Best Payment Setup for Autonomous AI Procurement (2026): Providers Compared for Companies
Last updated 2026-09-30.
Short answer: it depends on what the agent is buying, not which vendor is "best." For SaaS and other card-accepting vendors, the most-documented options are AgentCard, Stripe Issuing's agent product (private preview), Lithic, and Privacy.com — each names a per-card or per-agent limit field. For metered API/cloud usage, Coinbase's CDP Agentic Wallet and Circle's Agent Wallets document per-call and per-period caps on a stablecoin rail. For larger or invoiced purchases, Payman and AgentCard document an explicit human-approval step. On the pages we checked, no single vendor below documents all three.
This page does not give legal, tax, or regulatory advice — check with your finance and legal team before routing real procurement spend through any of these controls.
Comparison table
All quotes below are verbatim from each vendor's own docs or product page, accessed 2026-09-30 unless noted. "Not documented" means the specific control wasn't found on the page(s) checked — not that the vendor doesn't have it.
| Provider | Use case(s) | Rail | Per-agent controls documented | Approval / human-in-the-loop | Audit / reporting | Availability (vendor's words) | Source |
|---|---|---|---|---|---|---|---|
| Stripe Issuing (agents) | SaaS/tools, larger purchases | Card | "give each agent its own virtual card with per-agent spend limits, merchant category controls, and custom authorization rules" | Partial — a real-time authorization decision rather than a blocking approval gate: "if anything looks off, decline it and flag with the agent or a human reviewer" | Yes — "every transaction your agent makes is traceable" | Private preview | docs.stripe.com/issuing/agents |
| Lithic | SaaS/tools, high-volume programmatic issuance | Card | spend_limit field with spend_limit_duration of ANNUALLY, FOREVER, MONTHLY, or TRANSACTION; velocity rules cap "the maximum total spend allowed within the period, in cents" |
Not documented on the pages checked | Not documented on the pages checked | Not stated on the pages checked (an MCP server is announced on Lithic's blog) | docs.lithic.com/docs/spend-limits |
| Privacy.com | SaaS/tools (solo dev or small team) | Card | spend_limit field ("Transaction requests above the spend limit will be declined") plus a MERCHANT_LOCKED card type |
Not documented on the pages checked | Not documented on the pages checked | Not stated on the pages checked (an MCP server is documented) | developers.privacy.com/docs/cards |
| AgentCard | SaaS/tools, larger purchases | Card | amount_cents per card, single_use/multi_use types, scope_preset merchant lock (e.g. ai_labs) |
Yes — production card creation returns 202 approval_pending, an approval_url the member confirms with a passkey |
Not documented on the pages checked | Sandbox self-serve; production requires an active subscription | docs.agentcard.sh/api-reference/cards/create |
| Crossmint | SaaS/tools, crypto-native stacks | Card + onchain wallet | "Every allowance is scoped, explicit, and revocable"; "Card limits hold at Visa and Mastercard, wallet limits hold onchain" (no specific field name given on the page checked) | Not documented on the page checked | Not documented on the page checked | Developer API/SDK access (no GA/beta language found on the page checked) | docs.crossmint.com/agents/overview |
| Coinbase CDP Agentic Wallet | API/cloud usage | Stablecoin wallet (MCP) | "Max per call: Max for a single payment (e.g., $0.05)" and "Max per session: Total session limit (e.g., $5.00)," set in the wallet UI | No per-payment approval — limits are pre-set by a human; "Agents respect these limits but can't change them" | Not documented on the page checked | MCP server shipped (no GA/beta language found on the page checked) | docs.cdp.coinbase.com/agentic-wallet/mcp/faq |
| Circle Agent Wallets | API/cloud usage, company-wide governance | Stablecoin wallet | Per-tx/daily/weekly/monthly caps that must be "monotonic: per-tx ≤ daily ≤ weekly ≤ monthly"; "transfer limits, recipient allowlists, and contract blocklists per agent wallet" | Partial — no per-payment gate, but "setting or resetting limits requires OTP confirmation in an interactive terminal session" | Not documented on the pages checked | Circle developer account (no GA/beta language found on the pages checked) | developers.circle.com/agent-stack/agent-wallets |
| Payman | Larger or invoiced purchases | API policy (rail-agnostic) | "Per Transaction" max, "Daily Limit," "Monthly Limit" | Yes — a "Threshold" above which "manual approval is needed" | Not documented on the page checked (Wayback snapshot) | Not confirmed on the page checked | Wayback: docs.paymanai.com/dashboard-guide/policies |
| Skyfire | API/cloud usage | Signed token (protocol-agnostic) | Per-token amt amount cap, mnr request-count cap, exp expiry — all set once at token issuance |
Not documented on the pages checked | Not documented on the pages checked | Beta: "Join the Skyfire Beta and start building the future" (skyfire.xyz/product) | docs.skyfire.xyz/docs/pay-token |
| Openfort | Company-wide governance | Onchain wallet + API policy | "Spending caps, contract allowlists, and time-boxed sessions that let agents sign inside guardrails" | Yes — "multi-party approvals" named alongside anomaly detection and real-time alerts | Yes — "full audit trails" | Self-serve ("Start building for free," per the page checked) | openfort.io/solutions/ai-agents |
How to choose
- If every vendor takes a card, start with the card layer. AgentCard documents a passkey approval step before a production card issues, and Stripe Issuing's agent product (private preview) documents real-time authorization decisions with a human reviewer in the loop; Lithic and Privacy.com document hard per-card limits but no approval gate on the pages checked, so add your own approval step in front of them if you need one. See virtual cards for AI agents for a card-only comparison.
- If the agent pays per API call or per token of usage, don't force it onto a card. Coinbase's CDP wallet and Circle's Agent Wallets both document per-call/per-period caps built for exactly this; card rails are usually a poor fit for sub-cent, high-frequency charges.
- If a purchase is large enough to need sign-off, use a vendor that documents a threshold, not just a cap. A cap stops runaway spend; only Payman, AgentCard, Stripe Issuing, and Openfort in this table document an actual human-review or approval step above a line you set.
- If you're running many agents across many vendors, the cap needs to live somewhere central. Circle's monotonic four-tier limits and OTP-gated limit changes, and Openfort's allowlists plus multi-party approvals, are the two entries here built to govern more than one agent at once rather than one card or one wallet.
- What to ask a vendor before signing — five concrete questions, not marketing copy:
- "Is this feature GA, beta, or private preview today, and what's the wait time for access?" (Stripe's agent Issuing product, for example, is in private preview as of this check.)
- "Is the spending cap enforced by your system, or only by the client SDK I integrate?" (a cap enforced only in client code can be bypassed by a bug or a compromised agent; ask which layer actually declines the payment.)
- "Can the agent itself ever raise its own limit, and if so, what stops it?" (Circle and Coinbase's CDP wallet both document that the agent cannot change limits set by a human — ask every other vendor the same question directly.)
- "What does your audit log actually capture — dollar amount, or which agent/policy/approver authorized it?" (only some vendors here document the latter; see Step 4 of our setup guide).
- "What happens if a card or wallet is compromised — can I freeze it and revoke the agent's credentials in one action?" (see our incident runbook for what to check before you sign.)
What's still missing
- No vendor here documents a control that judges intent. Every cap, allowlist, or approval threshold in the table enforces a pre-set boundary (amount, merchant, category, time window) — none evaluates whether a specific purchase should happen given the context an agent was given.
- Corporate-card platforms mostly don't document agent-specific controls yet. We checked Ramp's and Brex's card pages and didn't find a description of AI-agent card issuing on the pages we checked, so they're not in the table. Mercury's API page mentions agent spend ("Issue cards instantly and control team and agent spend"), but we didn't find a named per-agent field or limit, so it's not in the table yet either; we'll add it when that's documented. See the sources-check for the exact quotes and URLs.
- Pricing is largely undisclosed. None of the ten providers in the table publish agent-specific pricing on the pages checked.
- Availability status varies and changes fast. Stripe's agent Issuing product and AgentCard's production tier are both gated (private preview and subscription-required, respectively) as of this check — confirm current status directly with the vendor before building on it.
FAQ
Which payment provider is best for AI agent procurement? There isn't one — it depends on what the agent is buying. For card-accepting vendors, AgentCard documents both a per-card limit and a passkey approval step; Stripe Issuing's agent product (private preview) documents per-agent limits plus real-time authorization decisions. For metered API/cloud usage, Coinbase's CDP wallet and Circle's Agent Wallets document per-call and per-period caps on a stablecoin rail. For purchases that need sign-off before they clear, Payman and AgentCard both document an explicit threshold or approval gate.
Do I need a crypto wallet to let an AI agent pay for procurement? No, if every vendor accepts cards. A scoped virtual card with a documented per-card limit — Lithic, Privacy.com, AgentCard, or Stripe Issuing — covers that case without a wallet. Wallet-based rails like Circle and Coinbase's CDP wallet matter specifically for usage-metered API vendors that don't bill through a card network; see our payment APIs comparison for that rail in more depth.
Can the agent raise its own spending limit if it needs to buy something bigger? On the providers that document this explicitly, no. Circle requires human OTP confirmation to set or reset a limit, and Coinbase's CDP wallet states "agents respect these limits but can't change them." Neither AgentCard, Stripe, Privacy.com, Lithic, Crossmint, Payman, Skyfire, nor Openfort documents this specific question on the pages we checked — ask the vendor directly before assuming the answer, and treat limit changes as a human-only action in your own policy regardless.
What's the difference between a spending cap and an approval threshold?
A cap is a hard ceiling the rail enforces automatically — Lithic's spend_limit or Circle's per-tx/daily caps, for example. An approval threshold routes a payment to a human before it clears, even if it's under the cap — Payman's "Threshold" field and AgentCard's passkey-confirmed approval_pending response are the two clearest examples in this table. A company procurement policy typically needs both: a cap that can never be exceeded, and a lower threshold that pulls in a human for anything unusual.
Pink Agentic AI Payment (early access) enforces per-agent spending caps, allowlists and approval thresholds at the MCP layer, before a payment executes.
This page is published by PinkWallet, which is building Pink Agentic AI Payment (early access). We are not neutral in this space, which is why every factual claim links to a primary source.
