How to survive the AI spend hangover

It's 6:30am and you hear the door of the nightclub you've spent the last 8 hours inside shriek as it closes behind you. You watch bleary-eyed as an overly bright sunrise illuminates the business-suited people as they glide effortlessly along the sidewalk, their obnoxiously well-rested faces talking about work on their fully charged phones. You wonder, "Where did all the fun people go? And what happened to my wallet?"

This feeling is what many CFOs, CTOs,  CEOs, and AI program managers will imminently be experiencing in their board rooms, as they finally wake up to the realities that unrestricted and unmoderated AI use has wrought on their bottom lines and the stability of their core technical assets. 

You can already feel the party ending and the hangover setting in. The first warning sign came when Uber’s engineering org burned through its annual AI budget by April, and then capped its engineers at $1,500 a month per tool. At Meta, an internal leaderboard nicknamed "Claudeonomics" turned token spend into a status game. The company was on pace to spend billions, and the CTO's eventual memo had to spell out that token usage on its own measures nothing. Two of the most sophisticated engineering organizations on the planet have arrived a half step ahead of where we will all be soon: facing down a shocking bill and scrambling to tie it to any real ROI. 

Worse, many organizations would be hard pressed even to say which teams spent their tokens, on which models, and on what projects. Tokens spent wisely on complex problems, and tokens burned writing personalized fanfic all look the same on an invoice. But untangling them just became an urgent priority for everyone who shares responsibility for their company’s AI bill.

Like any hangover, this one is going to hurt. But we don’t have to wait for the club to close down to start sobering up. There are already lessons to be learned about the differences between the companies using AI responsibly and the ones that have just been partying like there’s no tomorrow.

Budgeting for a world where LLMs stay and AGI never arrives

Paradoxically, the better LLMs get, the easier it is to see how far away they are from replacing the average knowledge worker. Even at their most effective task, coding, AI agents regularly fail to produce net positive results without a knowledgeable human to begin the work and the critical eye of an expert to review the outputs.

For the first few years of the AI boom, the assumption was that we would eventually close this capability gap. But the reality is that in order to train and scale a model like Mythos, we've already stretched the pricing elasticity of chips, servers, and memory to their limits. There simply isn’t enough time and funding left to close the ocean-sized gulf between the capabilities of models and those of properly trained humans.

Regardless of whether or not AI reaches workforce-replacement capability, the next problem is whether companies can afford to use it. Frontier model tokens are unlike any compute cost we've seen before. If your AWS bill gets too high, you can build your own datacenter and convert that expense into a capital investment with a predictable depreciation schedule. You can't do the same with frontier model inference. Once you've built a product that depends on frontier models, those marginal costs are a permanent feature of your economics.

And those costs only grow over time. Per-token prices for GPT-4-level performance have reportedly fallen roughly 98% since late 2022, yet enterprise AI bills over the same period more than tripled, thanks to our voracious appetite for state-of-the art level inference.

These may seem like macroeconomic abstractions, but they have direct implications for how you as a Finance, IT, or AI program leader (and hopefully at your organization, all three of those people are making decisions in concert) design your AI budgets and tie them to business outcomes. Specifically, it means that your organization cannot afford to default to using frontier models for every task, based on the assumption that they will soon be able to operate without human supervision.

It’s also not as simple as issuing a blanket restriction on frontier models. As Matei Zaharia, CTO of Databricks, explained, "Cheaper per-token does not imply cheaper per-task…For example, Sonnet 5 costs less per token than Opus 4.8 but used more tokens, resulting in higher cost and lower quality."

Tooling over tokens

Even if we all start aggressively monitoring and reining in AI spend, that doesn’t instantly revert us back to the good ‘ol days where artisan engineers painstakingly chiseled out software from silicic igneous rocks with their bare hands. These models are here and our reliance on them is permanent. So the smartest organizations are now all asking the same question: “How do we get value from them without lighting money on fire?” 

Here’s my advice to the leaders designing and approving AI programs: be judicious with inference, be generous with tool calling, and invest in the best harnesses possible that ensure those outcomes.

As we’ve learned in 2026, naive agents doing exponentially more work will burn exponentially more inference tokens doing it, so your capability curve and your cost curve are the same line. The fix is to stop using agents like a tourist and start using them like a resource-constrained engineer. We solved this forty years ago and called it platform engineering: every abstraction exists so the person above it does less. An agent's job, in a sane system, is to make the next agent need less inference. Build that scaffolding and the exponential curve bends logarithmic.

An executive recently shared an anecdote with me that showed the above problem in practice. The company's biggest token consumer (who had no idea they held the crown) was torching money every two days feeding the entire product’s compile log through a model just to see which errors were trending. What was likely tens of thousands of dollars in inference could actually be accomplished by leveraging that inference to build software that scans and parses the log for fifty cents of CPU a day. The moral of the story is that the easiest way to trim excess AI spend is not to ask the model to calculate the first ten million primes; ask it to write the code that does.

Software and people on the critical path

The last few years have taught us that AI is great at writing code and exceedingly mediocre at judgment. An engineering leader at Microsoft put the paradox to me this way: a year ago he told his org that if anyone was still hand-writing unit tests by the end of 2025, they'd failed, because the models are better at producing them than we are. And yet, point an agent at a repository, ask for "great test coverage," and what comes back is garbage.

In fact, it’s actually worse than garbage because garbage is generally easy to identify on sight. In this case, the model doesn't understand your codebase, so it tests what's easy rather than what matters, then fluffs up its suboptimal outputs with signals of quality, excessive comments, confident summaries, and impressive language. This is not an efficient use of your AI budget.

The version that works is one we are already familiar with and barely looks like it’s powered by AI. It’s using AI to create deterministic machinery that measures the coverage and targets the public interfaces that matter, then lets the model fill the real gaps before handing control back to the tools. And in the end, a real, bona fide human still owns the sign-off. 

Keep a person and a deterministic check on the path, or you'll pay premium prices for judgment that is flawed and never learns from its mistakes.

Measure idea-to-customer, not tokens

This brings us back to the thing every one of those budget blowups had in common: companies finding out too late that AI spend has a negligible relationship to AI payoff. It’s important to have visibility into your token spend so you can monitor usage and budget.. But tokens are close to meaningless as a measure of value, which is why a leaderboard ranking your engineers by consumption mostly teaches them to consume.

The metric that actually matters is calendar time: how long it takes to go from an idea, a PRD, a concept, to value in a customer's hands.

Call it idea-to-customer. It's brutally honest, because it starts from a baseline that can be measured independently of AI and doesn’t frame the problem as something only agentic AI can solve. If your AI investment isn't bending that number down, it isn't working, whatever the usage dashboard says.

Underneath it, three things are worth watching: speed, ease, and quality. 

• Speed is the idea-to-customer clock itself. 

• Ease is how much of an engineer's week goes to creating value instead of keeping the lights on and fighting their own tools. 

• Quality is whether what you ship survives users: How often defects escape, how fast you recover, whether anyone actually loves the result.

Closing time

So, back to the hangover. The companies that stagger out with empty wallets will be the ones who used AI like an open bar. The winners will be the ones that used it like an engineer, kept a human on the critical path, and measured the one thing that counts: how fast an idea now reaches a customer. 

The path out of the AI budget mess requires a nuanced approach that extracts maximum efficiency from spend. And the first step along that path is visibility into how tokens are being spent today, down to the level of each team, user, and model. That’s one of the challenges 1Password is leading the way on today. Consider it that crucial first cup of coffee the morning after a night out.

1Password can now give companies a holistic view of AI costs and usage, so they can set budgets, track burn rates, and get alerted before prepaid balances run out.