The Moment You Realize You're Paying for AI You Can't Account For
You open your credit card statement and count six separate line items with "AI" in the name. You know roughly what each tool does. You have no idea whether any of them are actually saving you money. That's the moment most business owners hit, and it's more common than you'd think.
The good news: you can fix this in about two hours without a developer, without a finance team, and without canceling everything and starting over. What you need is a simple audit framework that connects every dollar you're spending on AI to a specific workflow, a specific outcome, and a real number.
Here's how to run it.
Why AI Costs Feel Out of Control (It's Not Your Fault)
Most businesses end up with three categories of AI spend, and they usually don't realize it until they look.
First, there are flat SaaS seats. These are the monthly subscriptions where you're paying per user for tools like Notion AI, Jasper, or the AI add-on inside your CRM. Some of those seats are getting used daily. Some haven't been touched in three months.
Second, there are per-token API costs. If you've built any automations using OpenAI, Anthropic, or similar APIs, you're paying per call. Those costs are variable, they live inside your cloud or API billing, and they're almost never tracked against anything meaningful.
Third, there's redundancy. You're probably paying for two or three tools that do the same thing. One team is using ChatGPT for content drafts, another is using a separate writing tool, and someone in ops set up a Claude integration through Make last quarter. Nobody owns the full picture.
A one-hour audit of a typical 10-50 person business almost always surfaces 20-40% of AI spend that's delivering zero measurable output. Not "low ROI." Zero.
The Four-Step Audit (Run This This Week)
This doesn't require any special software. A spreadsheet works fine.
Step 1: List everything. Pull every AI-related charge from the last 60 days. Include SaaS subscriptions, API invoices from OpenAI or Anthropic, AI add-ons inside tools like HubSpot or Salesforce, and automation platform costs from Make, Zapier, or n8n. Don't skip the small ones. A $29/month tool nobody uses is still $348 a year.
Step 2: Tag each one to a workflow. Not a department. A specific workflow. "Lead intake," "proposal drafting," "customer support responses," "invoice processing." If you can't name the workflow a tool supports, that's your first red flag.
Step 3: Assign a weekly hours-saved estimate. Be honest here. If a tool is running an automation that takes a task from 4 hours a week to 30 minutes, that's 3.5 hours saved. If it's theoretically useful but nobody's actually using it, that's zero.
Step 4: Calculate cost per hour saved. Take the monthly cost, divide by 4.3 (average weeks per month), divide by your weekly hours saved. That's your cost per hour saved for that workflow. Now compare it to the loaded hourly rate for the role that workflow touches. If you're paying $40/hour saved for a task your $25/hour admin used to handle, that automation isn't pulling its weight.
Anything above your loaded hourly rate is a candidate for cutting, consolidating, or replacing with a cheaper approach.
Model-Routing: The Cost Lever Almost Nobody Uses
Here's something most small businesses don't know: not every AI task needs the same model.
GPT-4o and Claude Sonnet are powerful, but they're also the most expensive options per call. They're worth it for complex tasks like drafting a detailed legal summary, writing a nuanced client proposal, or generating a structured analysis from messy data. But document classification, FAQ responses, tagging support tickets, and simple data extraction? Those tasks run just as well on smaller, cheaper models. We're talking 10-20x lower cost per call.
The practical fix is called model routing, and a tool called OpenRouter already does this automatically. You connect your workflows to OpenRouter, set rules by task type, and it sends simple requests to lighter models and complex ones to frontier models. You don't have to rebuild anything. You just add a routing layer.
Cloudchipr's research shows that routing routine traffic to cheaper models can reduce inference costs by 30-50% without any meaningful drop in output quality at the application level. For a business running moderate API volume, that's real money.
If you're using n8n or Make to run automations that call AI models, this is worth looking at this week. The setup is not as hard as it sounds.
What a Healthy AI Budget Actually Looks Like
There's no universal number, but there are some useful benchmarks for businesses in the 10-200 person range.
For most small businesses, AI tooling should represent a measurable productivity return. Azure's guidance on this is straightforward: the goal isn't to spend less, it's to spend more efficiently in pursuit of measurable outcomes. Productivity gains, customer satisfaction, operational efficiency. If you can't point to one of those for a given tool, the tool is a cost, not an investment.
A reasonable target: every dollar of AI spend should save at least two to three dollars in labor time or generate equivalent revenue impact. That's not a hard rule, but it gives you a starting point for the "is this worth it" conversation.
The other thing to watch is allocation. CloudZero's framework breaks AI cost management into three layers: visibility, allocation, and optimization. Most small businesses skip straight to optimization (canceling stuff) without doing the visibility and allocation work first. That's why they cancel the wrong things.
Tag your AI costs by workflow before you start cutting. Otherwise you're guessing.
Building a Monthly Review That Takes 30 Minutes
The audit is a one-time reset. The monthly review is what keeps your AI spend from drifting back into chaos.
Set a recurring 30-minute calendar block. Pull your AI spend for the month. Check three things: total spend vs. last month, any new tools that appeared (it happens), and whether your top three workflows are still delivering the hours-saved numbers you estimated.
If a workflow's output dropped, find out why before you cancel the tool. Sometimes the automation broke. Sometimes the process it was supporting changed. Sometimes the tool is fine and the workflow needs to be rebuilt.
If a new tool appeared that nobody approved, that's a conversation about purchasing process, not a technology problem.
Notion AI or a simple shared doc works fine for tracking this. You don't need a dashboard. You need a habit.
Three Things You Can Set Up This Week
One: Run the four-step audit above. List every AI cost, tag each to a workflow, estimate hours saved, calculate cost per hour saved. This alone will tell you what to cut.
Two: If you're making API calls to OpenAI or Anthropic through any automation, sign up for OpenRouter and look at which of your workflows are using frontier models for simple tasks. Route the simple ones to lighter models. Check the cost difference after 30 days.
Three: Build your monthly review into your calendar now. Thirty minutes, recurring, monthly. Pull the spend, check the top workflows, flag anything new. That's the whole system.
You don't need to overhaul your AI stack. You need to see it clearly, connect it to outcomes, and check in on it regularly. Most businesses that do this find they're not spending too much on AI overall. They're just spending it in the wrong places.