there's an instagram post where someone gave two AI agents $1000 each to see how they'd spend it. it's a fun experiment. it's also a preview of the policy problem every company will hit.
the agents had a budget. they didn't have:
- category restrictions (can it buy software? hardware? subscriptions?)
- velocity limits (can it spend $1000 in one transaction or does it need to spread it out?)
- an audit trail that explains why it chose each purchase
when you give an agent a corporate card, you need all three.
most teams start with a spending cap and add rules after the first bad purchase. that's backward. the policy layer should exist before the agent sees the payment API.
MnemoPay enforces budget, category, and velocity rules before the transaction goes through. agent fico tracks spending behavior over time so you can tighten or loosen limits based on the agent's track record.
i built both because the $1000 experiment is cute. the $100,000 experiment is a compliance incident.
if you're testing agent spending in production, start with the guardrails. the agent will find creative ways to spend money. your job is to define what 'creative' means before it does.
Top comments (0)