Did you know that Uber used up its AI budget in just 4 months? Chamath Palihapitiya’s software startup, 8090, claimed that its AI expenses are ballooning rapidly and it could reach $10 million a year. Microsoft, Meta, Amazon, and many others shut down tools that were consuming AI tokens and revoked AI licenses because of rising costs. All these examples point to the same thing. AI spending is no longer predictable. You need to treat AI consumption like electricity. It can explode as adoption grows. It’s not a fixed line item in your budget statement. And most importantly, you need to build cost controls into every workflow that has AI’s footprint.
Key Takeaways
|
Before we understand the ‘variable’ part, let us first understand why AI costs are ballooning in the first place. In business, AI is not just about answering simple prompts. You don’t ask it to write a simple LinkedIn post. AI in business is all about completing tasks on its own. In the former case, it consumes a few thousand tokens. In the latter, you will burn 10X or even 100X that number.
Here is a table summarizing the average number of tokens used according to task type:
|
Type of Task |
Range of Tokens Consumed |
|
Simple task |
200–2,000 |
|
RAG query for retrieving documents |
2,000–12,000 |
|
Agentic tasks |
>1,000,000 |
Ultimately, it all boils down to how much work the AI performs under the hood. Now, imagine the number of tokens consumed when an agentic task is repeated across dozens of functions and hundreds of employees. Your AI bill will hit the roof before you even know.
Now, coming to the variability part, let us understand this with a simple example. An employee asks your HR bot a question about the leave policy. The bot has been trained to answer every query after reading everything it has been handed over in the past. Now there are two scenarios:
In both scenarios, the output is similar. The employee gets the answer. But behind the scenes, it's a world apart. The math is simple. 20 pages cost 10X what 2 pages do. But in scenario 2, the extra 18 added nothing to the answer. Just to your bill.
Now the agent will receive thousands of questions a day, especially if you have thousands of employees. This small difference in just ‘retrieval tuning’ adds a big chunk to your invoice. That’s why two companies running the same agent on similar workloads will end up with very different invoices. One is paying for answers. The other is paying(more) for clutter.
There is also one hidden factor that adds to your bill. It’s what we call the ‘success penalty’. It catches most business leaders off guard. Your pilot may be a grand success in light loads. Even the invoice might have been manageable. But push it to normal load. The usage will climb faster than you can comprehend. This is not an anomaly. It’s what adoption actually looks like when any tool is useful. Most AI budgets, unfortunately, do not leave this room for fast growth.
Also read: Why 80% of Your Team Feels More Productive With AI but Your P&L Hasn't Moved
Don’t get us wrong. We are not asking you to cut down on your AI investments. Cost should be part of how you design the system. Not something you discover after it’s already live. But there are certain simple steps you can take to control your AI bills.
First, not every task needs the most expensive model. You can send your most complex reasoning tasks to top-tier models like Claude Opus or GPT-5. But give your everyday jobs like summarizing or formatting to cheaper models like Claude Haiku or Gemini Flash. You’ll barely notice a drop in quality. But the price gap between tiers is more than 10x!
Secondly, just like an electricity bill that doesn’t tell which appliance consumed the maximum units, your monthly AI bill alone doesn’t tell you much. It doesn’t tell you that one of your workflows is eating 60% of AI tokens. That’s why you should tag your AI usage. Link every dollar spent to a team or a project. Then add budget alerts which will notify unusual spikes. For instance, a stuck agent will show up as an alert early. It won't be a nasty bad on the invoice. Essentially, when your invoice reaches unexpected numbers, you can always explain it to leadership with a cause.
If you think today’s AI prices are low and manageable, you are not wrong. The prices are low. That’s because AI vendors are still fighting for you in an attempt to gain market share. Behind the scenes, they are absorbing part of your costs. But such magnanimity will not last long. The race is going to settle soon, whether we like it or not. And you should expect pricing to move further north, probably close to what it actually takes to run these systems.
Some early signs are already visible. GitHub, the popular web hosting and collaboration platform, has moved Copilot to usage-based billing already. Others are likely to follow. They know that flat and predictable pricing will no longer hold, and paying for what you consume is a far better option.
So don’t build your AI budget around today’s rates. You need to assume per-unit prices will rise at some point in time. Based on that assumption, you need to test whether your AI plans still make sense at a higher cost.
If you are putting AI to work inside Salesforce, either through agents or copilots on top of your CRM data, the same cost problem we discussed above will apply. At Brysa, we help you plan for it before it shows up on an invoice. We follow a three-step approach. One, look at how your agents actually use your data. Identify tasks that could run on a lighter setup without anyone noticing a difference. Three, we help you put the basics in place. This includes tracking usage and setting up alerts that flag unusual spikes early. Our goal is simple: when your AI adoption and usage climbs faster than expected, your budget is ready for it. As we said earlier, our aim isn’t to hold back your AI investment. It’s to make sure every dollar is paying for an answer. If your Salesforce AI costs are getting harder to predict, talk to us before the next bill does.