Last month, a quiet rebellion erupted inside the walls of Sanofi, the French pharmaceutical giant with a market cap north of $130 billion. The company did what few enterprises dare: it ripped out its ServiceNow deployment—the gold standard for IT service management (ITSM) used by over 7,000 global organizations—and replaced it with a homegrown AI agent powered by Anthropic’s Claude and a niche automation platform called Elementum.
No press release. No fanfare. Just a quiet shift in a domain that rarely gets headlines outside of CIO conferences. But for those of us who have spent years watching the slow creep of centralized SaaS eat enterprise architecture, this is more than a procurement decision. It is a signal. A crack in the monolith.
I saw the news flash across my Telegram channels—those same channels where we debate validator sets and L2 sequencer centralization. And I couldn’t help but see the parallel. Sanofi isn’t just building a bot. It is unbundling trust. The same way we told the 2017 ICO crowd that “code is law” only to realize that multisig admins still held the keys, Sanofi is discovering that replacing a SaaS platform with an LLM-and-tools combo doesn’t automatically make IT governance more transparent. It makes it different. And sometimes different is scarier.
Context: The Unspoken Contract
For decades, enterprise IT has operated on a simple premise: you pay a vendor a subscription fee, and they handle the complexity. ServiceNow processes tickets, manages workflows, and gives you a dashboard. The contract is implicit—you trust ServiceNow to keep your data secure, to not go down, to not leak your employee’s password reset requests. That trust is built on SLAs, audits, and vendor lock-in.
But the economics of trust have shifted. Large language models—Claude, GPT-4, Gemini—have turned the “knowledge worker” function into a commodity API call. Suddenly, the value that ServiceNow extracted from bundling ITSM workflows with proprietary automation logic can be replicated with a few lines of Python, a vector database, and a well-prompted model. Sanofi realized that the cost of maintaining a ServiceNow instance for its 100,000+ employees could instead buy a dedicated team to build and run an internal agent that never needs a contract renewal.
According to the report, Sanofi paired Claude (likely through Amazon Bedrock or a private API deployment) with Elementum, a company that provides a platform for building and orchestrating AI agents. The combination allows Sanofi to handle IT requests—password resets, software provisioning, incident triage—without sending data to a third-party cloud. The pharmaceutical industry is heavily regulated (FDA, cGMP, HIPAA). Keeping data within their own network is not just a preference; it’s a compliance mandate. ServiceNow, for all its enterprise features, still operates as a centralized repository that handles sensitive employee and operational data. A self-hosted AI agent changes that.
Core: The Architecture of Faith
Let’s get technical for a moment. Sanofi’s AI agent is not a monolithic bot. It’s a composition: Claude provides the reasoning layer, Elementum manages the tool-calling and state, and Sanofi’s internal APIs wrap around Active Directory, ServiceNow’s own backend (yes, they kept the data but dumped the workflow), and asset management databases. Each tool call is an atomic action—reset password, grant access, escalate to human. The agent’s “brain” is Claude, but Claude cannot directly read the internal network. It must go through Elementum’s guardrails.
This is where the blockchain analogy hits home. In a DAO, we use smart contracts to automate treasury management. The code is visible, auditable, and deterministic. But here, the agent’s decisions are probabilistic. Claude might hallucinate a tool argument, or misinterpret an employee’s request. The “trust” is not in a contract’s logic but in a black box model shipping weights from Anthropic’s servers. We didn’t design our financial systems to rely on an opaque inference engine. Yet that is exactly what Sanofi is doing.
But here’s the kicker: the transition is economically rational. A former colleague at a Big Four firm once calculated that a global company with 50,000 employees spends roughly $8-12 million annually on ServiceNow subscriptions alone. Sanofi likely paid more. With an internal agent, the variable cost shifts to API calls (Claude) and platform fees (Elementum). Assuming Sanofi processes 100,000 IT tickets per month, each ticket requiring 500 tokens of context and 200 tokens of output, the monthly Claude bill might hover around $5,000-$10,000. Elementum’s licensing adds maybe $200k per year. The internal team of 3-5 engineers costs $600k. Total annual cost: under $1.5 million. That’s a 70-85% cost reduction.
But cost isn’t the whole story. The real transformation is in control. Sanofi now owns the agent’s behavior. They can tweak it, audit it, and—crucially—decide when to say “no” to a request. ServiceNow’s workflows are a black box unless you pay for premium support. With an open API stack (even if Claude is closed), the logic is debuggable. The training data for tool usage, the rollback strategies—all under Sanofi’s governance.
Contrarian: The New Centralization
Yet, let’s not romanticize this. I’ve been here before. In 2020, I lost $15,000 in a yield farming hack because I trusted a smart contract that “looked” decentralized. The multisig was held by three friends of the founder. Sanofi’s choice to use Claude and Elementum creates a different kind of dependency. Claude is a black box model trained on unknown data. If Anthropic changes the model’s behavior—fines a safety filter that blocks legitimate IT requests—Sanofi has no recourse. They are hostage to a another vendor’s alignment choices.
Elementum could pivot, be acquired, or go out of business. The platform that manages the agent’s state is as critical as the model. If Elementum shuts down, Sanofi’s agent loses its memory and tool-calling framework. The same lock-in that plagued ServiceNow now reappears with smaller, less stable players.
And let’s talk about security. Confidence is low that Sanofi’s agent has been properly hardened. Red teaming AI agents is still an immature science. A prompt injection attack could trick Claude into executing a dangerous tool call—like deleting a user directory. The attack surface is wider: the API gateway, the embedding database, the model itself. In a traditional ITSM, the workflow is deterministic; in an AI agent, it’s stochastic. The probability of a catastrophic failure might be low, but the impact is high.
I’m reminded of a conversation I had with a DAO developer in 2022. He said, “We replaced a CEO with a smart contract, and now we’re replacing the smart contract with an AI agent.” We laughed. But it’s not a joke. Truth in enterprise IT isn’t in the SLA anymore—it’s in the code that can’t be audited.
Takeaway: The Next Era of Enterprise Trust
Sanofi’s move will not be the last. Within the next two years, I predict at least three other Fortune 500 companies—in pharma, finance, and logistics—will announce similar transitions. The economics are too compelling. But the question we must ask as an industry is not whether to build agents, but how to build trust around them.
We need new primitives: verifiable logs of every agent decision, on-chain commitments of tool call history, maybe even decentralized governance of the agent’s prompt templates. Imagine a DAO that votes on which model to use, or a staking mechanism that penalizes the agent for incorrect actions. This sounds wild, but so did a decentralized exchange in 2015.
The Sanofi rebellion is a preview of the compositional enterprise—where companies assemble their own infrastructure from LLMs, automation platforms, and in-house APIs. It mirrors the crypto ethos of “don’t trust, verify.” But right now, the verification tools are missing. We have the first act of a new play. The ending is still unwritten.