AI tools are delivering real value for more and more companies, which has shifted the question many teams are asking from "Can we get AI to work?" to "How do we make this less expensive?".
73% of enterprises exceeded their AI cost projections this year, according to the FinOps Foundation's 2026 State of FinOps report. Uber blew through its entire AI coding budget by April and is now capping employee AI spend at $1,500 per month per tool. Accenture started rationing tokens to stop employees from using AI on basic tasks, not long after linking promotions to AI usage. One company reportedly discovered a $500 million Claude bill after forgetting to set usage limits.
Penny-pinching and tokenmaxxing are the same mistake with opposite signs. While Uber and Accenture were rationing tokens, Meta was tying AI usage to performance reviews, and its engineers burned through 60 trillion tokens in a single month. Both approaches optimize a usage metric instead of an outcome. The token number is never the target.
The instinct most organizations have is to restrict access: cap usage, limit seats, and require approval for AI tools. That works as damage control, but it also cuts into the productivity gains that justified the spend in the first place. There's another approach, which starts with understanding why the costs are growing the way they are.
The pattern that's burning money
Most teams use AI the same way every time: open a conversation, describe the task, get the result, close the conversation. The next time they have the same task, they do it all again. The AI starts from scratch each time, with no memory of yesterday's work, no awareness of the team's conventions, and no way to build on what it's already done.
Every one of those conversations is spending tokens on the same analysis, the same context, and the same patterns. A team running a daily report through an AI is paying the AI to figure out the data format, the metrics definitions, and the reporting conventions from scratch every single day. One CTO told Axios that employees at their company were using AI models to check the weather.
It's the AI equivalent of hiring a consultant to re-derive the same analysis every week rather than asking them to build a dashboard once.
The fix: build the tool, then stop paying for the tool
The most effective pattern we've seen both internally and across our client engagements is straightforward: identify if the task is consistent enough that it doesn't need an AI, and if that's the case build reusable software that accomplishes that task.
An example of this is a team using AI to reformat CRM exports every week. Each time, someone pastes the CSV into an AI conversation, explains what format the other system needs, and waits for the output. Each conversion costs tokens for the AI to parse the columns, understand the mapping, and produce the reformatted file. Multiply that by every export, every week, for every team.
The alternative: ask the AI once to write a script that does the reformatting. The script reads the CSV, applies the column mapping, and outputs the result. It costs zero tokens and produces consistent results. The AI's job was understanding the problem and writing the solution once rather than re-deriving the solution every time.
This works for a wide range of tasks:
- Data reformatting between systems. Instead of prompting AI every time a CSV needs transforming, ask it to write a script that does the transformation.
- Compliance and quality checks. Instead of AI reviewing the same checklist against every document, ask it to write a checker script that flags issues.
- Reporting and dashboards. Instead of generating reports through a conversation, ask AI to build the report generator and schedule it to run on a cadence.
- Inbox triage and classification. Instead of routing every message through a frontier model, use AI to build a classifier that runs on a smaller, less expensive model or a simple rule set.
The pattern holds whenever the task is repeatable and the criteria for "correct" can be clearly defined. Doing roughly the same thing each time means paying premium prices for commodity work.
The AI cost optimization hierarchy
The organizations getting this right treat AI models as a hierarchy, not a single tool. Using the smartest AI available to do things like simple document reformatting is overkill, and results in substantially higher spend for the same results. The price range makes this concrete: the least expensive production models cost around $0.10 per million tokens, while frontier reasoning models can cost up to $50 per million tokens for output. That's a 500x difference for tasks that might produce identical output.
At the time of writing, the top of this hierarchy is the frontier models (Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol, Gemini 3 Pro). Use these models for tasks that genuinely need reasoning, planning, and judgment. Architecture decisions, complex analysis, and novel problem-solving are expensive and worth it given the complexity.
In the middle: smaller models or one-shot API calls for tasks that need some intelligence but not the full frontier model. Summarization, code review, or draft generation all fall into this category.
At the bottom: deterministic code. Scripts, cron jobs, SQL queries, scheduled automations. These cost nothing to run and produce the same result every time.
The key insight: tasks should migrate down this hierarchy over time. Something starts as a frontier model conversation, gets understood well enough to be encoded as a script, and then runs as scheduled automation. The AI's job at each level is to make itself unnecessary for that specific task by producing something that runs at the level below.
The migration isn't always downward, though. As models improve, sometimes the right move is to reframe the task: a frontier model that owns a living script, watching the incoming data and revising the generator as the data changes, rather than processing every report itself. The frontier model's job shifts from doing the task to maintaining the code that does it.
If AI costs are growing linearly with usage, tasks aren't migrating down. That's coming from frontier token prices for work that should be running for free as code.
What this looks like in practice
I run a system that uses this hierarchy. Our operational overhead, covering daily reporting, data quality checks, monitoring, content management, inbox triage, and dozens of other recurring tasks all run on 65+ scheduled automations. Every one of those automations was built by an AI. None of them cost tokens to run.
When a new task comes in, the first question is: can a scheduled script handle this? If yes, we use AI to write the script and deploy it. If the task genuinely needs reasoning, it gets routed to the right model tier for the complexity involved. Mechanical work runs locally at zero cost. Judgment work runs on the frontier model. Nothing runs on the frontier model twice if a script could do it.
The result: our total AI spend for a team handling dozens of operational workflows hasn't risen with the added complexity because most of the work runs as code that an AI wrote once.
Three things to do this week
1. Audit repeatable AI tasks. Look at the last two weeks of AI usage across the team. Which conversations are people having more than twice? Those are the candidates. A daily standup summary, a weekly report, a recurring data pull. If the prompt is roughly the same each time, it should be a script.
2. Convert one task. Pick the most frequent repeatable task and ask the AI to write a deterministic version. "I run this prompt every Monday morning. Can you write a script that does the same thing automatically?" Most conversational AI tools can produce a working script in a single session, even for somebody without experience coding that can clearly describe what they need.
3. Measure the savings. Track what that one task was costing in tokens per week, then multiply by every repeat of the task that the team runs. The AI ROI case usually becomes obvious after one conversion.
Here's that measurement from one of our own processes. Every file that lands in our operations inbox gets screened for prompt-injection patterns before an agent reads it. Last week that was 408 files, roughly 450,000 tokens of content. Run through a frontier model, that screening would cost about $8 a week, or roughly $400 a year for one small task. Routed to a small model, still close to a dollar a week. As the deterministic script an AI wrote once, it costs nothing and finishes in milliseconds. Modest numbers for one task. We run more than 65 automations like it.
The economics get better, not worse
Inference costs for AI models have been dropping roughly 10x per year for equivalent capability. The blended cost of AI dropped 67% year-over-year, from $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026. Tasks that are expensive today will cost a fraction as much next year, but the deterministic versions built today will still cost nothing next year.
Organizations that build the right architecture now (routing work to the right tier, converting repeatable tasks to code, reserving frontier models for genuine judgment) see AI ROI improve as costs fall, without rebuilding. Organizations that keep routing everything through the frontier model will keep paying, just at a lower per-token rate on an ever-growing volume of tokens.
The economist William Stanley Jevons observed in 1865 that making coal more efficient didn't reduce coal consumption. It increased it, because cheaper energy made more applications viable. The same is happening with AI. Enterprise token usage grew over 1,000% from January 2025 to April 2026, while spend grew 497%. Tokens got cheaper, so people used more of them, and the total bill still grew. The only way off of that curve is to convert the repeatable work to code and reserve the tokens for work that actually needs them.
When token rationing isn't the answer
Restricting AI access is understandable when the bill arrives, but it trades a cost problem for a productivity problem. The people who were getting value from AI stop getting it, and the competitive advantage that was building disappears.
The better question isn't "how do we use less AI?" It's "how do we use AI to build things that don't need AI?" Every script, automation, and scheduled job is a task permanently removed from the token bill. The goal is to convert AI usage into AI-built infrastructure that runs on its own, for free.
