Praful Saklani is CEO of Pramata, a leading contract intelligence platform.
For decades, enterprise software followed a simple economic logic: pay a fixed cost, deploy at scale and watch your cost-per-unit shrink. The more you use your software, the better the ROI becomes.
But, the AI token model breaks that logic entirely. Even major enterprises like Uber and Microsoft are seeing the squeeze from ballooning AI costs. The irony is that most of the spend is waste. Not because the AI is faulty, but because people are using it the wrong way.
The Rising Appetite For AI, Budgets Can’t Match It
Budgets were set before people knew just how much AI could do for them. They don’t reflect the current, climbing demand. Enterprise leaders are being dazzled by promises of agentic tools that can automate every single time-consuming task without help.
But, the reality is that this demand needs to be checked. What I mean by this: Most people are relying on AI too much and have a false sense of confidence in what these tools can actually do without human intervention. Herein lies the reason AI budgets are becoming unmanageable for even the most tenured companies.
AI Echo Chambers Wasting Tokens
Companies are asking AI for answers, and rather than verifying outputs themselves, they’re relying on the AI to check itself or employing another tool to check it. Think of it like two confidently wrong people egging each other on. Both of them think they’re right and are going to keep iterating on their answers, but it doesn’t change the facts from not adding up.
Not only does this create an echo chamber of potentially wrong answers, it’s a major waste of tokens. Confident wrongness compounds: each round of AI self-review bakes in the original error more deeply, burning tokens to produce outputs that are further from the truth, not closer to it.
Stripping away the implications of inaccurate responses and looking at this from just a cost perspective, the gamble would be fine if we were dealing with traditional software where the five-millionth query costs a fraction of a penny as the first. But, no economies of scale exist in brute-force AI usage. The curve infinitely continues upward. And these skewed outputs aren’t going to get you the value you’re actually looking for.
There’s actually a very simple solution most are missing.
Solution Part One, Use The Right Tool
Here’s an analogy I keep coming back to: imagine hauling everything across town in a taxi when you own a perfectly good truck. The taxi works, technically. But why pay per mile when the truck is free?
Traditional software, think databases, rules engines and deterministic code, cost roughly 0.1% of what AI inference does. Inference is the phase where AI tools use its training to make decisions. For a huge category of tasks, these tools are perfectly capable of handling the work. They’re fast, reliable and very cost-effective at scale. But companies caught up in the excitement of AI are routing everything through inference when they don’t need to.
The smarter approach is knowing when to use AI and when to use something else. AI should be doing the things only AI can do: pattern recognition in unstructured data, nuanced interpretation, contextual reasoning. It should not be acting as its own quality control layer on top of work it already did.
What You Put In Determines Spend
The other side of this cost problem is precision, or often the lack of it.
Think of it like giving directions to a rideshare driver. If you say, “Take me somewhere in Downtown San Francisco and I’ll let you know when we’re close,” you’ll burn time and money wandering. If you say, “483 7th Street,” you’re going to get there efficiently.
Most companies are doing the equivalent of the first. They’re handing AI a vague objective without the specificity required to make it efficient. If you’re just prompting things like assess this risk, summarize these contracts or flag these issues with no context, the AI is bound to fail. It’s not because it isn’t powerful enough. It fails because it’s operating blindly or with limited context, iterating endlessly to compensate for guidance it was never given.
A day of thoughtful planning before you build an AI workflow can cut costs meaningfully. A week of it can cut them dramatically. The upfront investment in precision pays back in every token you don’t burn on iteration.
The Real Cost Of Getting This Wrong
There’s an economic cost here, and it’s real. But the deeper cost is the erosion of confidence in AI as a legitimate business tool.
When an enterprise spends heavily on AI and gets inconsistent results, the reaction isn’t usually “We need to build this more carefully.” It’s “AI doesn’t work.” That conclusion spreads and as a result, initiatives stall. In turn, the real opportunity gets written off before it has a fair shot.
The enterprises that are going to win with AI aren’t the ones spending the most. They’re the ones spending the most intelligently. They know which tasks warrant AI and which don’t. They’ve done the work to guide AI precisely rather than hoping it figures things out. And they’ve built systems where AI does what it’s best at, while other tools handle the rest.
AI is an accelerator. But you still have to do the driving.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
