Levee / AI agent card controls
Why card limits and MCC blocks aren't enough for AI agents
Static card controls answer "how much" and "what kind of merchant". AI agents need an answer to "is this the purchase a human approved?"
Short answer: Limits and MCC blocks can't express intent. An agent can stay under every cap and inside every allowed category while still buying the wrong thing. Intent-based spend authorization checks each payment against a signed, human-approved mandate at the moment of authorization, and records why it was approved or declined.
The controls most platforms use today
| Control | What it does | Gap for AI agents |
|---|---|---|
| Spending limits | Caps how much can be spent | Can't tell a correct purchase from a wrong one under the cap |
| MCC blocks and allowlists | Restricts merchant categories | Too coarse: a whole category is allowed or blocked, regardless of purpose |
| Single-use or virtual cards | Limits exposure per card | Still says nothing about what the agent intended to buy |
| Prompt or tool guardrails | Shapes agent behaviour | Lives inside the agent, so it can be bypassed or injected |
| Fraud scoring | Flags stolen-card patterns | Built to catch criminals, not an agent acting outside its approval |
| Intent-based spend authorization | Checks the purchase against a human-approved mandate at authorization | Adds purpose, approval and proof; works alongside the controls above |
Where static controls break
They can't say "for this purpose, this week, from these merchants"
A human approving an agent's purchase usually means something specific: these cloud credits, for this project, from this vendor, before Friday. Limits and categories flatten that into a number and a code.
They're all-or-nothing
Give an agent a card and it can spend anywhere the controls allow, for any reason. There's no middle ground of "ask me first if it's unusual".
They leave no proof
When a finance team, auditor or card network asks why a purchase was allowed, a spending limit offers no answer. Agent spend needs a reasoned, tamper-evident record for every decision.
What to add
- Human-approved, signed mandates that state purpose, merchants, amount and time window.
- A decision point inside the issuer's authorization path, outside the agent, that fails closed.
- A challenge path so edge cases go back to a human instead of silently passing or failing.
- A tamper-evident evidence ledger for chargebacks, finance review and audits.
That is what Levee provides. It works alongside existing limits and fraud tools rather than replacing them.
Related
Levee is running a free design partner pilot for platforms putting AI agents in charge of spend.
Become a design partner