The Cheapest Token Is the One You Never Waste: Why Intent Is a FinOps Discipline
As a FinOps advisor, I spend most of my time helping enterprises govern the variable cost of cloud. Over the past 18 months, a second variable cost has landed on every CFO's desk: the cost of intelligence itself. AI spend is now a standing agenda item in the executive reviews I run, and the questions sound familiar. Where does the money go? Who owns it? How do we know we are getting value?
We call this discipline Tokenomics: FinOps applied to AI. And the more assessments I run, the clearer one insight becomes. The biggest driver of wasted AI spend is not the model you choose or the infrastructure you run it on. It is the quality of intent that goes into every single prompt.
The intention-expression gap has a price tag
There is almost always a gap between what you want from an AI system and what you manage to say up front. In IBM's we call this the intention-expression gap. You know more than you can articulate, your intent evolves as you see drafts, and your assumptions hide in plain sight.
In a world of flat-rate chat subscriptions, that gap costs you time. In an enterprise consuming AI through metered APIs, it costs you money, and it compounds fast. Every misaligned draft is a full round trip of input and output tokens. Every regeneration resends the entire context window: system prompt, conversation history, retrieved documents, all of it, with every request. Practitioners building production agents report single tasks consuming 50,000 to 100,000 tokens once multi-step reasoning is involved. If the intent behind the task was fuzzy, a meaningful share of those tokens produced something nobody wanted.
The research backs this up. Studies on prompt optimization show that context-aware, well-framed prompts reduce token consumption by 15 to 30 percent while simultaneously improving output quality. Semantic caching of repeated queries can cut API costs by up to 73 percent in high-repetition workloads. These are not marginal gains. They are the same order of magnitude as the savings we chase in classic cloud FinOps through rightsizing and commitment management.
Three tactics that turn clarity into savings
The good news: closing the intention-expression gap does not require new tooling. It requires a different first move. Three tactics, all of which I now use daily in client work, make the difference.
- Socratic prompting: Instead of asking AI for the deliverable, ask it to interview you first. Before you write anything, ask me what you need to know. In my FinOps capability assessments, I use exactly this pattern before building any client deck. A handful of cheap clarifying questions, a few hundred tokens each, routinely saves me 3 or 4 full regeneration cycles of a 15-slide deliverable. The cheapest tokens I spend all day are the ones the model uses to ask me questions.
- Meta prompting: When you cannot articulate what you want, ask the AI to write the prompt for you. This works because describing the shape of a good request is often easier than making the request itself. One well-constructed prompt that runs once beats a vague prompt that runs 5 times.
- Generate-and-react prompting: Humans recognize what they want faster than they describe it. Asking for 3 genuinely different framings of a problem, then reacting, converges on the right answer in fewer total tokens than iterating blindly on a single draft. Divergence up front is cheaper than correction at the end.
What this means for FinOps teams
If your organization is building an AI cost governance practice, I would argue prompt discipline belongs in it, right next to model tiering, caching, and routing.
Three concrete implications:
- First, measure the ratio, not just the spend. Tokenomics is a value question. The goal is not fewer tokens, it is more value per token. A team whose spend doubles while their regeneration rate collapses is winning.
- Second, treat prompt quality as an optimization lever. The established five-layer view of AI cost optimization starts at the application layer, and intent quality sits at the very top of it. It is the one lever every single user in your organization pulls dozens of times a day, no platform migration required.
- Third, make it a capability, not a memo. In the FinOps Framework, this lands squarely in workload optimization and in FinOps education and enablement. Teach the tactics. Put them in onboarding. Track whether they move your unit metrics.
The intention-expression gap will never fully close. Intent evolves through dialogue, and that is precisely how it should work. But there is a difference between productive iteration and expensive guessing, and the line between them is drawn before the first token is ever generated.