Why Companies Are Cutting AI Spending (and How to Measure What You Actually Get)

For about two years, the corporate attitude toward AI was roughly "try everything, worry about the bill later." Early in 2026 that attitude even got a nickname: tokenmaxxing, the habit of encouraging people to use AI for as much as possible, as often as possible.

Now the mood is shifting. As of early November 2026, reports from major business and tech outlets describe companies tightening budgets, capping usage and asking a question that used to be unfashionable: what are we actually getting for this?

This article looks at why that's happening, where AI money really goes, and how to think about return on investment without fooling yourself.

What's being reported

A few details give a sense of the picture. Reporting this year described Uber as having used up its entire 2026 AI budget within a few months, with its leadership saying the spending had become harder to justify. Other coverage says Meta built an internal dashboard to monitor token usage and set budgets after projecting very large AI costs. Several large firms are reported to have adjusted internal policies to limit usage, and some have told employees to use cheaper or open-source models for simple work.

Two caveats are worth keeping in mind. First, many of these accounts come from secondhand reporting, and exact figures vary between outlets, so treat specific numbers cautiously. Second, "cutting AI spending" doesn't mean AI is going away. Industry forecasts still expect enormous total AI investment over the next few years, much of it in infrastructure. Gartner, for example, projected global AI spending above two trillion dollars for 2026. What's changing is how individual companies manage the usage side: more tracking, more limits, more questions.

So the better description is not a retreat but a correction. The free-for-all phase is ending, and cost discipline is arriving.

Why the bills got so big

If you've read our guide to AI tokens, you already know the basic mechanics: you pay for what the model reads and writes. A few things about how companies use AI push that number up faster than people expect.

Usage was encouraged without limits. When employees are told to use AI everywhere and nobody sees the cost, consumption climbs. Individual power users can run up large monthly token bills, especially with coding assistants and agents.

Agents multiply calls. A single request to an AI agent can trigger dozens of model calls, each carrying the entire conversation history. One task that looked cheap on a pricing page ends up consuming a surprising amount.

Frontier models for grunt work. Many teams default to the most capable and most expensive model for everything, including simple tasks like classification, formatting and summarization that a much cheaper model handles well.

Reasoning tokens. Models that "think" before answering generate extra hidden output that is billed. Leaving reasoning settings high for easy tasks is pure waste.

Pilots that never ended. Experiments launched with enthusiasm can keep running, and billing, long after anyone checks whether they delivered value.

Prices drop, but volume rises faster. Cost per token has generally fallen over time, yet total spending can still grow because usage grows even faster. Cheaper often means more of it, not a smaller bill.

What companies are doing about it

The responses fall into a handful of practical patterns.

Visibility first. You can't manage what you can't see. Companies are building dashboards that show tokens and dollars by team, product feature and sometimes by individual. Tagging requests by purpose is the foundation of everything else.

Budgets and caps. Teams get monthly allowances. Individual limits on expensive tools prevent a few heavy users from consuming a large share of the total. Some firms require approval for higher-cost usage.

Model routing. Instead of sending every request to the top model, a router sends easy tasks to a small, cheap model and reserves the expensive one for hard problems. Some companies are moving simple work to open-source models they run themselves, which changes the cost structure from per-token fees to infrastructure costs.

Prompt and context hygiene. Shorter system prompts, trimmed conversation history, retrieval instead of pasting whole documents, and caching of repeated content all reduce input tokens without changing the result.

Batching and scheduling. Work that isn't urgent can often be run in batches at a discount.

Killing projects. Perhaps the least glamorous step is also one of the most effective: reviewing pilots and shutting down the ones with no measurable benefit.

The hard part: measuring return

Cost is easy to measure. Value is not, and this is where a lot of organizations stumble.

The tempting shortcut is to treat usage as success. "Our teams used four billion tokens this month" sounds impressive, but it says nothing about outcomes. A team burning huge volumes on poorly designed prompts may get less done than one using a fraction of that. High consumption can even signal inefficiency.

Better measures tie AI to something the business already cares about:

• Cost per completed task. What does it cost, in tokens translated to money, to resolve a support ticket, review a contract or produce a first draft?
• Time saved, honestly measured. Compare the actual time for a task with and without AI, including the time spent checking and fixing AI output. Self-reported savings tend to run optimistic.
• Quality and error rates. If AI speeds up work but increases mistakes that someone has to fix later, the net benefit may be small or negative.
• Throughput or revenue effects. Did the team handle more volume, close deals faster or reduce churn?
• Adoption that sticks. Are people still using the tool after the novelty fades, or did usage spike and die?

A good rule is to measure at the level of a specific workflow, not the company as a whole. "AI improves productivity" is a slogan. "AI cut our average claim processing time from 12 minutes to 8 while keeping error rates flat" is a result.

A simple framework for deciding what to cut and what to keep

If you're responsible for an AI budget, here's a straightforward way to sort things.

Keep and expand use cases where the cost per task is low, the value is clear and the quality holds up. Customer support triage, document extraction and code review assistance often fall here, though it depends on the company.

Optimize use cases that work but cost too much. This is where model routing, shorter prompts and caching pay off. You keep the benefit and trim the waste.

Pause or redesign use cases where the benefit is vague, the output needs heavy human correction or the architecture is wasteful. Agents that loop without limits are a classic example.

Stop projects nobody can connect to an outcome after a fair trial.

Mistakes to avoid when cutting

Cost-cutting can backfire if it's done bluntly.

Cutting everything equally punishes the good use cases along with the bad ones. Target the waste, not the whole category.

Moving to a cheaper model without testing can quietly lower quality. A cheaper model that needs three attempts, or produces errors people must fix, may cost more overall.

Ignoring the hidden costs of people and process is common too. Training, oversight, security review and integration work all belong in the total, not just the API invoice.

And don't let caps destroy learning. If employees are afraid to experiment at all, you may save money now and fall behind later. The goal is controlled experimentation, not none.

What this means for individuals and small teams

You don't need to be a large company to apply this. If you use AI tools heavily, check what you're paying and for what. Match the model to the task. Keep conversations short instead of dragging a huge history along. Turn off extended reasoning when you don't need it. A little attention goes a long way.

The takeaway

The current pullback isn't proof that AI doesn't work. It's what happens when a technology moves from exciting experiment to ordinary line item. Finance teams start asking for numbers, and the strongest answers come from companies that can point to specific workflows, specific costs and specific results.

The organizations that handle this period best will probably be the ones that stopped counting how much AI they use and started counting what it produces.