AI in business often starts out beautifully. The team tries ChatGPT, developers plug in Cursor, marketing generates copy, and operations roll out agents. Everyone is busy, and everyone seems to be moving faster. Then finance sees the bill and asks an uncomfortable but valid question: what exactly did we buy for this money?
That’s where things get interesting. AI costs don’t look much like a regular SaaS subscription. Some of the money goes to licenses, some to tokens, and some quietly burns up in agents that run through multiple steps in a row. If you don’t track that, the business quickly starts paying not for results, but for the pleasant feeling that “we’re adopting AI.”
Zapier doesn’t estimate AI costs by feel
According to the Zapier Blog, the company first gave employees open access to AI tools and measured almost nothing. Later, it changed course. Now every employee can see their own AI spend by tool and model, and the system can automatically add budget if the work genuinely requires it.
Usage data for Anthropic, Cursor, OpenAI, and Sidekick at Zapier flows into an analytics system every day. Each person gets a monthly limit for each tool. Not right at the edge, but above their historical usage. There’s a standard tier and a higher one for people who use AI more intensively.
When an employee reaches 75% of their AI limit, a Slack bot asks what they’re working on. They can request a budget increase or flag a suspicious expense for the security team. To get an increase, one sentence explaining the reason is enough.
From there, an AI step reads the response and adds the budget. A manager gets notified if a person has requested an increase two or more times in a month. If the higher limit is needed three months in a row, the user is moved to a higher tier.
This is not a story about tightening the screws. On the contrary, Zapier says it clearly: the goal is for teams to use more AI, but to do it consciously. Expensive models should be used where complex reasoning is actually needed. Simpler tasks are better handled by cheaper models or plain automation.
According to figures cited by the Zapier Blog, 52% of managers say their organizations spend more than $100,000 a month on AI, but only 34% of executives understand where that money goes.
Small businesses need visibility too, even without Zapier’s budget
For a Ukrainian agency, clinic, online store, or delivery service, the lesson is simple: don’t start with bans. Start with visibility.
Imagine an online store. A marketer uses AI to write product descriptions, support drafts replies, and an operations manager compiles reports. At first glance, everything looks fine: AI helps, and people save time. But look closer, and you may find that an expensive model is writing template emails that could be generated more cheaply. Or an agent keeps reworking the same file because the process is set up poorly.
A practical starting point could look like this:
- collect all AI tools the company pays for, including token-based usage;
- break down costs by team, person, model, and task type;
- set up a simple monthly report for each user;
- set soft limits by role, not the same cap for everyone;
- review tasks where an expensive model is being used for routine work.
The catch is obvious: control can easily turn into bureaucracy. If people have to file a request for every AI prompt, they’ll simply stop experimenting. It’s better to follow Zapier’s approach: keep limits high enough for normal work, and only trigger review when spending gets close to the threshold.
There’s one more thing businesses often forget. You don’t need an LLM everywhere. If the process is predictable — say, you need to pass a lead to a CRM, send a Slack notification, or update an order status — plain automation through n8n, Make, or Zapier will be cheaper and more reliable. AI is better reserved for places where you need to interpret text, classify requests, draw conclusions, or prepare decision options.
Our position at Artica is simple: AI budget should live alongside business metrics. If spending doesn’t affect revenue, speed, productivity, waste reduction, or risk, it should be reviewed. No need to cut immediately. But you definitely need to know what you’re paying for.