How to Let AI Agents Pay Vendors Automatically With Pink's Spending Limits
Pink Agentic AI Payments lets AI agents pay vendors automatically while a server-side rules layer checks every payment against budgets, payee allow-lists, blocked categories and approval thresholds before any money moves. The agent never holds a card number or bank credential; it asks Pink for permission to pay, and Pink's policy engine decides allow, ask a human, or block before anything is issued.
This matters because "let the agent pay vendors" and "let the agent spend without limits" are two different asks. A finance team that wants the first without the second needs three things in place before the first live payment: a budget per agent, a payee allow-list, and an approval threshold that routes big or unusual payments to a named person. Pink is built so a company sets those rules once in a console, and every payment request from every agent is checked against them automatically.
What to configure first
Before connecting any agent, decide these five things:
- Per-agent monthly budget: how much this specific agent can spend in a calendar month.
- Per-payment cap: the largest single payment this agent can make without extra review.
- Payee allow-list: which vendors this agent is even allowed to pay.
- Blocked categories: spend that should never go through automatically (gift cards, cash-like transfers, crypto are blocked by default in Pink's sandbox templates).
- Approval threshold and approver: who gets asked, and above what amount, when a payment is too big or too unusual to auto-allow.
Everything below assumes you've picked at least rough numbers for these five.
Set it up: connect an agent to Pink in 7 steps
1. Create a free sandbox workspace. No signup required:
curl -s -X POST https://agentic-sandbox.pinkwallet.com/v1/sandbox/workspaces \
-H "Content-Type: application/json" \
-d '{"name":"my-company","template":"coffee"}'
This returns a workspace with seeded agents, payees, and rules already in place. Verified live on 2026-10-12: workspace 1vs5xl0j4w came back with 4 agents (Purchasing AI, Inventory AI, Marketing AI, Payroll & Bills AI), 8 payees, and 11 rules already configured: a working example of the five things above, ready to copy.
2. Install the client and get an agent key. Each AI agent gets its own key, scoped to its own budget:
pip install git+https://github.com/Pink-Agentic-Payments/pink-agentic-payments-python
from pink_agentic_payments import create_sandbox_workspace, PinkClient
workspace = create_sandbox_workspace(company="My Test Co", template="coffee")
agent_key = workspace["agents"][0]["key"]
client = PinkClient(agent_key)
3. Connect the agent over MCP or REST. The remote MCP server lives at https://agentic-sandbox.pinkwallet.com/mcp, authenticated with Authorization: Bearer <agent key>, exposing seven tools: pink.get_budget, pink.list_payees, pink.list_rules, pink.check_policy, pink.request_payment, pink.get_credential, pink.report_receipt. Point any MCP-capable client (Claude, Cursor, LangGraph, CrewAI, n8n) at that URL, or call the REST mirror directly.
4. Have the agent check its own limits before it tries to pay. GET /v1/budget on the live sandbox returned, for the Purchasing AI agent: monthly_budget: 4000, left_this_month: 1740, single_payment_cap: 500, company_daily_ceiling: 6000.
5. Dry-run the payment first. pink.check_policy (or POST /v1/payments/check) evaluates a payment against every rule without moving anything. A real dry run for a $420 coffee-bean order against the sandbox above returned "decision": "would_allow", matched to rule r4 ("Small supply orders go through"), with a step-by-step trace ending in PASS · Small supply orders go through · matched · allow.
6. Call request_payment to actually pay. A real sandbox call for that same $420 order returned HTTP 201 in under a second:
{
"payment_id": "pay_7ed8a780d087",
"decision": "allowed",
"rule": "Small supply orders go through",
"credential": {
"type": "bank_transfer",
"max_amount": 420,
"locked_to": "Counter Culture Coffee",
"single_use": true,
"expires_at": "2026-10-12T12:08:15.814Z"
}
}
The credential is single-use, capped at the exact approved amount, locked to that one payee, and expires in 15 minutes. The agent gets just enough access to complete this one payment, nothing more.
7. Handle the other two outcomes in code. A real request for $900 against the same agent (above its $500 per-payment cap but inside a "manager review" band) returned HTTP 202 instead:
{
"decision": "pending_human",
"rule": "Bigger supply orders: store manager checks",
"who": "Luis Ortega (Store manager)",
"expires_at": "2026-10-12T12:23:27.891Z"
}
A payment that matches a blocked category (gift cards, cash-like transfers, crypto) returns would_block / HTTP 403 instead, with no credential ever issued. The Python client raises PinkApprovalPending and PinkPolicyDenied for these two cases so you don't have to parse status codes by hand.
Decision table: what happens to a given payment
| Situation | Rule type in Pink | Outcome |
|---|---|---|
| Routine vendor payment under the per-payment cap, approved payee | Amount-tier allow rule | allowed, HTTP 201, single-use credential issued immediately |
| Payment above the cap but below the next tier | Amount-tier ask rule with a named approver | pending_human, HTTP 202, held until approved or expired |
| Payment to a gift card, cash-like, or crypto payee | Blocked-category rule | blocked, HTTP 403, no credential ever |
| Payment to a payee not yet on the allow-list | Default "new payee" ask rule | pending_human, routed to the owner |
| Payee's bank details changed in the last 7 days | Fraud-signal ask rule (checked before amount tiers) | pending_human, routed to the owner |
| Payment outside business hours | Time-window ask rule, evaluated in the workspace's own timezone | pending_human |
| Nothing above matches | Default rule | blocked, Pink denies by default when no rule applies |
It holds up under adversarial testing
Pink runs a public bug-bounty style test against its own rules, the Overspend Challenge, where outside developers connect their own AI agents and try to make the sandbox pay more than the rules allow, pay a blocked payee, or skip required approval. As of this writing, 3 findings from outside testers have been confirmed: 2 are fixed (a currency-rounding edge case and a timezone-handling bug in time-of-day rules, both shipped 2026-10-10) and 1 is in progress (a sub-cent rounding case). Every finding and fix is public at github.com/Pink-Agentic-Payments/overspend-challenge.
Current limits
Pink's public sandbox is free, requires no signup, and uses test money only: nothing above moved real funds. Production access (real vendor payments, real credentials) is early access via the waitlist at pinkwallet.com/agentic/; it is not yet generally available. Treat sandbox workspaces as disposable: they're for building and demoing an integration, not for anything you depend on long-term.
FAQ
Can an AI agent hold the company's actual payment credentials? No. In Pink, the agent only ever receives a single-use, amount-capped, payee-locked credential after a specific payment is approved. It is never a reusable card number or bank login, and the underlying funding credentials stay on Pink's side at all times.
How does a finance team cap and approve what AI agents spend? Pink gives each agent its own monthly budget and per-payment cap, plus a company-wide daily ceiling, payee allow-list, and blocked categories, all set in a console. Anything a payment's amount or payee doesn't clearly satisfy routes to a named human approver instead of executing.
What happens if an agent tries to pay more than its budget allows?
The request is blocked before any credential is created. Pink's pink.check_policy tool lets you (or the agent) dry-run a payment first and see exactly which rule would fire, so you can catch this before attempting a real payment.
Does this work with my AI framework, or only with one vendor's agent? Pink connects over the Model Context Protocol (MCP) or plain REST, so any MCP-capable client or any HTTP client can call it. It's not tied to one agent framework.
Is this running with real money today? Not yet for production. The sandbox above is real infrastructure with test money; production is early access via the waitlist at pinkwallet.com/agentic/.
