Last Call for Tokenmaxxing: Why CFOs Are Cutting Off the AI Open Bar
I've spent the last year sitting across the table from companies at almost every stage of the AI curve: an early-stage LLM infrastructure startup trying to prove its technology actually saves money, a wellness business layering AI into a growth model, and a veterinary company using AI to accelerate its diagnostic capabilities in local clinics. Before that I also partnered with customers at Microsoft, building AI and cloud go-to-market strategy for Fortune 500 companies, including the commercialization case for a custom AI deployment at a top-10 global pharmaceutical company that had to prove a real payback period before anyone in finance would sign off.
The pattern I keep running into is the same one, whether the company is a five-person startup or a household-name enterprise: everyone can tell you what their AI does. Almost no one can tell you what it's actually worth.
This Isn't a Small-Company Problem or a Big-Company Problem.
Here's what I find genuinely interesting: the companies burning the most money on unmeasured AI spend aren't necessarily the ones you'd expect. I've seen it happen inside AI-native startups that should, in theory, know better than anyone how to use the technology, and I've seen the exact same pattern inside large enterprises with mature finance functions and layers of governance. The size of the company doesn't protect you from this problem. If anything, it just changes how many zeroes are attached to the mistake.
And the finance function has noticed. Gartner projects worldwide AI spending will hit $2.52 trillion in 2026, a 44% year-over-year increase, but fewer than one in three organizations can tie their AI investments to an actual P&L change, and just 15% of AI decision-makers reported any EBITDA lift in the past year (Intraverse AI). Forrester research found that enterprises are now postponing 25% of planned AI spend into 2027 specifically because the financial scrutiny has caught up to the enthusiasm (AI Business Weekly). The first wave of AI adoption happened fast, with CTOs and AI teams making a simple argument: move now, or fall permanently behind. And finance mostly went along with it. That era is over. CFOs are in the room now, asking harder questions, and a lot of teams don't have good answers.
This isn't abstract. It's showing up in real budget decisions at real companies, and it's costing people money, credibility, and, increasingly, their jobs on the project.
The 5 AI ROI Traps I See Over and Over
1. No Clear Target Outcome, Defined Up Front
This is the trap that makes every other trap on this list worse, so it goes first. If you can't say, in plain, tangible business terms, what you're trying to improve before you start building, you don't have an AI project. You have an expensive hobby.
The outcome has to be measured in something a CFO would recognize without translation: throughput, productivity, customer satisfaction, cost per transaction, time-to-resolution. Not "we're exploring what's possible." If you can't make the case, before you spend a dollar, that this will move one of those numbers, that's a gating factor, not a detail to figure out later. I've watched this exact gap sink otherwise well-funded AI initiatives: the technology worked fine, but nobody had defined what "working" was supposed to mean in dollars or hours.
2. Automation for Automation's Sake
Related to trap #1, but worth calling out on its own because it's so common: teams start "experimenting" with AI because it feels like the responsible, forward-looking thing to do, without ever defining what the automation is supposed to achieve. There's no end state, no success metric, no point where the team can say "we're done, and here's what we got." It just becomes an open-ended initiative that consumes budget indefinitely because nobody set the boundaries in the first place.
Experimentation has its place. But experimentation without a defined destination isn't strategy. It's just spending with better PR.
3. “Tokenxmaxxing”, a.k.a., Using the Wrong Model for the Job
This is the one I probably see most often, and it's usually invisible until someone actually looks at the bill. Companies default to the biggest, most capable (and most expensive) model for every task, regardless of whether the task actually needs that much horsepower. You don't need a tank to mow your lawn.
The fix here is genuinely more accessible than most people realize. There's now a real, competitive market of model routing and selection platforms built specifically to match the right model to the right task automatically: simple classification jobs go to a fast, cheap model; complex reasoning goes to something more capable. Done well, this alone can cut inference costs by up to 90% without touching the quality of the output. If nobody on your team can tell you why you're using a frontier model for a task a much cheaper one could handle just as well, that's real money sitting on the table.
4. Not Choosing the Right Data to Tune the Model
This is a harder trap to fall into gracefully. But the principle is simple even when the execution isn't: a generic model fed generic prompts gives you generic output. The more of your own proprietary, high-quality data you responsibly bring into the loop, through fine-tuning, retrieval, or context, the more the output actually reflects your business instead of the internet at large.
This is also where the ROI math gets real. A model tuned on your own claims data, your own service tickets, your own customer interactions will simply outperform an off-the-shelf model on your specific use case, and that performance gap is exactly what turns into measurable business value.
5. Automating the Wrong Workflow: Usually One That's Too Complex
It's exciting to build an agent that does something impressive-looking end to end. It's much less exciting, and much more valuable, to use AI to knock out the low-level, repetitive, well-bounded tasks that quietly eat hours every week: document parsing, data entry, first-pass customer support triage, invoice processing, basic reporting.
AI is extremely good at bounded, high-volume, clearly-defined tasks. It's much less reliable the moment you ask it to handle open-ended, judgment-heavy, multi-step complexity end to end. The highest and most consistently measurable ROI I see, and this shows up in the broader research too, comes from exactly this kind of "boring" back-office automation, not from ambitious agentic builds that look impressive in a demo and fall apart in production. Know what the model is actually good at, and don't push it past that just because the more ambitious version makes for a better slide.
Avoiding the trap: starting small and building a clear ROI case from day one
Start with a clear definition of the business outcome your AI investment is supposed to drive. If you can't make the case in either productivity, customer satisfaction, cost, or throughput terms this investment will move a number that impacts the P&L, that's your signal to stop and rework the plan, not push forward and hope the ROI shows up later.
Once you've cleared that bar, get thoughtful about your data and your model selection, and optimize deliberately for cost as well as capability. That discipline is what earns you the credibility, and the budget, to scale AI into your next process, instead of spreading yourself thin across a dozen half-finished experiments and burning through tokens with nothing to show for it.
The companies winning the AI ROI conversation with their CFOs right now aren't the ones with the most ambitious AI roadmap. They're the ones who can say, in one sentence, exactly what they got for what they spent.
Erin Guthrie is the founder of Greenwood Point Strategy, where she advises founders and operators on growth strategy, margin improvement, and AI ROI, drawing on her background at McKinsey, Microsoft Cloud & AI, and as an operator who owned a $2B P&L at Uber.