AI token costs
AI inference

How much are companies spending on AI inference?

hero image for blog post

By Michael Domanic, Section AI

Most finance leaders can tell you what their company spends on AI. What they can’t tell you is whether that number is normal, because there are almost no benchmarks to compare against.

That matters, because the bills are real now - into the millions of dollars in inference spend for large companies. 

So we looked for a benchmark ourselves. We analyzed real billing data from companies using Section HQ, matched to how skillfully each person uses AI. The headline: the average company spends about $150 a month per AI user. Let’s break down what drives AI spend, who amasses it, and what it says (and doesn’t) about your company. 

Spend per user: About $150 a month

Across all companies, spend per month (over the month of October) worked out to $150 per person who used a paid AI model. The median company spent $125 per user.

The per-employee figure is lower because at most companies, only a share of staff generate any metered spend. The rest either aren't using AI yet, didn't use a paid model that month, or are on plans that aren't billed by usage.

Because we work with a range of companies, from those just starting out to the quite advanced, this provides a decent proxy for budgeting. $100 to $150 per active user per month is a reasonable middle, but plan for some users to cost 10 to 100 times that.

The more skilled the user, the more they spend

Inside a company, per-user spend climbs steeply with AI proficiency. A typical Expert (L4) spends about $167 a month, roughly 60 times a typical Novice (L1).

This makes sense: In our Section methodology, an Expert is defined as a user who builds advanced workflows for themselves and others. They’re more likely to incur higher inference costs because they’re building agents, chaining them together and overall using more advanced capabilities.

Section HQ spend and AI Fitness data · October 2026 · person-level data from a subset of companies

That makes skilled users the bulk of the bill. In every company where we looked person by person, people at Practitioner level or above drove 91% to 99% of spend. Novices and Experimenters barely register.

Heavy users and tooling choices matter more than overall company fitness

You might expect companies with more skilled AI users to spend more per user. Partly, they do. But in our data, two other factors explain far more of the difference between companies:

Which provider you buy from. Companies on Anthropic alone spend a median of about $150 per user a month; companies on OpenAI alone, about $40. Part of that is skill: the Anthropic-only companies have more Practitioners and Experts. But a bigger factor is likely what people use AI for. At Anthropic-only companies, about 70% of spend (median) goes to premium models, and where we could break spend down by team, technical teams usually account for more than half of it. That points to heavy coding and agentic work, which uses far more compute per task than everyday chat. Provider choice may also be a signal in itself: in our experience, companies that pick Claude are often making a more deliberate bet on AI. Differences in pricing and rate limits between providers may play a part too.

A few power users. Spend is concentrated at the top at companies of every size. Across the companies we looked at, the top 5% of users typically drove 35% to 50% of the month’s spend. At one large company with low average fitness, a few highly skilled, high-usage people produced about 64% of the month’s spend, and the top 10% of users produced 83%. That company had the highest per-user spend we saw, despite one of the lowest overall fitness scores.

Overall fitness still matters. Person by person, more skilled users spend far more, and they get far more value from what they spend. But at the company level, spend is set mostly by a small group of heavy users and by your tooling choices. That's also why spend alone tells you little: two companies can have the same bill and get very different value from it.

Where the spend goes: technical teams spend twice the average

Engineering and IT teams spend about twice their company's average per person. Back-office teams spend a third or less. In dollar terms, that's roughly $290 per person a month for technical teams versus about $35 for finance, legal and HR (medians). The pattern held across company sizes and providers. 

This makes intuitive sense - coding work generally uses more compute, so technical teams spend more. And the most regulated teams - finance, HR, and legal - may have stronger data privacy restrictions, making it hard to enable connectors or agentic capabilities. 

Low-spend teams aren’t necessarily non-users - they’re just not using lots of agentic AI. At one company, the legal team had more AI users than any other department but generated only 3% of spend, about $56 per person. The technology team at the same company spent about $1,000 per person.

Section HQ spend and AI Fitness data · October 2026 · teams grouped into functions by name

What this means if you're setting an AI budget

  1. Budget $100 to $150 per active user per month as a minimum - but don’t stop there. It’s what the average company spends today, and most companies are still early in adoption. Treat it as a floor, not a target: budget $300 or more for technical teams and your most advanced users to start, and plan to be at $300-$500 for non-technical users once your adoption has matured. If your spend is well below these numbers, it means people aren’t yet using AI for heavier work like agents, which means you’re unlikely to get business value.

  2. Expect spend to grow as people get better. Practitioners typically spend about $70 a month, versus a few dollars for Novices. That is the cost of AI becoming part of real work, so plan for it (and celebrate the success) rather than treat it as overspend.

  3. Watch your top users, not the average. A handful of power users can drive most of the bill. Know who they are, what they are building, and whether the value matches the spend.

  4. Don't use spend as a proxy for cross-workforce skill. A high organizational bill can come from a few heavy users in an otherwise early-stage workforce. Measure proficiency and spend directly.

Have a great week,

Michael

Your fellow Head of AI

A note on how we measured

Spend is inference billed through Anthropic and OpenAI admin data for October 2026, the latest complete month, as reported in Section HQ. "Per user" divides spend by the people who generated any spend that month; "per employee" divides by total headcount. AI Fitness is Section's measure of how skillfully each person uses AI, from L1 Novice to L4 Expert. Person-level comparisons use a subset of companies where we could match each person's spend to their fitness level. All companies are anonymized.

‍

Greg Shove
Michael Domanic
Open laptop on a blue fabric surface displaying a user dashboard with welcome message and options for a 60-day plan, team fluency, and section insights.

Section HQ

For any

(and every)

leader responsible for AI success

From org-wide impact to department-level use cases, Section HQ gives every leader the insights they need to drive progress.