What is it about AI that's actually bothering you?

Somewhere in the last two years, "we need to figure out AI" became a standing agenda item. The features shipped, the customers used them, the bill showed up. And now there is a vague unease in the room that nobody has quite named.

Here’s the problem with vague unease: you can’t fix it. "AI is troublesome" is not an actionable statement. It’s six different problems wearing one trench coat, and they have different owners, different price tags, and different solutions. Before you buy anything, hire anyone, or form a committee, it’s worth pulling the trench coat off and looking at what is actually underneath.

In our conversations with SaaS companies, the troubles sort into six piles.

1. Model selection and routing. Which model should handle which task? The expensive frontier model writes beautiful summaries, but so does one that costs a tenth as much. Teams burn real money sending simple jobs to premium models, and real engineering time arguing about it. A growing set of gateway and router products exists to solve exactly this.

2. Overages and enforcement. One customer discovers your AI feature and runs it ten thousand times overnight. One retry loop misfires. Without rate limits, quotas, and guardrails, a single bug or a single enthusiastic user can torch a month's budget before breakfast. Again, there are good tools for this, mostly living at the gateway layer.

3. Performance and caching. Users notice latency. Finance notices that you paid full price to generate the same answer four hundred times. Caching and performance optimization are real disciplines with real vendors behind them.

4. Quality and reliability. Is the output good enough to put in front of a customer? How do you know it stayed good after the model provider shipped an update you didn’t ask for? Evaluation and monitoring platforms have grown up around this question.

5. Governance and risk. Where is customer data going? Which vendors are you dependent on? What does your auditor need to see? This one lands on the desks of legal and security, and the tooling here is maturing fast.

6. Cost. Not the bill itself. The bill is easy to find. The trouble is that the bill tells you what you spent and nothing about where it went. Which customers drove it? Which features? Which of your pricing plans is quietly underwater because a handful of accounts use the AI feature fifty times more than the average you priced against? Your model provider's dashboard shows tokens by API key. Your finance team needs margin by customer. Between those two views sits a canyon, and most companies are standing at the edge of it squinting.

Six piles. Different problems, different buyers, different tools. So here is the first useful question: do you actually care about all six?

Probably not equally. If your outputs are embarrassing you, quality is your fire. If a customer just blew through your budget, enforcement is. But in most of the conversations we have, when the vague unease finally gets named, it gets named in dollars. The bill is growing faster than the revenue attached to it, nobody can say which customers or features are responsible, and the CFO is being asked to defend a number nobody can explain. If that’s your situation, your problem is not AI in general. Your problem is cost attribution, and you should look for tooling that connects AI spend to customers, features, and margins, not just tooling that makes individual API calls cheaper. A router can cut the price of a call by thirty percent while you keep losing money on your biggest account. Cheaper is not the same as explained.

Now the second question, and this is the one I would ask you to sit with. Imagine you solve it. Imagine your AI costs are attributed, forecasted, and under control, and the hit they were making on your profitability is fully managed. Would you be satisfied?

Really think about it, because I suspect the honest answer is no. And the reason is worth noticing: what was bothering you was never actually the AI line item. It was that a growing chunk of your cost of goods sold varies wildly by customer and you cannot see it. AI just made that variance too big and too fast-moving to ignore. But it was never the only variable cost. Compute varies by customer. Storage varies by customer. Data transfer, third-party APIs, and support load all vary by customer, and they always have. Some of your customers have always cost multiples of what others cost to serve on the same plan at the same price. AI did not create that problem. It turned the volume up loud enough that you finally heard it.

To be clear, the other five piles don’t vanish just because cost is the deepest one. You may still need a router, a cache, an evaluation harness, and answers for your auditors, and you should go get them. Those are AI problems, they live inside the AI layer, and the tooling for them is good and improving. Cost is the odd one out. It only looks like an AI problem. Scratch it and you find a business problem that was there before AI and will be there after.

This means that if you treat cost like the other five and chase the AI slice alone, you will do this again. Next year there will be a new line item growing uncomfortably fast, and the year after that another one, and each time you will shop for a tool shaped like that layer. Cost layers change every few years. The question underneath them does not: which customers, features, and products are profitable, and what should we do about it?

So flip your axis. Stop organizing your cost intelligence by layer, where AI gets one tool, cloud gets another, and nobody owns the whole picture. Organize it by the business question, and let every layer, including whatever arrives after AI, feed the same answer. The companies that do this will set their prices with their eyes open. Everyone else will keep treating each new cost layer as a brand new crisis, when it was the same old question all along.