In August 2025, my cofounder and I sat down to figure out why our monthly software bill had jumped by $3,400 in six months. We sell project management software to mid-size construction companies. Twelve employees. Nothing extravagant. We should not have been spending more on AI tools than on our actual office lease.
But we were.
So I did something that, in hindsight, I should have done from day one: I built a spreadsheet and tracked every single AI-related expense for six months. Every subscription. Every API call. Every "free trial" that auto-converted. Every time someone on the team said "let me just throw this into ChatGPT real quick" and it hit our company OpenAI account.
The results were... educational. And by educational, I mean I wanted to scream into a pillow.
The Spreadsheet of Shame
Here is what we were paying, categorized by function. These are monthly averages over the six-month tracking period (September 2025 through February 2026):
Content and Marketing
- Jasper AI (Business plan): $299/month
- Midjourney (Standard): $30/month
- Grammarly Business (4 seats): $60/month
- Copy.ai (Growth): $49/month
Subtotal: $438/month
Development and Engineering
- GitHub Copilot Business (3 seats): $57/month
- OpenAI API (GPT-4o): ~$340/month (this one fluctuates wildly)
- Anthropic API (Claude): ~$180/month
- Cursor Pro (2 seats): $40/month
Subtotal: $617/month
Operations and Support
- Intercom with AI features: $189/month
- Notion AI (8 seats): $80/month
- Otter.ai Business (3 seats): $60/month
- Fireflies.ai (Pro, 2 seats): $36/month
Subtotal: $365/month
Analytics and Research
- Perplexity Pro (2 seats): $40/month
- ChatGPT Team (5 seats): $125/month
Subtotal: $165/month
Grand total: $1,585/month. That is $19,020 per year.
For a 12-person company. On AI tools alone. Not counting our actual SaaS stack (Slack, GitHub, AWS, etc.), which is a whole separate budget item.
My cofounder, Derek, looked at the spreadsheet and said — and I quote — "We are idiots."
He was not wrong.
Where the Real Money Was Bleeding
The subscriptions were bad. But the API costs were worse, because they were invisible.
The OpenAI Bill Nobody Was Watching
Our OpenAI spend averaged $340/month, but one month it spiked to $612. Why? Because our senior developer, Jake, had built an internal tool that used GPT-4o to summarize customer support tickets. Good idea. Except the tool was sending the entire conversation history as context every time — sometimes 15,000+ tokens per request. And nobody set a spending cap.
Jake's defense: "I thought the free tier covered it." The free tier covers personal ChatGPT usage, Jake. Not API calls running 200 times a day. We love Jake. But Jake cost us about $800 in unnecessary tokens before anyone noticed.
The Overlapping Tools Problem
This one hurt the most. When I mapped out what each tool actually did, I found massive overlap:
- Jasper AI + Copy.ai — both doing marketing copy generation. Why did we have both? Because marketing signed up for Jasper and I signed up for Copy.ai three months later without checking. Classic.
- ChatGPT Team + Perplexity Pro — both used for "research." In practice, 90% of what people used Perplexity for, they could have done in ChatGPT.
- Otter.ai + Fireflies.ai — TWO meeting transcription tools. Two! Our ops manager signed up for Otter, and our sales guy signed up for Fireflies because he "liked the interface better." We were paying $96/month for the privilege of transcribing meetings twice.
- Notion AI + ChatGPT — Notion AI was being used almost exclusively for summarizing meeting notes. Which ChatGPT could already do.
Total waste from overlap: roughly $425/month. That is $5,100/year thrown into a hole because nobody maintained a central "what AI tools are we paying for" document.
The Audit: Cutting Without Crying
After the initial existential crisis, I sat down with each team and asked one question: "If you could only keep ONE AI tool for your work, which one?"
The answers were revealing:
- Engineering: "Copilot. Everything else is nice to have."
- Marketing: "Jasper, probably. But honestly, Claude is better for long-form."
- Operations: "Otter, but only for the auto-summary. The transcription we never read."
- Everyone else: "ChatGPT."
What We Cut
- Copy.ai — gone. Jasper covers everything Copy.ai did, plus some.
- Fireflies.ai — gone. Otter stays.
- Perplexity Pro — gone. ChatGPT Team handles research.
- Notion AI — gone. Nobody was using it for anything ChatGPT could not do.
- Midjourney — downgraded to Basic ($10/month). We were generating maybe 20 images a month. Did not need Standard.
What We Kept (and Why)
- Jasper AI — marketing relies on it daily. Templates save them hours.
- GitHub Copilot — developers would mutiny without it. Genuinely speeds up coding.
- ChatGPT Team — the Swiss Army knife. Everyone uses it.
- OpenAI API — powers our internal tools (with spending caps now, Jake).
- Anthropic API — our support summarizer works better with Claude. We tested.
- Otter.ai — but downgraded to Pro (fewer seats, $40/month).
- Intercom AI — the AI features genuinely reduce support volume by ~30%. Worth every penny.
- Grammarly Business — stays because our client-facing writing needs to be spotless.
- Cursor Pro — 2 devs swear by it. $40/month for productivity gains? Fine.
The New Monthly Bill
After cuts and downgrades:
- Before: $1,585/month
- After: $1,070/month
- Savings: $515/month ($6,180/year)
That is a 32% reduction. And honestly? Nobody noticed. Not one person came to me saying, "Hey, where did Fireflies go?" Because nobody was using it.
But I was not done. The API costs still bothered me.
Taming the API Beast
Here is what I implemented to stop the bleeding on variable API costs:
1. Hard Spending Caps
OpenAI lets you set monthly limits. We set ours at $400. If we hit it, the internal tools gracefully degrade to cached responses. Has it triggered? Twice. Both times, nobody complained because the cached responses were 90% as good.
2. Model Tiering
Not everything needs GPT-4o. I audited every API call and switched what I could to GPT-4o-mini:
- Support ticket classification → GPT-4o-mini (works perfectly, 10x cheaper)
- Data extraction from emails → GPT-4o-mini
- Customer-facing summaries → stayed on GPT-4o (quality matters here)
- Complex analysis → stayed on Claude (it is better for nuanced stuff, fight me)
This alone cut our OpenAI bill from $340 to about $190/month. Model tiering is the single highest-ROI optimization I made.
3. Caching
Jake — redemption arc — built a simple Redis cache for repeated queries. If someone asks the same question within 24 hours, it serves the cached response instead of hitting the API. Our cache hit rate is about 35%. That is 35% fewer API calls for free.
4. The "Do You Actually Need AI for This?" Check
We started asking this question before building any new feature. You would be surprised how often the answer is "no, a regex would work fine" or "this is literally just a database query."
Last month, someone proposed an AI-powered feature to categorize incoming emails. After 10 minutes of discussion, we realized Gmail filters could do 90% of what they wanted. The remaining 10% was not worth $50/month in API costs.
The Real Cost Nobody Talks About: Time
Here is the part that does not show up on any invoice. In our six months of tracking, I estimate our team spent roughly 15-20 hours per month just managing AI tools. That includes:
- Crafting prompts (and re-crafting when they do not work)
- Switching between tools to find which one gives the best output
- Debugging AI-generated code that almost works
- Reviewing and editing AI-written content that sounds like it was written by a very confident robot
- The meetings about which AI tool to use for what
At an average loaded cost of $75/hour for our team, that is $1,125-$1,500/month in labor. Add that to the $1,585 in tool costs and our actual monthly AI spend was closer to $2,800-$3,100.
For a 12-person company.
I am not saying AI is not worth it. Our Intercom AI alone saves us probably $4,000/month in support costs. Copilot genuinely makes our developers faster. The math works out in our favor — barely. But the myth that "AI is basically free" needs to die. It is not free. It is not even cheap. It is just less expensive than the humans it partially replaces, and even that equation is closer than the marketing materials suggest.
What I Would Tell Past Me
If I could go back to August 2025, here is what I would say:
- Maintain a central AI tool registry from day one. A shared spreadsheet. Who signed up, when, how much, what it does. Update it monthly. This alone would have saved us $5,000.
- Default to GPT-4o-mini (or equivalent) and upgrade only when needed. Start cheap. Most use cases do not need the flagship model.
- Set API spending caps before you write a single line of code. Not after Jake runs up a $600 bill.
- One tool per function. You do not need two meeting transcribers. You do not need two copywriting AIs. Pick one and commit.
- Ask "could we do this without AI?" before every new feature. Half the time, the answer is yes.
AI is incredible. It is genuinely changing how small companies operate. But it is also a line item that grows silently, subscription by subscription, API call by API call, until one day you look at your software bill and realize you are spending more on robots than on rent.
Track it. Audit it. Be ruthless about what stays and what goes. Your accountant will thank you.
I built a simple AI spend tracking template in Google Sheets that we use internally. If you want a copy, drop a comment — it is nothing fancy, but it does the job.
📚 Related reading:
