Token Waste Is a Symptom

Giving AI to employees without training has cost your enterprise in time, money, and people. Studies from GitHub suggest between 40% and 60% of all the money spent on AI tokens is wasted. Burned away in repetitive steps, unclear instructions, and sloppy outputs to rework. And the answers on offer for reducing token waste are unconvincing:

• Talking like a “caveman”
• Limited access hours
• Cutting context every other step

From simple token waste to growing internal and structural tensions, a lack of honesty and knowledge around AI is eating away at financial and cultural stability. If this continues, enterprises will find themselves under worsening financial and operational pressure, forcing more of them to shutter operations. The solution is simple, if not pretty or popular: enterprises must seriously educate themselves and their employees on AI.

Rules written in blood vs cash

“Safety rules are written in blood,” my boss at Quiznos chided me after I sliced my fingers open cleaning a deli slicer the wrong way. Trust me, that mistake was never made again.

But enterprises still find themselves bleeding bit by bit, using AI in shockingly inefficient ways. Instead of blood, enterprises are bleeding cash, and because of that, the mistakes and bills that continue to pile up are written off as the cost of doing business.

Metric	April 2026	Estimated waste
Median monthly cost	$2,246	$898 to $1,348
Average monthly cost	$140,842	$56,337 to $84,505
75th percentile (top 25%)	$14,843	$5,937 to $8,906
90th percentile (top 10%)	$73,030	$29,212 to $43,818
95th percentile (top 5%)	$211,409	$84,564 to $126,845
99th percentile (top 1%)	$831,338	$332,535 to $498,803

But it doesn’t need to be this way. By investing in the machine’s operator, we can reduce the human and financial costs of AI transformation.

Drill, baby, drill!!!

There are plenty of other blogs that focus on the individual steps and tactics to reduce token expense. We’re going to talk about a few philosophical and strategic mindsets that reduce token costs:

• Training should focus on competency, not familiarity.
• An operator who understands their machine uses it better, more creatively and more efficiently.
• Enablement is a force multiplier; mystery is a force divider.

These mindsets are not anti-AI propaganda. They come from accepting that AI is a commoditized technology to be used by people, so it is the quality of the people that affects outcomes, not the models. Giving your employees continued, in-depth, role-specific training is the single most effective way to reduce token waste in the short and long term.

“If you needed me to tell you that, I’m glad I did”

Teaching and training people is difficult, in part because everyone makes assumptions about what is common knowledge. So why do we assume employees truly understand a technology that has only recently become commonplace? If you didn’t know that an engine uses gasoline, would you think to look at the fuel gauge? Would its purpose be clear?

The assumption of common knowledge is at the heart of token waste. Enterprises and employees were not taught how a model works, what it costs, or its strengths and weaknesses. Instead, the hype around AI’s speed and capabilities led to a “move fast and break things” approach. That led to runaway agents repeating doomed tasks in the dark of the web, employees rolling over each conversation, and investment in fruitless AI projects. When people are taught better, they tend to do better. In an industrial-sized rollout, enablement can be a matter of life and death.

Enable or die: the M-14 debacle and AI

The M-14 rifle was a manufacturing disaster that cost billions of dollars and untold lives, and led directly to the end of the US Springfield Armory. While there were many contributing factors, the most important one was the rushing of the Materials Requirement Document, or MRD. An MRD should contain everything a manufacturer needs to know to start mass-producing an item. In other words, it was an enablement package. The rushed MRD for the M-14 led to deep, fatal flaws in the manufacturing process for the barrel. Instead of merely slowing down delivery, the lack of enablement led to incorrect work being done at scale, turning a single issue into a large-scale mystery and crisis.

The lesson is that slow work can be fixed, but incorrect work is harder to catch, causes more problems, and introduces uncertainty. Each time an unenabled team uses AI, that incorrect work shows up on the bill as token waste, and in the other costs: time spent fixing errors, parsing hallucinated responses, or running additional review cycles. This changes the math. It’s not just compute spend vs productivity gain, but compute spend vs the cost of incorrect work delivered at machine speed.

The right people with the right tools

The point of role-based, tool-specific training isn’t better prompt engineering. It’s having an engineer who knows what they are optimizing for. It’s having an analyst in the workflow who can call out incorrect answers. It’s having a marketing manager who knows which models to use for analysis and which for generative tasks. It’s having employees knowledgeable enough to know when to override a model’s choice. Those capabilities are not developed in general AI literacy courses. They come from structured readiness built around the tools your enterprise has actually approved and deployed.

Token waste is a symptom. The cost you can’t see is what an unenabled workforce produces with those tokens: confident, fast, and sometimes wrong.

If your token bill is growing, the symptom is visible. Clairet treats the cause. See how.

Sources: GitHub, “Improving token efficiency in GitHub Agentic Workflows,” May 7, 2026; Ramp, “How Much Do AI Tokens Cost Businesses? 2026 Spending Benchmarks,” June 8, 2026 (April 2026 data).