Baxter International: verified AI for 5,000 employees in 6 months
Baxter International runs its own proprietary AI alongside Claude and Copilot. After a 3-month vendor evaluation, Baxter chose VarsaAI as the trust and verification layer across that stack. The integration went live in under 3 weeks. Within 6 months it covered 5,000 employees, and it caught 174 fabricated claims before they could be relied on.
Results at a glance
- Under 3 weeks
- to integrate the VarsaAI MCP across Baxter's AI workflow, end to end
- 5,000 employees
- covered at full deployment, in month 6
- 174
- fabricated claims caught
- 908
- claims flagged for human review
- 1,631
- sources referenced and checked
- 479
- AI-cited sources (29%) that could not be traced back to the first author or original work
- Customer
- Baxter International
- Teams
- IT in North America and Europe
- AI stack
- Baxter proprietary AI, Anthropic Claude, Microsoft Copilot
- Product
- VarsaAI Verification MCP (VeriCore™, VeriRef™, Receipts, Claims Library)
The challenge: AI everywhere, proof nowhere
Baxter has been an early adopter of AI. It uses specialised generative AI and machine learning models to drive innovation and process efficiency across the organisation. Its stack combines Baxter's own proprietary AI with Anthropic Claude and Microsoft Copilot.
More models meant more AI-generated content, and every model could produce a claim that sounded right but wasn't. Baxter's existing tooling could log every prompt and response. It couldn't say whether those responses were true.
Baxter needed one trust and verification layer that worked across every model in its stack, not a separate fix for each one.
A rigorous choice. Baxter ran an extensive 3-month vendor evaluation before selecting VarsaAI to provide that layer.
The integration: live in under 3 weeks
VarsaAI worked with Baxter's IT teams in North America and Europe to plug its MCP into Baxter's AI workflow.
“Typically, integrating a third-party software into our technical estate is a time-consuming process, implemented across multiple phases. I was surprised at how easy and fast it was to set VarsaAI's MCP up and have it working across our workflow. I think end to end it took less than 3 weeks.”
Technical configuration
VarsaAI connects to Baxter's AI stack as a native MCP integration, using Streamable HTTP secured with OAuth 2.1. It plugs in at the MCP integration layer, behind Baxter's existing secure AI gateway, so it inherits the single sign-on, role-based access and PII redaction already in place. Nothing changed for users: they keep working in Baxter's proprietary AI, Claude and Copilot. As each model generates a response, VeriCore™ gives a verdict on every claim. VeriRef™ attributes each sentence to its source, and each verified claim gets a receipt that third parties can check independently. Every verified claim and source flows into an organisation-wide claims library, giving IT and compliance one view of what Baxter's AI has said and the evidence behind it, across every model.
Baxter's logging already recorded every response; the VarsaAI layer adds verdicts, sources and proof to each one.
Staged deployment: 300 to 5,000 employees in 6 months
Baxter rolled VarsaAI out in three stages, proving value with a first cohort before scaling to the full workforce.
| Stage | Employees covered |
|---|---|
| First 6 weeks | 300 |
| Month 3 | 1,100 |
| Month 6 (full deployment) | 5,000 |
Because verification runs inside the AI tools employees already use, each stage meant switching more users on. There was no new tool to train them on.
174 fabricated claims stopped
VarsaAI caught 174 fabricated claims in AI output: statements with no supporting evidence that would otherwise have gone straight into Baxter's work. A further 908 claims were flagged for human review, so experts spent their time where judgement was actually needed.
29% of AI-cited sources weren't original
VarsaAI checked 1,631 sources referenced in AI output.
| Source check | Sources | Share |
|---|---|---|
| Original, traced to the first author | 1,152 | 71% |
| Not original: could not be traced back to the first author | 479 | 29% |
| Total referenced | 1,631 | 100% |
Those 479 sources looked like citations but led to secondary or derivative material, not the original work. Without sentence-level source attribution, they would have passed as evidence.
Same AI stack. Now every answer carries proof.
Baxter didn't replace a single model. Its people use the same proprietary AI, Claude and Copilot, through the same sign-on and the same gateway. What changed is that every response now comes with verdicts, sources and proof.
Put the same layer under your AI stack
See how fast VarsaAI's MCP plugs into your models, and what it finds in your AI output.