Companies pay for enterprise software, then often pay outside providers to help employees navigate the applications they already license. Even a routine IT request can mean contacting an outsourced service desk, where someone follows the company’s guidelines to resolve a problem and close a ticket, generating another charge.

At Sanofi, those recurring costs helped expand the ambitions for Concierge, the healthcare and pharmaceutical company’s internally developed agentic AI assistant. It began by answering employees’ questions about the company, but Sanofi’s chief digital officer Emmanuel Frenehard saw an opportunity for it to complete tasks that his company paid outside providers to handle.

“We were jumping from one system to another because each system managed only part of the process. I started thinking, what if we could bring all that data into one interface and create new processes that were no longer constrained by the barriers between those systems?,” Frenehard tells me in an exclusive interview. “Agentic AI is extremely powerful, but it is also costly. It should be used where the returns are high, not for small automations.”

McKinsey’s Technology Trends Outlook 2026 reports that 89% of organizations regularly use AI, yet only 37% attribute any positive impact on earnings before interest and taxes to their AI programs. The report argues that the biggest gains will come from redesigning workflows around employees and AI agents working together, an approach Sanofi is pursuing through Concierge’s support and purchasing capabilities.

The company reports 65,000 weekly users and 24 million conversations. Employees use it to submit purchase requests, while sales representatives turn to a specialized version to prepare for customer interactions. Moreover, it also claims 1.2 million agent invocations and 90% positive feedback. However, Concierge is not Sanofi’s first AI implementation. The company launched plai with Aily Labs in 2023, an analytics companion that 20,000 employees use daily to analyze company data. It also developed Muse with OpenAI and Formation Bio to help recruit patients for clinical trials.

As agentic AI becomes more capable and widely available, how much of the underlying software can enterprises actually stop paying for? An AI assistant can handle an employee’s request while existing applications still execute the work, leaving companies with a new AI platform to fund and the same software contracts to renew.

Frenehard says employees often struggled to find clear answers about their benefits and workplace rules because the company’s own documentation was difficult to navigate. Building a better alternative proved harder than choosing a model. An unnamed Seattle technology company worked with Sanofi on an early version that collapsed after roughly 10 simultaneous conversations, prompting the company to rebuild around AWS services and Claude models while focusing its own efforts on the employee experience.

That experience also depended on connecting Concierge to systems such as Workday for organizational context, Snowflake and ServiceNow for support information. Sanofi used AI to rewrite knowledge articles into forms models could interpret, a process that forced people to resolve ambiguities in the company’s own policies.

“We worked with our legal team to clean up all our procedures and policies in the company and create a single source of truth,” Frenehard says. Sanofi’s experience shows that faster AI service depends as much on resolving the complexity buried in a company’s own information as on the model used to interpret it.

Moving AI Beyond Enterprise Software Interfaces

Concierge runs on Amazon Bedrock, which lets Sanofi swap foundation models rather than commit to a single provider, while its data sits on Snowflake . Elementum AI helps execute workflows, beginning with software license management.

“Sanofi realized pretty quickly that putting an AI agent on top of legacy software doesn’t fundamentally change anything. The data and business logic still live in those systems, and employees are still dependent on them,” says Nader Mikhail, Elementum AI’s founder and CEO. “With Concierge, the goal was to move toward owning the data and workflows underneath the experience, starting with software license management and expanding from there.”

Frenehard described employee onboarding as an example of the fragmentation he wanted to overcome, with managers moving between service requests and HR records because applications divide up the work.

Sanofi says Concierge’s purchasing workflow reduces the time needed to prepare and submit a request from a three-day baseline to less than five minutes. Employees upload a quote and answer questions in ordinary language. The five-minute figure describes request submission, leaving out the time the subsequent purchasing process may require.

“The savings disappear when AI automates one step but leaves people doing the same manual work around it,” Mikhail says. “At Sanofi, the goal wasn’t to make individual tasks faster. It was to redesign the workflow end to end, including the data, business rules, approvals, and exceptions behind it. That’s how you get from a great demo to savings that actually hold up in production.”

The company also reports 80% touchless support resolution and more than 20,000 years of support waiting time returned in 2025. For customer-facing teams, Sanofi claims a 40% reduction in time previously spent on back-office activities. Frenehard says 400 sales representatives now use Concierge for Field daily. His business case anticipates more sales from the time recovered, but he does not provide evidence of that revenue gain.

He also resists converting every saved minute into money. “We don’t quantify productivity as a dollar metric because I don’t think you can quantify it,” he says. “We could build an intake form with an agent at a cost. Of course, there’s always a cost. But when we monitor it and when we pilot it, the cost is lower than the cost of a third-party service because they also have to have a markup. And we can do it in a way that the experience with the supplier is much better.”

Sanofi’s AI board, which Frenehard co-chairs with its chief financial officer, meets about every six weeks. The company evaluates individual activities against their costs, he says. The company did not provide total development or operating costs for Concierge, however.

Where Sanofi Draws The Line On Agentic AI

Frenehard sees another hidden cost in agentic AI. Companies do not need a model to reason through every decision when a simple rule can handle it more cheaply. If a U.S. employee requests an application available only in France, Sanofi can check that rule directly instead of paying an AI model to reach the same answer.

“The agent will wrap up that no into an elegant no to explain to you why,” he says. “I think a large portion of what we do in the end is deterministic. It’s not agentic. But the glue, the experience layer, starts by being the agent itself.”

The distinction also matters when an agent encounters information it should not disclose. Mikhail describes an early support-agent trial in which the system revealed private information to someone without access, although he does not identify the organization involved.

“Since then, we have reinforced the deterministic platform to auto-inherit user permission. No agent gets to decide if it’s convinced you need that information,” he says. “This is the part of the AI conversation I think we’re getting wrong. The scary question isn’t whether a model can complete a task. It’s what the model is allowed to touch and what happens when it’s wrong.”

At Sanofi, employee roles determine which of roughly 40 agents people can access, Frenehard says. Elementum’s Mikhail argues that other companies can begin with a single workflow, but the more important question is who defines how that workflow should operate. “The part the industry underplays is that AI can automate execution, but humans still have to define what good execution looks like.”

Sanofi’s experience offers other enterprises a working example of how AI can change their relationship with software. But getting there required the company to clarify its own rules and assemble outside technology around the workflows it wanted to improve. The broader promise of agentic AI is that enterprises can gain more control over how they operate, with software adapting to their processes rather than forcing those processes to conform to the software.