Pegasystem's Ken Stillwell on Tokenomics and Outcome-Based AI

Somewhere along the way, the AI industry started keeping score by the wrong number. Token leaderboards, token maxing, ever-larger consumption as a badge of sophistication. In a recent conversation with Constellation Research, Pega's COO and CFO, Ken Stillwell, argued that this scoreboard was rigged from the start and that the smarter path forward is to stop measuring activity and start measuring value.

Speaking with Liz Miller, VP and Principal Analyst at Constellation Research, Stillwell laid out a refreshingly blunt view of AI economics from the seat of the person who has to answer for both cost and results.

Nobody should be surprised that AI is expensive

Stillwell's starting point is almost disarmingly simple: you have to pay for the capacity you use, and that capacity is not cheap. The industry is spending trillions on infrastructure that will need to be refreshed every year or two. Many assumed the cost would fall quickly. Instead, it has climbed.

But expensive is not the same as wasteful. Things that cost money can still deliver enormous value. The problem is not the price tag. The problem is paying for value you never actually receive.

Who built the leaderboards, and why

Here, Stillwell makes his most pointed observation. Token leaderboards were created by the very companies that profit when you consume more tokens. It should surprise no one that AI model providers want you to use a lot of their product. And it should surprise no one that the companies with nothing to gain from token maxing were never the ones promoting it.

Those leaderboards faded fast once a few voices in the market pointed out how backward the whole exercise was. The lasting question is the more useful one: how do you connect the value of AI to its cost?

AI is a utility, so treat it like one

Stillwell's preferred analogy is electricity. AI is a powerful utility, and like any utility, you use it when you need it. You don't leave the air conditioning running with the windows open. You don't cool an empty house to 65 degrees for two weeks when no one is home. These are intuitions people already carry through their daily lives, and Stillwell believes organizations will develop the same instincts for tokens and AI.

That mindset leads to a more deliberate approach. In design work, for example, token use is heavy but concentrated and short-term, because you are not designing forever. There, it can make sense to lean in hard, because faster iteration and faster decisions are worth the spend. The value of the acceleration justifies the cost.

Right-size the model to the task

Not every job needs a frontier model. Using the most advanced, most expensive model for a simple probabilistic task is its own form of waste. Part of a disciplined AI strategy is knowing when a more static, less costly model will do the job perfectly well.

Just as important is knowing where AI does not belong at all. Stillwell draws a clear line: anywhere a human would otherwise make a judgment, do analysis, or reason is in scope for AI. But in a deterministic workflow built specifically to remove human judgment, introducing that judgment back in does not help. The entire reason the system exists is to avoid it.

The confusion at the heart of the market: outputs vs. outcomes

The conversation's sharpest point is a distinction that Stillwell and Miller both see being blurred across the industry: outcomes versus outputs. Plenty of vendors promise outcome-based pricing, then define the outcome as something like the model returning an answer without hallucinating. Stillwell argues that it is not an outcome at all. It is an output, and frankly, one you should expect to be correct every single time simply by virtue of doing business with them.

A real business outcome is something you have to step back and define. For Pega, that means a completed case: a loan origination, a client onboarding, a dispute, a fraud investigation. The kernel of work that actually needs to get done, ideally automated end-to-end. The value is in completing that work, not in racking up activity along the way.

Stillwell frames it as the difference between activity-based and outcome-based measures. Activity-based metering, counting microtransactions at various skew levels, makes sense when you are selling raw cloud compute. But value-added work is different. He points out how telecom and internet pricing evolved away from counting minutes and toward simply providing access. The lesson for AI is the same. Tokens measure how much processing is happening, not the value being created. And that disconnect is exactly the trap.

The takeaway

Stillwell's message lands as a useful corrective for any leader feeling pressure to prove AI maturity through sheer consumption. Pay for value, not activity. Match the tool to the job. Know where AI helps and where it doesn't. And above all, get clear on the difference between an output you should take for granted and an outcome worth paying for.