I open today with two recent stories for your consideration.
The first, from Bloomberg, titled AI-Fueled Profit Growth Set to Flow Through Every S&P 500 Sector, details the impact AI is having across industries, with third-quarter earnings expected to have grown in every S&P 500 sector for the first time since 2021.
Highlights: the data center and capex spend is playing out far and wide, of course, but also consumer firms are seeing impacts from AI shopping integrations, customer service, and distribution boosts; and healthcare and finance are both seeing significant gains and willing to quantify their AI-related growth.
AI success is spreading beyond tech, in other words, and achieving enough stability to begin yielding fruits for leading players across industries.
The other story is from the Wall Street Journal. It details how the SaaS-provider Retool recently discovered an out-of-office email responder created using AI was costing the company more than $10,000 a day.
For autoreply.
The tool was unnecessarily processing millions of tokens daily, a discovery CEO David Hsu called “really shocking.”
Today I’m focused on this complicated reality of AI automation costs vs gains.
Enterprise AI continues to be an organizational challenge
As McKinsey Senior Partner and Global Leader of Technology and AI Kate Smaje noted at the WSJ Leadership Institute’s Technology Council:
“The ROI is incredibly concentrated at the moment. And what you find is most people, like 80% plus, will say, I get personal productivity out of this, right? … But if you then translate that down to, OK, is there something that the CFO can actually see… That number drops very, very fast…”
Her bottom line: “I think the reality is, no, most people are not getting value from it, but those that are are incredibly concentrated, and it’s about learning what they do differently.”
The banner stat above (that around one in 10 CEOs say their AI is both generating more revenue and lowering costs) is from PwC’s 2026 Global CEO Survey, which also found that 56% say AI is still not showing any significant financial benefits.
Around one in eight are seeing both benefits, with organizations that have strong AI foundations unsurprisingly 3x more likely to report significant returns.
It’s not a new story, but business moves far slower than AI does, and organizational challenges (like lacking a working AI implementation strategy) continue to be a drag on the flywheel.
Julie Averill was the CIO of Lululemon when the company grew from $2 billion to over $10 billion and expanded over four continents. Before that she led tech transformations at companies like Nordstrom and REI.
She recently wrote an editorial for the New York Times (called Companies Need to Stop Kidding Themselves About AI) that presented many of the challenges as she’s experienced them.
Businesses, she noted, are extremely fast to proclaim their AI use and to put a lot of energy and even money behind it, but often without the strategy to really support it.
The pilots are more amazing than ever before, drawing leaders into what she calls a state of “AI wishing.” This is moving too fast and unevenly into projects that either aren’t planned out well enough or are expected to miraculously leap over real-world challenges (like lacking clean, connected data).
Some leaders are caught up in the belief that AI, as amazing as it is, is faster and easier than is truly possible of any technology, and in the process, their companies are missing out on many of its top benefits.
Some of the things called out: spending too much money on the wrong things, cutting workers while they’re still needed, ignoring real-world exceptions and workarounds that are present most everywhere, and also “AI washing,” or doing AI more for show and to keep up with rivals than to genuinely transform the business.
Of course, at PTP we have experienced many of these same things.
And in the end, they add up to make the difference between riding the AI wave and still treading water with AI pretending.
Tokenmaxxing, tokenminning, and AI tokenomics
I can’t write about AI costs without at least touching on tokens, that often-vexing measurement sometimes wrongly compared to the metering of a utility.
As economist Howard Rubin told Lydia DePillis of the New York Times: “It’s a currency where you have no instinct to know what you’re using, and the accounting practices aren’t even there for it.”
As a measure of AI compute needed, tokens have a real-world impact on electricity and chip use at data centers but are actually a measure of chunked data being ingested and returned.
‘Tokenmaxxing’ was the popular term for a window of AI wishing (and valid experimentation, to be fair) this spring. It saw companies pushing their developers to spend as many tokens as possible in their coding work, with some crafting leaderboards to spur competition.
Fast-forward a few months, after the bills came in (much higher than expected), and the entire time looks some like a tech company fever dream.
Uber CTO Praveen Neppalli Naga explained to The Information how he’d had to go “back to the drawing board” after developers had burned their entire annual budget in just a few months with Claude Code.
And they weren’t alone. But when the dust cleared, many (like Uber) admitted to having learned valuable lessons (if costly) that have helped them adapt and slash costs. One takeaway cited by Uber is to treat efficiency as an engineering problem. Another is to use intelligent model routing (see below).
(One AI consulting firm told Axios about a client whose unbridled Claude Code enthusiasm caused a $500 million AI bill in just one month. As in the Retool case above, some employees were using high-end AI agents to do simple things like check the weather all day long.)
‘Tokenminning’ was maybe jokingly coined to refer to the backlash, with CFOs tasking teams to watch, weigh and compare their token use, in a bid to better understand what the business is getting for its money.
For most, it became immediately clear that more tokens doesn’t necessarily mean more value.
Today this has informed the AI cost optimization practice many call ‘tokenomics.’
What is AI tokenomics and why does it matter for enterprise AI costs?
In June, the Linux Foundation launched the Tokenomics Foundation with the goal of helping businesses make sense of AI tokens, through open standards, benchmarks, and infrastructure best practices.
The need is clear: every AI interaction has a cost, but it can be very difficult to estimate (and opaque to chase).
Reasoning models, for example, use far more tokens than non-reasoning (needed or otherwise). You can see this by giving the same prompt to GPT-5.5, GPT-4o, and GPT-3.5.
In a New York Times example, GPT-5.5 used 14 times as many tokens as GPT-3.5, and 8 times as many as GPT-4o.
Also important here: based on the chips and energy used, newer models like Anthropic’s Mythos, for example, can cost 10 times as much per token than cheaper Claude models.
Meaning it’s critical to ask it for the right things in the first place.
And even though AI token costs have continued to drop since August (overall they’ve trended down since March), the age of the one-size-fits-all LLM is officially over.
Still, as the foundation demonstrates, most get hung up on tokens and API costs simply because they’re what’s most visible.
How should companies calculate the true cost of using AI across business workflows?
The full picture has to include things like infrastructure costs, compute, memory, storage, data, energy, software licenses, and outcomes, labor, value realization, and of course revenue.
AI model routing is part of the solution
Many companies have documented the challenges they faced shifting approaches to suddenly alert, gate, and channel tokens so that they only go to most expensive frontier models when the need is really there.
The popularity of model routing options has risen from this challenge. So what is AI model routing and how does it reduce LLM costs?
It’s a gateway that sits between application and models and is born from the goal of not sending every request to the same place.
Routers can evaluate incoming requests and use preset preferences (or knowledge) to determine which is the best fit. This can be based on the task or allowed cost, and can take into account things like availability, latency, data risks, privacy policies, and more.
They’ve shown promise bringing significant savings for businesses. Amazon Bedrock noted its Intelligent Prompt Routing, for example, cost 60% less in internal testing, and that just by routing between Anthropic models.
As pointed out by the Madrona Venture Group’s Vivek Ramaswami (covered in one of our recent AI news roundups):
“Do you need to drive a Ferrari everywhere?
Probably not.”
What is the best strategy for implementing AI that delivers measurable business ROI?
Another PwC study from April put numbers on this trend that’s being widely discussed: 20% of companies are capturing 74% of AI’s economic gains.
How?
By redesigning workflows with AI, instead of approaching it as a tool adoption; by applying AI to growth; and by increasing automation while at the same time utilizing real, working AI governance.
At PTP, our Enterprise AI VOICE Framework came about for these reasons.
It was built from the lessons we’d learned first-hand, observed through work with partners, and seen validated across the industry at large.
Effective AI use almost without fail begins with business need, not with a technological capability or impressive demonstration.
This means understanding where an organization really is in terms of readiness (not where they wish they are) and starting with use cases that are both easy to implement and support and which also bring real value.
Depending on capability and need, ambitions can scale, but action should always follow an honest, effective understanding of what’s needed, what AI can actually do, and how it’s going to happen.
Work also should not begin while leaders are still wondering how can businesses measure the ROI of AI investments?
Metrics must be established so that success and failure are clear to everyone involved.
This should drive scaling and ongoing improvements, and begin with measurable business outcomes, including costs removed, revenue gained, quality impacts, errors avoided, and hours re-applied.
Consider the full picture as discussed above. It’s not just tokens or licenses, but also costs around human review and redoing work, governance, and infrastructure requirements that make up AI charges.
Businesses also must match the capability to the job, and here I mean both with AI overall (repeatable, verifiable, acceptable risk) and with the model and software harness used.
An AI agent probably doesn’t need to handle your out-of-office replies, and likewise boilerplate code adjustments may not need to go to the most sophisticated solution.
You also shouldn’t be attempting to automate exception-laden workflows cobbled together over the past decade. Be realistic about the work and orchestrate the right workflows that don’t burn money on AI agents to spin their wheels trying to figure out the best next step to take.
Process is critical to success, as is, of course, the state of necessary data. Don’t be unrealistic about AI’s ability to use data effectively from a variety of sources which may together introduce duplication or formatting challenges or expose the organization to too much risk.
When everyone’s using the same handful of AI models, an organization’s competitive advantage will come, in part, from how they integrate the people and their skills.
I believe it’s essential to use them correctly from the start. Don’t hold back AI while a human checks every single thing painstakingly, nor expect real, detailed review of every output to be reasonable or effective.
Voice systems require clear escalation paths that are fast and effective, and the same logic applies throughout the business.
And ultimately, it’s essential to ditch the losers and scale the winners.
As Retool CEO David Hsu said: “you have to enable a lot—and then you have to kill fast and also invest further.”
So how can companies reduce AI costs without sacrificing performance or quality?
If you’ve read this far with this question still in mind, my short answer is this: know yourself, know what AI can really be expected to do, pick the right use cases, and govern very well.
Accurately predicting all the costs is still problematic, as we’re all still learning to meter, route, optimize, and standardize.
It’s one of those AI truisms that if the world stopped all AI development today, we’d still have at least a decade of work just to figure out how to use what we have.
I’d say it’s more like two decades. The business is what moves slower, and so does a lot of the dirty work that still needs to be done.
The winners with AI right now understand this and are in the process of getting it done.
References
An AI Out-of-Office Reply Cost This Company $10,000 a Day, The Wall Street Journal
CEO confidence in revenue outlook hits five-year low – as AI becomes a defining divide between leaders and laggards: PwC 2026 Global CEO Survey, PwC
I Helped Run Lululemon. Companies Need to Stop Kidding Themselves About A.I. and What Are Companies Getting for All That A.I. Spending?, The New York Times
After blowing through its entire 2026 AI budget in months, Uber CTO says, ‘We’re coming to the end of the so-called tokenmaxxing era’, Fortune
Cracks in the AI Thesis Part 2, Ramp AI Index
