SYSTEMS
Audit Your AI Spend Before It Audits You 📚
Companies are burning entire yearly AI budgets in months, and most can't even see it happening. This is the exact 3-step audit Somya runs across her.
Source: KPMG survey: 26% AI cost visibility →
The numbers behind the video: a KPMG survey reported by the Wall Street Journal found only 26% of companies have a comprehensive view of their AI costs. Half have partial visibility. The rest find out when the bill lands. KPMG's global AI head described clients who exhausted their entire annual AI budget in months, including one whose token usage jumped sixfold.
The reason is structural: AI billing quietly shifted from predictable subscriptions to usage-based pricing (tokens, credits, compute). Subscriptions behave. Usage doesn't. The audit below takes about an hour to set up, and after that an agent does the monthly work for you.
Let Claude Build The Sheet From Your Real Receipts
Do not type your costs in from memory. You will forget the AI feature buried inside your CRM and the trial that started charging in March. Point Claude at where the money already shows up, easiest path first:
Connect Gmail (most people start here): Claude searches your inbox for every receipt and renewal notice and builds the sheet from real emails, real amounts, real dates. Drop in your card statement (CSV or PDF, into the chat or Google Drive): Claude pulls the AI line items and catches anything with no email receipt. Power move with Claude Code: have it read the actual usage numbers straight off your Anthropic and OpenAI billing pages, so usage tools are exact instead of estimated.
The sheet has these columns: Tool | Type (subscription / usage / both) | Plan + seats | This month | Last month | Change | What it produces | Cap | Alert set? The two people skip are the two that save you: Change (so a creeping cost gets caught at +$40, not +$400) and Type (usage lines are the only ones that can explode).
Connect Gmail first, then paste this prompt (and save it as a monthly scheduled task so it reruns itself). Claude finds the costs for you instead of asking you to list them.
Price Every Task So You Know What It Actually Costs
A total spend number is useless on its own. What you actually want to know is the cost of the work: what does one blog post, one edited video, one agent run cost you? Most people have no idea, so they cannot tell the $60 tool that replaces $900 of freelance editing from the $40 one that does nothing.
The prompt below takes the sheet from Step 1 and prices it: cost per output (monthly cost divided by what it made) and cost versus doing it by hand (a freelancer's rate, or your own hourly rate). It ranks every tool from best return to worst and names the dead weight.
Paste this right after the sheet exists. It turns a list of charges into a ranked ROI report, so you know exactly what each tool buys you.
Set The Caps And Alerts (Exact Clicks)
Subscriptions behave. Usage is what blows up budgets, so cap every usage-based account the day you add it. Here is exactly where:
Anthropic Console: Settings then Workspaces, open your workspace, go to the Limits tab and choose Change Limit to set a monthly spend cap, then Add notification for an email alert at, say, 80%. Under Extra Usage, turn Auto-Reload off (or cap it) so credits don't silently top themselves up.
OpenAI: in your org settings open the project, click Limits, set a Monthly budget and Add Alert. Know this: OpenAI's budget is now notification-only, it does not stop API calls, so treat it as a warning, not a wall.
The backstop nobody tells you about: auto-reload is the real budget killer, so switch it off wherever you can. For any tool that won't enforce a true cap, put it on a virtual card with a fixed monthly limit (Privacy.com, Revolut, or your bank's virtual cards). That is the one cap that physically cannot be exceeded. And if a tool has neither cap nor alert, your monthly scheduled audit from Step 1 is the alert.
One hour, then it runs itself
Build the sheet once, schedule the monthly prompt, set your caps. After that the audit is a 10-minute read each month, and your AI spend becomes a number you manage instead of a bill you discover.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.