The Signal

A weekly read for operators thinking seriously about AI. Less noise, more signal.

The Lead

Stop measuring usage. Start measuring uplift.

Earlier this month, Amazon quietly pulled back internal visibility on a metric it had been pushing for months: AI token consumption per developer. Leadership had set a goal of more than 80% weekly AI usage. Employees responded the way employees always respond to a vanity metric — they gamed it. Engineers built scripts that pinged the internal AI agent to burn tokens without doing any real work. The behavior had a name internally: tokenmaxxing.

The story is being read as a culture problem. It's not. It's a measurement problem, and it's the dominant measurement problem in enterprise AI right now.

Look at the rest of the week. Ramp's latest AI Index, the one that tracks actual paid business spend, shows Anthropic overtook OpenAI in enterprise adoption for the first time — 34.4% to 32.3%, with Claude usage quadrupling year over year. Anthropic shipped twenty-seven vertical-specific workflow bundles for small business and legal — packaged outcomes, not general-purpose chat. Mira Murati's Thinking Machines Lab launched "interaction models" that explicitly reject the autonomous-agent direction in favor of real-time human-AI collaboration. Three different stories, one underlying shift: the market is finally moving from "how much are we using AI?" to "what is it actually producing?"

Usage is a vanity metric. It is easy to measure, satisfying to put on a slide, and almost completely uncorrelated with whether the AI investment is paying back. Hours logged in Copilot, tokens consumed in Claude, percentage of seats activated, weekly active users — every one of these is the AI-era equivalent of measuring a salesperson by how many calls they made instead of how much revenue they closed. They tell you motion. They do not tell you outcomes.

Uplift is the harder measurement. It asks a different question entirely. Not "are people using the tool" but "what is being produced now that wasn't being produced before?" Not "how many hours did we save" but "what did we do with the hours, and was it worth more than the hours cost?" The honest answer to those questions is almost always uncomfortable. Most AI deployments today are saving hours that get reabsorbed into the same low-value work, or producing outputs that don't change a single downstream decision. The Amazon tokenmaxxing story is not embarrassing because employees gamed a metric. It is embarrassing because the metric was measuring the wrong thing in the first place.

The operators who get this right are going to look strange on the dashboards for a while. Their token consumption per employee might be lower than peers'. Their seat-activation percentages might be flat. They will be measuring something else: new proposals generated, decisions made faster with better information, deals qualified that would have died in the queue, customer questions resolved that would have escalated. The kinds of numbers that don't fit cleanly on a usage chart.

Anthropic took the enterprise lead this quarter because Claude is producing outputs people are using to make money. That is the only metric that matters. Everything else is tokenmaxxing with extra steps.

Three Things Worth Knowing

AI is being sold by the workflow now, not the seat

Anthropic shipped twenty-seven vertical-specific bundles this week — fifteen agentic workflows and twelve one-click templates for small business and legal professionals, integrating QuickBooks, PayPal, DocuSign, and the rest. The shift this represents is significant. General-purpose AI assistants are being repackaged as pre-built outcomes. Payroll-run. Invoice-chase. Contract-redline. If you are still pitching AI to your team as "a smarter chatbot," the packaging has already moved past you. The buyers buying right now are buying jobs done, not capabilities.

DeepMind's co-mathematician is a pattern, not a product

Google DeepMind's AI co-mathematician scored 48% on FrontierMath Tier 4 this week — more than double Gemini 3.1 Pro's raw score — and already helped an Oxford professor crack an unsolved problem. The benchmark is impressive. The architecture is more useful. A coordinator agent decomposes a question, parallel sub-agents propose approaches, a review cycle filters the strongest, and the loop iterates. That pattern is now public, reproducible, and applicable to anything that looks like multi-step expert work. R&D, equity research, legal due diligence, complex underwriting. The operators who treat this paper as an architecture blueprint will move six months faster than the operators who treat it as a frontier-lab curiosity.

Not every workflow should be agentic

Mira Murati's Thinking Machines Lab launched "interaction models" this week — a new paradigm explicitly built for real-time human-AI collaboration across voice, video, and text, with 0.4-second response times that replace turn-taking with micro-turns. The framing matters. The industry has been racing toward maximum autonomy for two years; Murati's bet is that some of the most valuable AI is the AI that stays in conversation with the human rather than running off to act alone. Both directions will coexist. The mistake operators are about to make is assuming the agentic direction is the right answer for every workflow. It isn't. The workflows where speed of human-AI exchange matters more than autonomy — coaching, design review, anything where judgment lives in the loop — will not look like agents.

From the Practice

The revenue uplift frame

Most AI ROI conversations get stuck in cost displacement. Hours saved. FTEs reduced. Dollars taken out. The math is easy because it is subtraction. Two hours per proposal times four proposals per week equals eight hours saved. The story is clean. It is also small.

The harder math, and the one that actually justifies the investment, is uplift. Same engagement, different question. The system did not just save eight hours. It removed the bottleneck that was capping proposal volume at four per week. Now the operator runs twelve. Three of those new proposals close. That is the number that goes on the board's slide. The hours-saved figure is a rounding error against it.

Honest uplift measurement is harder than cost displacement for two reasons. It requires you to define what the system unlocked that was not possible before — which forces an uncomfortable conversation about how much latent capacity was actually being constrained. And it requires a conservative attribution rate, because not every new dollar of revenue is downstream of the AI. Skip either of those and you end up with overclaimed ROI that does not survive the first executive review.

The frame that works in practice is two questions. What did the system unlock that was not possible before, and what is the defensible attribution rate? Operators who answer both honestly will be ahead of the room. Most are still counting hours.

If you are about to be asked to defend your AI investment to a CFO or a board, the conversation is much easier if you have already done the uplift math. Happy to walk through the frame with you in a 30-minute Discovery call.

Book a Discovery Call

Or read this issue on LinkedIn.

Applied AI strategy, training, and implementation.

Ryan Rezeika · Founder · Vancouver, BC
ryan@resonai.ai · resonai.ai · LinkedIn

You're receiving this because you subscribed at resonai.ai/signal. · Unsubscribe