AI Pulse of 2,131 leaders found that "only 12 percent consistently assess AI value against cost." (KPMG Global AI Pulse)

The reason is simple. AI billing moved from flat subscriptions to usage, like tokens and credits. Subscriptions behave. Usage doesn’t. This takes about an hour to set up, then a monthly prompt does the work.

Don’t type costs in from memory. You’ll forget the AI add-on buried in your CRM and the trial that started charging in March. Point Claude at where the money already shows up.

Connect Gmail

Most people start here. Claude searches your inbox for every receipt and renewal and builds the sheet from real amounts and dates.

Add your card statement

Drop in a CSV or PDF. Claude pulls the AI line items and catches anything that never sent a receipt. For API accounts, attach a screenshot of the billing page.

The two columns people skip are the ones that save you. Change catches a creeping cost at +$40, not +$400. Type matters because usage lines are the only ones that can explode.

The AI Cost Audit

You are my AI cost auditor. Build and maintain one AI cost sheet. Find the numbers yourself. Only ask me to list things from memory when you truly can't find them. FIRST RUN: 1. Search my connected Gmail for AI receipts and renewals. Use terms like receipt, invoice, "your payment", "subscription renewed", "your plan", "usage credits", plus tool names like Claude, Anthropic, OpenAI, ChatGPT, Midjourney, ElevenLabs, Descript and any others you spot. Pull the tool, amount, date and plan from each email. 2. If I attach a card statement or a screenshot of a billing page, read it and add any AI line item that had no email receipt. Flag AI add-ons hiding inside non-AI tools (CRM, email platform, design tools, website host). 3. Build the sheet with these exact columns: Tool | Type (subscription / usage / both) | Plan & seats | This month | Last month | Change | What it produces | Cap | Alert set? 4. Output it as a table I can paste into Google Sheets, plus the same data as CSV. At the end, list only the items you couldn't find or weren't sure about and ask me to confirm them. EVERY MONTH AFTER: 1. Re-scan Gmail (and any new statement) for this month's charges. Update the sheet and calculate Change for every line. 2. Flag, most urgent first: any line up more than 15% month over month (with the dollar amount), any usage-based line with no cap or alert, and any line whose "What it produces" has been empty 2 months in a row (name it as a cancellation candidate). 3. Close with the verdict: total AI spend, change vs last month, spend with a measurable return vs spend without, and the single action that most improves next month (a cap to set, a tool to cancel or a duplicate to consolidate). Never cancel, change or pay for anything yourself. Only report and recommend.

On a paid plan, have it rerun itself. Describe the task and "the 1st of every month" in a chat, or click Scheduled in the sidebar, then New task. (Scheduled tasks)

A total spend number is useless on its own. You want the cost of the work. What does one blog post, one edited video or one agent run cost you? Without that, you can’t tell the $60 tool that replaces $900 of freelance editing from the $40 one that does nothing.

Price The Work

Using my AI cost sheet, price the work each tool does. Be concrete. When you have to estimate, say so out loud. For every line: 1. Read from the sheet, or ask me, roughly what it produced this month: posts, videos, hours saved, agent runs, whatever fits. One quick number is fine. 2. Calculate COST PER OUTPUT = monthly cost / units produced. For usage-based tools, also estimate COST PER RUN so I know what a single job or agent run costs me. 3. Calculate the MANUAL ALTERNATIVE: what this output would cost from a freelancer, or in my own time at $[MY HOURLY RATE]. Show the gap. 4. Label each line KEEPER (return clearly beats cost), WATCH (unclear) or CUT (cost with no real return). Then give me a ranked table, best return at the top and worst at the bottom, and one line on where my AI money works hardest and where it's leaking. If two tools do the same job, name the one to keep and why.

The 2-Month Rule

If a tool can’t show an output for two months straight, it’s a cancellation candidate. Subscriptions survive on forgetfulness, and the sheet takes that away.

Cap every usage-based account the day you add it. Here’s where.

Claude Console (API)

Go to Settings, then Workspaces, pick your workspace and click Spend limits. Under Monthly limit, click Change limit. Under Email notification, click + Add notification for an alert at, say, 80%. (Workspace limits) Then on the Billing page, click Manage auto-reload, then Turn off, so credits don’t quietly top themselves up. (Auto-reload)

Claude app usage credits

On Pro, Max or Ultra, go to Settings, then Usage. In the Usage credits section, click Adjust limit to set a monthly cap. This matters on Pro, where Fable runs on credits. (Usage credits)

OpenAI API

Open Project settings, then Limits. In Spend, click Edit spend limit, enter a monthly amount and turn on Enforce a hard limit. A spend alert alone only warns you and won’t stop traffic, and even a hard limit can run slightly over. (OpenAI spend limits)

The backstop

For any tool that won’t enforce a real cap, pay with a virtual card that has its own spend limit, like Privacy.com in the US or your bank’s virtual cards. If a tool has neither cap nor alert, your monthly audit from Step 1 is the alert.

One Hour, Then It Runs Itself

Build the sheet once, schedule the monthly prompt and set your caps. After that, the audit is a 10-minute read each month.

Audit Your AI SpendBefore It Audits You

Audit Your AI Spend
Before It Audits You