What we’re getting wrong about AI efficiency


From Tokenmaxxing to Tokengeddon

Every technology boom develops its own way of measuring progress. The internet rewarded page views. Social media rewarded followers and engagement. Streaming platforms rewarded watch time. AI has its own scoreboards: benchmark rankings, response times, computing costs and token usage. Silicon Valley loves a leaderboard, and the competition to build faster, cheaper and more capable AI models has become particularly intense.

Spend enough time in AI circles and you’ll encounter tokenmaxxing, a term associated with maximizing AI usage and getting as much work as possible out of available tokens. Developers compare usage statistics, experiment with longer prompts and push models through increasingly demanding tasks. Meanwhile, AI companies are working to reduce the cost of processing those tokens. Both behaviors reflect the unusual economics of AI, where greater usage can create greater value while also increasing the cost of providing the service.

For model providers serving millions of users, these calculations matter. Small improvements in computing costs can produce substantial savings. The concern begins when organizations adopt the industry’s obsession with measurable efficiency without examining what those numbers tell them about the quality of work being produced.

Silicon Valley’s Measurement Problem

The technology industry has a long history of confusing measurement with progress. Page views became a substitute for audience value. Monthly active users became a proxy for business health. Social platforms rewarded engagement even when that engagement reflected outrage rather than satisfaction. Each metric captured something useful, but each also created incentives that sometimes worked against the outcome it was supposed to measure.

Economist Charles Goodhart identified this problem decades ago. His observation, commonly summarized as “when a measure becomes a target, it ceases to be a good measure,” remains relevant to the AI industry. Benchmarks help researchers compare models, but performance on standardized tests doesn’t necessarily predict how well a system handles the messy, ambiguous work people do every day. A model can climb a leaderboard without becoming meaningfully better at understanding an organization’s history, advising an executive or recognizing a sensitive stakeholder issue.

Communicators should recognize this pattern. We’ve spent decades explaining that media impressions don’t equal reputation, follower counts don’t equal influence and engagement doesn’t necessarily indicate trust. AI is creating another situation in which the numbers are easier to measure than the outcomes we care about.

The Cost of Context

Consider a communications team using AI to prepare a CEO speech. A short prompt can produce a perfectly acceptable draft in seconds. A stronger draft requires considerably more information: previous speeches, company performance, employee concerns, investor expectations, regulatory developments and the organization’s history. The model needs enough context to understand the audience and the circumstances surrounding the message.

All of that information consumes tokens. Additional reasoning, revisions and alternative approaches consume more. Some of that usage will be unnecessary, and better systems can reduce waste without sacrificing quality. But removing relevant context simply to lower the token count can produce work that requires more human correction or, worse, misses something important.

This is where the definition of efficiency becomes consequential. A cheaper AI response may save a few cents while creating an hour of additional work. A more expensive analysis may help a team identify a risk before it reaches the executive committee. The appropriate measure is the value of the completed work relative to its total cost, including the time people spend checking and correcting it.

When Cost Control Becomes Counterproductive

There’s another development worth watching. As organizations move beyond AI experimentation, finance and procurement teams are examining what they’re spending. Enterprise subscriptions, premium model access and usage-based charges all require justification. Some companies may respond with tighter budgets, usage limits or restrictions on access to more expensive models.

Financial discipline is necessary, especially when organizations are paying for tools that employees rarely use or cannot demonstrate a benefit from. But indiscriminate restrictions could discourage the very experimentation companies have been encouraging. Employees may begin avoiding complex tasks, shortening useful conversations with AI or choosing less capable models because they worry about exceeding a budget.

The consequences can be difficult to measure. A finance report will show reduced spending. It won’t necessarily capture the research someone decided against conducting, the alternative strategy that went unexplored or the employee who returned to a slower process because accessing AI had become too complicated.

Organizations need to distinguish between the cost of using AI and the cost of getting work done. They also need to measure whether AI is improving the work itself. Time saved, errors reduced, decisions improved and employee capability developed are more meaningful indicators than token consumption alone.

Heading Toward Tokengeddon?

I find the tension between tokenmaxxing and what we might call tokengeddon particularly revealing. One part of the AI world is pushing models to do more and consume more computing resources. Another is working to make each interaction cheaper. Inside organizations, leaders are trying to reconcile growing demand with budgets and uncertain returns.

These pressures are legitimate. They also expose the limitations of treating efficiency as a single number.

The most capable organizations will understand when additional AI usage is justified, when a simpler model is sufficient and when human judgment should carry the decision. They’ll examine the full cost of producing useful work and give employees enough room to experiment, question and learn.

For communications leaders, the lesson should be familiar. We’ve spent years challenging measures that reward activity without adequately capturing impact. AI deserves the same scrutiny.

The cost of a token is measurable. The value of the thinking it contributes to is much harder to calculate. Our decisions about AI spending need to account for both.

Comms FutureLAB

Comms FutureLab is a media platform for communications and marketing leaders who are endlessly curious about the future of our profession. Through thoughtful analysis of AI, media, leadership, trust and organizational change, we help today's leaders prepare for tomorrow's challenges and opportunities.

Read more from Comms FutureLAB
AI Workflow Automation: Examples, Tools, Best Practices

We're Changing How We Use AI. Are We Changing How We Manage It? AI is becoming more involved in how we work. A recent announcement from OpenAI raises questions about what happens when the technology we've come to depend on changes. If you've spent hours building a custom GPT, you might want to pay attention to December 11, 2026. That's the date OpenAI plans to retire custom GPTs and move users toward plugins. For anyone who has invested time creating a specialized AI assistant, this is a...

When Every Meeting Is Recorded, What Happens to Candor

When Every Meeting Is Recorded, What Happens to Candor? AI meeting assistants are becoming part of everyday work. New research raises questions about what happens when those tools start evaluating the people in the room. I recently attended a meeting with almost as many AI agents as human participants. 😅 Between meeting assistants, recording devices and smart glasses, I’m starting to wonder how much of our working lives are being documented, analyzed and stored somewhere. A few weeks ago, I...

Your Embargo Needs an Internal Comms Plan I’m working on a project with an organization that has a longstanding practice: employees hear company news 24 hours before it’s announced publicly. I understand the reasoning. Employees shouldn’t routinely learn important news about their company from a news alert, LinkedIn post or reporter. But 24 hours can be a long time when you’re also working with media under embargo. An embargo gives reporters access to information before an announcement with...