Subscribe to Our AI Marketing Newsletter

Issue #31: AI Usage Is Growing. So Are the Bills

TL;DR

AI cost overruns are becoming a real business risk.
The scary cases involve stolen API keys, runaway agents, and accidental loops.
The quieter problem is steady cost creep from unplanned AI usage across teams.
Provider dashboards and alerts help, but they’re not enough by themselves.
SaaS companies need user-level, tenant-level, app-level, and provider-level controls.
Smart model routing, caching, batching, and document preprocessing can reduce costs dramatically.
The goal is not to use less AI. The goal is to use AI with business discipline.

1. The AI bill is not the problem.

The real problem is uncontrolled AI.

That distinction matters.

AI usage often starts small. A few ChatGPT accounts. A few API calls. A developer using Claude Code. A marketing workflow here. A chatbot experiment there.

Then, a few months later, someone looks at the invoice and asks the question nobody wants to answer:

“What exactly are we getting for this?”

That’s where many businesses are heading right now.

Not because AI is bad.

Because AI is being adopted faster than it’s being managed.

2. The horror stories are real.

The obvious risk is the dramatic one.

A stolen API key.

A public repo leak.

A runaway agent that gets stuck in a loop.

A workflow that keeps calling the model again and again because nobody gave it a stopping point.

These are the stories that grab attention because the bills can explode fast. One bad key or one badly designed agent can turn a normal monthly AI bill into a painful finance problem almost overnight.

And yes, these cases are important.

You absolutely need to protect against them.

But they’re not the only problem.

3. The quieter cost problem may be bigger.

Most companies won’t get hit by one giant AI disaster.

They’ll get hit by hundreds of small, normal decisions that nobody measured.

A developer asks an AI coding tool to do something that would have taken 30 seconds manually.

A team member uses a premium model for every task, even when a cheaper one would work.

A support workflow regenerates answers from scratch instead of using reusable templates.

A document workflow sends full PDFs to a model when clean Markdown would have been enough.

A few people do this, and it’s no big deal.

Twenty people do it every day, and suddenly the monthly spend starts looking very different.

That’s the quiet creep.

And it’s easy to miss because everyone can honestly say they’re “using AI to be more productive.”

But are they?

That’s the question.

4. Exploration is good. Unlimited exploration is not.

I’m not a fan of locking teams out of AI.

That’s the wrong move.

Businesses need to experiment. Employees need to learn. Developers need better tools. Marketing, sales, operations, and support teams should absolutely be testing how AI can improve their work.

But “go explore” is not a strategy.

It’s an expense category.

If you give everyone access to AI tools with no usage expectations, no review process, no budget visibility, and no ROI conversation, don’t be surprised when the bill grows without a clear business result attached to it.

Exploration should have a purpose.

What workflow are we improving?

What time are we saving?

What customer experience are we improving?

What cost are we reducing?

What new capacity are we creating?

If nobody can answer those questions, the AI spend may still be useful, but it’s not being managed.

5. Start with the basic security controls.

The first layer is boring, but necessary.

Never expose API keys in frontend code.

Not in JavaScript.

Not in mobile apps.

Not in public repositories.

Not in a demo you think nobody will find.

API keys should live behind your backend, server-side proxy, or AI gateway. That’s where you can monitor usage, control access, and shut things down quickly if something goes wrong.

Then separate your keys.

Don’t use one master key across every app, every customer, every environment, and every developer. That’s convenient until something breaks. Then you have no idea where the problem started.

Use separate keys or projects for different applications, teams, and environments.

Rotate them regularly.

Revoke suspicious keys quickly.

And make sure each key has only the access it actually needs.

This is not advanced security.

This is basic hygiene.

But a lot of AI projects skip it because the prototype worked, and everyone rushed to production.

6. Budgets need layers.

Provider-level budgets are useful.

But they should not be your only protection.

Some provider budgets are alert-based. Some act more like hard caps. Some have delays before enforcement. Some are tied to usage tiers. Some focus more on rate limits than dollar caps.

That means you can’t assume the AI provider will perfectly protect your business from your own application behavior.

You need your own controls too.

For a serious AI product, especially a SaaS product, I’d think in layers:

User-level limits.

Tenant or workspace-level limits.

Application-wide limits.

Provider-level budgets, quotas, or rate limits.

That way, if one user abuses the system, intentionally or accidentally, only that user gets stopped.

If one customer account starts burning through too much usage, that account gets flagged.

If the entire app goes sideways, the application-wide limit protects the business.

And if everything else fails, the provider-level setting becomes your final backstop.

That’s a much better design than discovering the problem when the invoice arrives.

7. SaaS companies need cost controls built into the product.

This is where things get serious.

If you’re building AI into a SaaS application, AI cost control is not just a finance issue.

It’s product architecture.

Let’s say you have hundreds of customers using your app. One customer creates a workflow that accidentally loops. Or one user starts testing edge cases all day. Or one account uses your AI feature far more than the pricing model supports.

If your only limit is at the app level, you have a bad choice.

Let the cost continue.

Or shut down AI for everyone.

Neither is acceptable.

The better approach is to track usage by user, customer, feature, workflow, and model.

Then you can answer questions like:

Which customer is driving the cost?

Which feature is expensive?

Which model is being overused?

Which workflow is creating unusually long prompts?

Which user is suddenly consuming 50 times more than normal?

Without that visibility, you’re guessing.

And guessing gets expensive.

8. The model choice matters more than people think.

A lot of teams make the same mistake.

They start with the strongest model, the workflow works, and then they leave it there forever.

That’s expensive.

The better approach is to start with the best model when the feature is new. Get the user experience right. Prove that the workflow works. Reduce product risk first.

Then quietly test cheaper models behind the scenes.

Run the cheaper model on a sample of the same tasks.

Compare the output.

Check quality, accuracy, formatting, tone, latency, and failure rate.

If the cheaper model performs well enough for that specific task, route some traffic to it.

Keep the premium model for complex jobs, high-value users, edge cases, or fallback.

That’s how you reduce cost without making users feel like quality dropped.

Not every task needs the most powerful model.

Some tasks need reasoning.

Some need extraction.

Some need classification.

Some need formatting.

Some need a short JSON response.

Treat them differently.

That’s where real savings show up.

9. Sometimes the scaffolding matters more than the model.

This is especially true with coding agents and complex AI workflows.

Tools like Claude Code, Open Code, and other agent frameworks are not just models.

They’re scaffolding.

They manage files, tools, context, terminal actions, planning loops, retries, and workflow structure.

In many cases, that scaffolding is the real value.

So instead of asking, “Which expensive model should power everything?” a better question may be:

“Can we keep the workflow structure and route some tasks to a cheaper model?”

That’s where AI gateways, endpoint swapping, OpenRouter, Ollama, LiteLLM, and self-hosted models become interesting.“Can we keep the workflow structure and route some tasks to a cheaper model?”

You may be able to keep the workflow your team likes while changing the inference endpoint underneath it.

The user experience stays familiar.

The cost structure changes.

That’s a powerful lever.

But it should be tested carefully. Cheaper is only better if the output still does the job.

10. The easy savings are usually hiding in the workflow

Not every cost reduction requires a model swap.

Some of the best savings come from basic workflow design.

Use prompt caching when you send the same context repeatedly.

Use batch processing for work that doesn’t need to happen in real time.

Trim old conversation history instead of sending everything forever.

Use structured outputs when you only need specific fields.

Set output token limits so the model doesn’t write more than anyone needs.

And please, don’t send full PDFs to a model unless the visual layout actually matters.

PDFs can be token heavy. If all you need is the text, convert the file to clean Markdown first. Tools like Microsoft’s MarkItDown can help turn PDFs and other business documents into LLM-friendly Markdown before processing.

That one change can reduce waste in document-heavy workflows.

It’s not glamorous.

It’s just smart.

11. Want the full story?

This newsletter is the short version.

The full blog goes deeper into the details we couldn’t fit here, including:

Real examples of AI bills caused by stolen keys and runaway usage.
Which AI providers offer project budgets, spend caps, or rate limits.
How to think about user-level and tenant-level controls for SaaS products.
Why provider budgets should be treated as the last line of defense, not the first.
How model routing, endpoint swapping, caching, batching, and Markdown conversion can reduce cost.
What a practical AI cost-control playbook looks like for a growing business.

Read the full blog here:

AI Cost Control: How to Stop Runaway AI Bills Before They Hit Your CFO
https://kuware.com/blog/ai-cost-control-runaway-ai-bills/

Go to the blog if you want the provider comparison, implementation details, cost-saving techniques, and a more complete framework for building AI usage guardrails before the bill becomes a problem.

12. Final Thought.

AI should create leverage.

It should save time, improve customer experience, increase output quality, reduce friction, or unlock capacity your team didn’t have before.

But AI should not become an uncontrolled operating expense.

That’s the line businesses need to draw now.

Use AI.

Experiment with it.

Build with it.

Train your team on it.

But put real controls around it.

Because the future of AI in business is not just about who uses the most advanced model.

It’s about who builds the smartest systems around it.

Thanks for reading Signal Over Noise,
where we separate real business signal from AI noise.

See you next Tuesday,

Avi Kumar

Founder: Kuware.com

Subscribe Link: https://kuware.com/newsletter/

Signal Over Noise: AI Insights for Business Leaders