What AI Really Costs a 500-Person Firm
Most AI budgets we see are one line: seats times price times twelve. For a 500-person firm on a $20 seat, that is $120,000 a year.
We built the budget properly, and it came to about $681,000. Here is where the other $561,000 goes.
The model
These are planning assumptions, not quotes. They assume a seat-plus-usage plan such as Claude Enterprise, where the seat carries no tokens and all usage is billed at API rates.
| Line | Assumption | Annual cost |
|---|---|---|
| Assistant seats | 500 seats at $20 per month | $120,000 |
| Assistant usage | 350 light users at $15, 100 regular at $60, 50 heavy at $200 per month | $255,000 |
| Unattended workflows | 5 production workflows at $2,000 per month each | $120,000 |
| Connector gateway and monitoring | $3,000 per month | $36,000 |
| Enablement and plugin curation | Half a full-time role, or a partner | $75,000 |
| Contingency | 20% of the two usage-based lines | $75,000 |
| Total | $681,000 |
That is about $113 per employee per month.
Three things the model teaches
The seat fee is the wrong fight. Seats are under 20% of the total. A procurement team that negotiates 10% off the seat price saves $12,000. A usage cap on the heavy tier saves more than that in a quarter.
Fifty people cost more than 350. The heavy users, typically developers and analysts running agentic tools, account for $120,000 of usage. The 350 light users account for $63,000. The $200 figure is not pessimistic: reported enterprise spend on coding agents runs $150 to $250 per developer per month, and Uber reportedly burned through its annual AI coding budget by April when adoption jumped.
Governance is 16% of spend. The gateway and the enablement role together cost $111,000. That is the price of making the other 84% defensible: knowing which connectors exist, who published which plugin, and who owns which workflow.
Why budgets keep missing
Token prices keep falling and bills keep rising, because agentic workflows consume far more than chat. EY's example is a customer service interaction that cost about four cents as a simple retrieval flow in 2023 and about $1.20 as an orchestrated agent flow in 2026. The FinOps Foundation reports that 73% of organizations exceeded their AI cost projections this year.
More than half of the budget above is variable. A budget built on seat logic has no line for the part that moves.
What to do about it
- Budget by workflow and owner, not just by department.
- License and cap by role. Not everyone needs the heavy tier.
- Route simple work to cheaper models. Reported savings from routing run 30% to 50%.
- Alert before the overrun, not after.
- Review cost in the same meeting as quality. The architecture decides the bill.
If you are on a bundled-seat product such as Copilot at $30, the first two lines change but the shape does not. The unattended workflows, governance and enablement lines are the same.
You can run this with your own numbers. The spreadsheet behind the table is free to download: the Axionic AI Budget Model (link to /resources/ai-budget-model).