Inside OpenAI and Anthropic's conventional enterprise stack: Salesforce, Workday and Stripe

AI labs are widely assumed to live in the future, with agents writing everything they use internally. A new reconstruction of their stacks, based on job postings, vendor case studies and trust centers, suggests the opposite: they look like conventional enterprises.

What the researchers found in the stack

Both companies use Salesforce for CRM and Workday for HR and payroll. The difference in finance is minor: Anthropic uses NetSuite, while OpenAI uses Oracle Fusion. Zip handles procurement, Stripe processes payments, and OpenAI also uses Metronome for usage-based billing.

There are some intriguing details. Anthropic maintains four recruiting tools at once: Greenhouse, Gem, ModernLoop and BrightHire, apparently for different roles and markets. In support, Anthropic uses Intercom Fin, while OpenAI has assembled a combination of Salesforce, Pylon and TaskUs. The last of these provides outsourced human staff. Workato connects the various previous-generation systems.

Why they did not vibe-code their own stack

An obvious question follows: if AI is so good at writing code, why not build a CRM over a weekend? The analysis offers several reasons.

• These are systems of record. An error in payroll, taxes or revenue costs more than any software license, and auditors want to see Workday and NetSuite, not the phrase “the agent calculated it.”
• Compliance is built into the product. Payroll across dozens of countries, SOC 2 and tax rules require years of work and legal expertise.
• Network effects. Slack, Stripe and Salesforce are valuable because everything else is already connected to them.
• Focus. An engineer at an AI lab is the most expensive resource on the market, and assigning one to build an internal CRM makes no economic sense.
• Data. These systems contain years of history, and migration is more daunting than the subscription cost.

What is already being replaced

The layer above the process is being replaced: interview coordination, interview notes, frontline support, lead research and documentation. This is where Fin, BrightHire, Clay and Mintlify operate—the new generation of AI agents inside SaaS. In other words, it is not the core that is breaking apart, but the interface and the routine work around it.

What product teams and founders should do

This leads to a set of practical rules for SaaS products.

1. Be a system of record, not an interface. The vendor that stores the “truth” about money, people and contracts survives. The one that merely presents someone else's data attractively is next in line for replacement.
2. Take responsibility: certifications, guarantees and legal risk. AI can write the code, but it cannot sign off on an audit.
3. Build a proper API and MCP. Agents will work through your product, and a tool that is convenient for agents will not be discarded—it will be used more often.
4. Move from per-seat pricing to outcome-based pricing. There will be fewer human users, and the seat-based model will die with them.
5. Build AI into the product yourself, as Intercom did with Fin. Otherwise, someone will build it on top of your API and take the customer. If you have unique data from thousands of customers, you still have a viable business.

Limitations

This is a reconstruction based on indirect sources: job postings, vendor case studies and trust centers. Some tools may have been pilots or used by individual teams rather than company-wide. The specific product names should therefore not be treated as definitive. But the broader conclusion aligns closely with what the market shows: for now, AI is taking over the routine work surrounding systems of record, rather than replacing the systems themselves.