Sapiom Raised 15M to Build Stripe for AI Agents - Should AI Agents Have Corporate Credit Cards?

finance_carlos · February 12, 2026, 11:04pm

I spent the last week digesting the Sapiom announcement – a $15M seed round led by Accel to build what they’re calling “Stripe for AI Agents” – and I can’t stop thinking about the implications for every finance and operations team in tech.

The Core Proposition

Sapiom, founded by former Shopify payments director Ilan Zerbib, is building financial infrastructure that lets AI agents autonomously purchase software, APIs, compute, and data services. Their pitch: every time an AI agent connects to an external tool (Twilio for SMS, AWS for compute, a data vendor for enrichment), it needs authentication and a micro-payment. Sapiom makes that seamless.

The investor lineup tells you this isn’t fringe: Accel led, with Okta Ventures, Menlo Ventures, Anthropic, and Coinbase Ventures participating. When an identity company (Okta), an AI company (Anthropic), and a crypto company (Coinbase) all write checks into the same payments startup, the thesis is clear: non-human economic actors are coming, and they need financial plumbing.

Why This Should Terrify Every VP of Finance

Here’s what keeps me up at night. Right now, if an engineer wants to spin up a new SaaS tool, there’s a procurement process. Maybe it’s lightweight (expense it under $500) or maybe it goes through formal approval. Either way, a human makes the decision, a human swipes the card, and there’s an audit trail tied to a person.

Sapiom’s vision removes the human from the loop. An AI agent decides it needs a service, provisions it, pays for it, and starts using it – all within milliseconds. The enterprise dashboard lets you set budgets and approval thresholds (say, autonomous spending up to $500 per transaction), but the fundamental shift is this: your AI agents are now spending entities with their own purchasing authority.

From a unit economics perspective, this is fascinating. If an agent can autonomously purchase the cheapest compute for a specific workload, switch between providers based on real-time pricing, and negotiate volume discounts programmatically – the efficiency gains could be massive. Imagine your cloud bill optimized not by a FinOps team reviewing dashboards monthly, but by agents making thousands of micro-decisions per hour.

The Numbers Problem

But here’s where my finance brain starts breaking. How do you budget for autonomous spending? Traditional SaaS procurement has predictable costs – you sign an annual contract, you know the line item. Agent-driven procurement is inherently variable. Your costs fluctuate based on:

How many agents you deploy
What tasks they’re assigned
What services they discover and decide to use
How aggressively they’re configured to spend

This is usage-based pricing on steroids. And we already know that most finance teams struggle with usage-based models (ask any company that switched from seat-based to consumption-based pricing how their forecasting changed).

The Governance Gap

Sapiom’s dashboard offers budget caps, spending rules, and approval thresholds. But I’ve seen enough corporate card programs to know that controls at issuance don’t prevent problems at scale. When you have 50 AI agents, each with $500 autonomous spending authority, that’s $25,000 in potential uncontrolled spend per transaction cycle. Scale to 500 agents and you’re looking at $250,000.

And unlike human employees, agents don’t have judgment about appropriateness. An agent optimizing for speed will pick the most expensive option if it’s the fastest. An agent optimizing for cost might pick a vendor with terrible security practices. An agent given broad authority will find creative ways to spend that no one anticipated.

The Opportunity

That said – I think this is inevitable. The agentic economy is projected to generate $1-3 trillion in orchestrated revenue by 2030. McKinsey estimates autonomous procurement agents can capture 15-30% efficiency improvements. Visa and Mastercard are already preparing payment infrastructure for AI agents.

The question isn’t whether AI agents will have spending authority. It’s whether finance teams will be ready to govern it.

What I Want to Know

For those of you building AI agent infrastructure or deploying agents in production:

How are you handling the procurement question today? Manual approval for every tool an agent uses?
What’s your comfort level with autonomous spending limits? $100? $500? $5,000?
Who owns the agent spending budget – engineering, product, or finance?
Are you thinking about agent-to-agent commerce (your agent buying from another company’s agent)?

I have a feeling this is going to be the FinOps 2.0 conversation, and most of us are not ready for it.

cto_michelle · February 12, 2026, 11:05pm

Carlos, this is the post I didn’t know I needed to read today. I’ve been evaluating Sapiom-like capabilities for our platform and the CTO perspective is slightly different from the finance angle.

The part you’re right about: governance is the hard problem, not payments. Payments are solved – Stripe exists, APIs exist, ACH exists. What doesn’t exist is a coherent framework for “who is accountable when an agent makes a bad purchasing decision?”

At my company, we’re running about 15 AI agents in production right now. Every single one of them has hardcoded vendor lists. Agent X can call Twilio and OpenAI, period. There’s no “discovery” – we’re nowhere near letting agents browse a marketplace and pick vendors autonomously. And honestly, I think the Sapiom vision is 2-3 years ahead of where most enterprises actually are.

Where I disagree with you slightly: I don’t think this is primarily a finance problem. It’s an architecture problem. The reason agents need spending authority is because we’re building them as autonomous actors instead of as orchestrated workflows. If you design your agent system with a central orchestrator that manages all external integrations, you don’t need per-agent spending limits – you need system-level budgets managed by the platform team.

The real question is: should agents be economic actors at all, or should they be execution engines within a governed system? I lean toward the latter for enterprise. Sapiom’s consumer play (agents ordering Ubers, shopping on Amazon) is a different story entirely.

That said, the investor mix you mentioned is telling. Anthropic investing in agent financial infrastructure means they see a future where Claude agents have wallets. That’s not 2030 – that’s probably late 2026 or early 2027.

security_sam · February 12, 2026, 11:06pm

I’m going to push back on both of you from a security perspective, because I think you’re having the wrong conversation.

The question isn’t “how do we govern agent spending?” The question is: “how do we prevent the largest automated fraud surface ever created?”

Think about what Sapiom is proposing. AI agents with authentication credentials, payment capabilities, and autonomous decision-making authority. From a threat model perspective, this is a dream target:

Credential theft – If an attacker compromises the agent’s API keys, they now have a spending entity they can drain. Unlike a stolen corporate card that gets flagged after unusual purchases, an agent making API calls to cloud services looks like normal behavior.
Prompt injection as financial attack – If an agent can be manipulated through prompt injection to purchase services from an attacker-controlled endpoint, you’ve just created a money printer. “The cheapest compute available” could be an attacker’s honeypot offering below-market rates.
Supply chain poisoning – Agent-to-agent commerce means an attacker doesn’t need to compromise your agent. They just need to offer an attractive service on whatever marketplace agents use to discover vendors. Slopsquatting for services instead of packages.
Audit trail corruption – When a human makes a purchase, you can question the human. When an agent makes a purchase, you have logs. Logs can be tampered with. Who verifies that the agent’s stated reasoning for a purchase matches reality?

Carlos mentioned $500 autonomous spending limits. Multiply that by thousands of transactions per day across hundreds of agents, and even a 1% fraud rate represents massive losses. And unlike credit card fraud, there’s no chargeback mechanism for API consumption – once compute is used, it’s used.

I worked at Stripe on payment security. The entire payment industry’s fraud prevention was built around detecting unusual human behavior. None of those models work for agents. We need an entirely new fraud detection paradigm, and I don’t see Sapiom talking about this.

product_david · February 12, 2026, 11:06pm

Fascinating thread. I want to bring the product strategy lens to this because I think everyone’s missing the market timing question.

Carlos framed this as “FinOps 2.0” and Michelle called it “2-3 years ahead.” I think the truth is somewhere in between, and the reason is the platform war that’s happening right now.

OpenAI just launched Frontier – their enterprise agent platform with HP, Uber, and Oracle on board. Anthropic invested in Sapiom. Google is building agent infrastructure into Vertex AI. Salesforce has Agentforce. Every major platform player is racing to own the “agent runtime” layer.

Here’s the product insight: whoever controls the agent’s wallet controls the platform. Think about it. If Sapiom becomes the standard payment layer for AI agents, they have visibility into every agent’s purchasing behavior across the ecosystem. That’s the most valuable data asset in enterprise software – not what agents produce, but what agents consume.

This is exactly what Stripe did to commerce. By processing payments, Stripe didn’t just become a payments company – they became a data company with unmatched visibility into business health, growth patterns, and market trends. Sapiom is making the same play for AI agents.

Sam’s security concerns are valid but also somewhat manageable. Payment security has solved similar problems before with tokenization, fraud scoring, velocity checks, and merchant verification. The difference is that AI agents transact at machine speed, so the detection systems need to operate at machine speed too.

The real product question for me is: will enterprises actually want autonomous purchasing, or will they want “recommended purchasing” where the agent finds the best option and a human clicks approve? My bet is that the second model wins for the next 3 years, and Sapiom’s dashboard becomes less of a governance tool and more of an approval workflow tool. That’s a much smaller TAM than autonomous commerce.