AI budget as % of payroll (4% rule)

Set a single rule: the company centrally pays for all AI tools (chat seats, coding agents, API credits, transcription, dictation, design tools) up to a target percent of total payroll. No reimbursement workflow. No per-employee approval. Power users spend $300–600/mo, casual users $20–50/mo; the distribution is supposed to be uneven. Track monthly spend as a single line item.

A small AI media-buying startup runs at ~4% of payroll on AI subscriptions and tokens. A large fintech budgets several hundred thousand dollars/year for IT AI (against several million in targeted effect — better than 2× payback). The argument is structural: at <5% of payroll, the friction of asking permission costs more in lost experimentation than the spend itself. The win comes from people experimenting freely; budget caps starve the exact employees who deserve more headroom.

1. Centralize billing on every AI tool — chat, coding agents, API, transcription, dictation, design.
2. Default approval policy: anyone can request any standard tool, ops auto-provisions.
3. Budget the line as a percent of payroll. Start at 3%, target 5–8% as agentic workflows go autonomous.
4. Track monthly per-tool spend in a single dashboard; review quarterly.
5. Do not cap individuals. Surface the top-10 spenders to leadership; they're typically the highest-leverage employees, and the data is feedback, not a threat.
6. Allow one founder-experiment line for one-off bursts (the founder who burned $2K in a week testing fast modes flagged it as one-off, not in the 4%).

• Employees paying out-of-pocket and "expensing later" is the wrong model — friction kills adoption silently. Make it a tooling-procurement line, not a perks line.