Playbook
Writing spending policies for agents
A practical template for your first agent policy — limits, allowlists and the approval rules that keep finance happy without slowing anyone down.
Most teams write their first agent policy in one of two ways: so loose that it doesn't matter, or so tight that every payment needs a person. Here's the structure we recommend instead.
Start with the job, not the agent
Describe what the agent is for in one sentence — "books economy travel for employees" or "renews approved cloud and SaaS contracts". Every rule should follow from that sentence. If a payment doesn't fit the job, it shouldn't go through automatically.
Three layers of limits
- Per payment — the largest single amount the agent can spend without asking.
- Per period — a daily or monthly ceiling, so many small payments can't add up to a surprise.
- Per merchant — a cap for any one vendor, which catches duplicate or runaway renewals.
Allowlist first, then loosen
New agents should only pay merchants you've explicitly approved. After a few weeks, look at what was held and decide which vendors to add. It's much easier to loosen a policy with data than to tighten one after an incident.
Approval rules people will actually use
Route approvals to someone who knows the context, with a backup. Keep the threshold high enough that approvers see a handful of requests a week, not a hundred — otherwise they start approving without reading.
Here's what a first policy looks like in Halden:
agent: ap-travel-03
purpose: Book economy travel for employees
limits:
per_payment: 1500 USD
per_day: 5000 USD
merchants: allowlist # airlines, rail, approved hotel chains
block_categories: [gift_cards, crypto, cash_equivalents]
approval:
over: 1000 USD
approvers: [people-ops-leads]
backup: finance-on-call
Review monthly
Put a 30-minute review in the calendar. Look at holds, declines and anything approved over the threshold. Policies are living documents, and the ledger tells you exactly where yours is wrong.