A woman manages a secure, AI-powered healthcare workflow on a computer, with connected clinics, robots, and patient records.
WTW, the benefits administrator formerly known as Willis Towers Watson, has put a multi-agent platform into production on Microsoft Foundry and Microsoft Agent Framework. The company says humans stay in charge of the decisions that matter. Microsoft's customer story, dated November 5, 2026, describes the project in detail. WTW's own investor call a week earlier gives a firmer picture of its scale.

The case is worth a look for IT teams. It is a real workflow in a regulated industry, with named Azure services, defined exception paths, audit retention and a hard limit on what agents may decide. It is not a chatbot demo.

A woman manages a secure, AI-powered healthcare workflow on a computer, with connected clinics, robots, and patient records. What WTW built, and why​

WTW's Benefits Delivery & Outsourcing unit runs healthcare, insurance and pension benefits for large U.S. employers. Microsoft's account says it serves about 12 million employees, dependents, and retirees. Through its proprietary BenefitConnect platform it handles roughly 1 million procedures and 2.2 million cases a year. It also makes sure participant data reaches more than 2,000 vendors.

Routine work was already automated. The trouble was the small share of cases that fall off the "happy path," as WTW's Jeff Commander put it. An analyst had to review each exception, decide what to do, act across several systems and document every step. That pulled experienced people away from harder problems. The stakes are real: one missed file or step can interrupt a participant's benefits or access to care, or delay a pension payment.

WTW's answer was a governed multi-agent layer inside BenefitConnect. It recognizes exceptions, investigates them and recommends next steps, while a person makes the final call. Commander compared it to an airport control tower. Specialized agents do the work, and a central control decides who goes and when to stop.

The Microsoft stack, piece by piece​

WTW built the platform under its Violet AI suite, working with implementation partner AIS. Per Microsoft's description, the pieces are:

  • Microsoft Foundry and Microsoft Agent Framework: the agent foundation. The agents run inside a WTW-governed Azure environment.
  • Azure Container Apps and Azure Container Registry: agent services run as containers, and versioned images support consistent deployment, scaling and updates.
  • Azure Service Bus: messaging between agents and processes. Work is queued and throttled, and it survives a failed step.
  • Azure Cosmos DB: platform state and activity records.
  • Microsoft Entra ID: identity and role-based access for both people and services.
  • Azure Key Vault: credentials, secrets and encryption keys, so no agent holds sensitive credentials directly.
  • Microsoft 365 Agents SDK: when a step fails, it surfaces context and asks clarifying questions. These help decide whether to continue, retry, defer or escalate.
  • Microsoft Teams: escalations land with an analyst, who can act from the message using Adaptive Cards.

Microsoft does not disclose the model choices, the number of agents, the network topology or performance benchmarks. It also does not say how client data is isolated. The story says client data must be separated and decisions must be auditable, but it doesn't name the technical controls that do that.

The vendor-export proof of concept​

The platform was running within five months. The first process WTW moved over was vendor exports. This packages, formats and sends participant election data to third-party vendors, which sets up healthcare coverage for participants and their dependents. WTW chose it because it exercised nearly every platform capability.

The automated flow runs like this:

  1. A specialized agent spots a scheduled export and checks the prerequisites.
  2. The platform launches the job and monitors each action.
  3. It validates key metrics and results, triggers encryption, delivers the files and confirms the vendor received them.
  4. The process closes with a full audit trail.

Four layers of human oversight​

Microsoft lists four layers of oversight:

  • Exception monitoring: the platform flags unusual responses, failures and unknown patterns.
  • Issue classification: agents analyze the situation and recommend an action.
  • Human-in-the-lead escalation: anything needing judgment goes to a business analyst in Teams.
  • Decision boundaries: agents never make final decisions on participant eligibility or coverage. A business analyst always does.

Violet also logs every human-in-the-loop event and keeps the logs for 10 years, which Microsoft ties to audit requirements. The story does not say which audit standard applies, where the logs are stored, or whether every record type gets the same retention.

Reuse is the real selling point​

Commander said the integrations, APIs, orchestration services, human-in-the-loop frameworks, throttling controls and audit logs already exist. New workflows only need procedure-specific logic and agent behaviors. WTW says this has cut onboarding time for new workflows from months to weeks. It gives no exact before-and-after figures.

For admins and developers, this is the transferable lesson. Governance, identity, secrets handling and escalation are built once and shared. Each new agent then inherits them instead of reinventing them.

The savings claim: expectation, then a later update​

The customer story says WTW expects to save tens of thousands of hours on vendor exports alone. It expects further savings from case management and HRIS, the two other processes it has since moved over.

A week earlier, on October 29, 2026, CEO Carl Hess gave a firmer update on WTW's third-quarter earnings call, according to published transcripts. Hess said the Microsoft-built agentic administration process is deployed and managing end-to-end workflows for 160 vendors across more than 20 clients. He added that agents complete vendor exports, HRIS data imports and case management with human oversight. Hess said the system has saved tens of thousands of human hours so far.

Treat these as company-reported figures. Neither source gives an absolute total, a method for measuring the hours or a breakdown by workflow. The sources also don't report error rates, cost savings, staffing changes, implementation costs or accuracy compared with the old process. Both the customer story and the earnings call are WTW and Microsoft describing their own results, and a customer story is by nature a favorable showcase.

Takeaways for IT teams​

  • Design the boundaries first. WTW defined what agents may never decide before building them.
  • Use the platform's own plumbing. Managed identity through Entra ID, secrets in Key Vault and queued messaging through Service Bus keep agents from becoming a security gap.
  • Put escalation where people already work. Teams with Adaptive Cards lets an analyst act without leaving the conversation.
  • Start with lower-complexity work. WTW says it is deliberately training the platform on simpler procedures before moving to higher-value ones.
  • Expect governance to be the long pole. WTW lists auditability and data separation as design drivers, and the 10-year log retention shows how seriously it takes the audit side.

WTW plans to extend Violet to more procedures and clients across the Benefits Outsourcing business. If the reuse claim holds, each new workflow should cost less to add. If it doesn't, the Azure plumbing will still be there to support the next attempt.

 

References

  1. WTW builds an agent platform on Microsoft Agent Framework, saving tens of thousands of hours - Microsoft Microsoft Thu, 05 Nov 2026 08:00:00 GMT
  2. Willis Towers Watson Public Limited Company (WTW) Earnings Call Transcripts roic.ai
  3. Willis Towers Watson (WTW) Q3 2026 Earnings Call Transcript fool.com