Inside MealSuite’s race to govern AI and take control of their data

The challenge: The need to protect employees, not just customers

“There is no playbook,” Xu told CTO Craig Wincey on her first day.

MealSuite handles data for healthcare-adjacent customers, such as PII and PHI, that senior living and skilled nursing operators are legally required to protect. But knowledge workers were using AI to connect directly to MCPs without access permissions, policies, or the ability to see what data agents were accessing.

“The tools [MCPs] are coming very rapidly and they’re ahead of what companies have thought about in terms of strategy. We had agents that had access to data, and in an industry like ours that deals with healthcare data, we have to protect the PII, PHI, and ensure it’s not getting leaked, accidentally or on purpose,” said Wincey.

Xu felt the weight of what she describes as “the fairness problem” when it comes to AI usage. Telling employees to simply “use AI” and police their own access to sensitive data is unrealistic and unfair.

“As a user, how can you even know, without checking all the data?” said Xu. “If a company hands AI to employees and holds them accountable, the companies themselves first have to centralize governance. They can’t expect employees to just take guesses.”

For Xu, governance was also personal: she didn’t want AI models knowing or accessing her personal data either. MealSuite needed one place to centralize that control, protecting both employees and customers.

The solution: Comparing the field

Wincey scanned the market, building a competitive comparison of gateway vendors, and came up with a shortlist to evaluate and test.

“It came across that Barndoor was not just meeting a lot of what everybody else was doing, but also leading in a number of different areas,” said Wincey.

Through her evaluation process, Xu identified Barndoor’s simple and intuitive user interface as a deciding factor. “For some platforms, the user interface wasn’t straightforward. There were many steps to get to the action you wanted,” she said.

As a result, the decision came quickly. “We couldn’t waste time. We needed to move fast to protect customer and employee data,” said Xu.

Outcomes: Meeting the dynamic and human demands of AI usage

While evaluating vendors, Xu took into account the change management aspect of onboarding employees to a governance platform.

“One of the problems with a complicated platform is it makes it difficult to manage. This means slower onboarding and rollout,” said Xu. “Business is dynamic and Barndoor is easy to manage, flexible, and helps me keep up with whatever new MCP my team needs.”

Productivity gains

Wincey built an internal dashboard pulling Jira data through the LLM gateway. Here, he’s tracking sprint output before and after MealSuite’s AI rollout. Engineering and QA are now completing about 200% more development output per sprint than before.

Better security

Xu’s team has rolled out PII and prompt-injection policies that are on track to block more than 1,500 risky tool calls a year. These are agent actions that are stopped before they reach and share customer data in tools like Salesforce and Gong.

“We applied a prompt injection policy and a PII policy against the servers that contain customer information. We don’t want that information to be exposed to our agents and we want to ensure the customer data is protected,” said Xu.

“We communicated these policies to the relevant teams so they are aware that if they see something redacted or masked, it’s because of this policy.”

Cost visibility

For MealSuite, the team is now able to understand how each team is spending on LLM usage, manage the complexities of how each model provider charges behind the scenes, and take control over how different models behave.

“For these models, their pricing strategies are so complicated — there’s about 8 different types of pricing, and if we as a company aren’t tracking the right data, how can we even establish our own strategy?” said Xu.

“With Barndoor, I’m getting a better understanding of the cost of how we’re doing a lot of this work. I can see what my engineering group costs me versus my marketing group versus my sales group,” said Wincey. “When I start to see those costs starting to rise within certain groups, I can go to them and ask the right questions to the right people, rather than just seeing our entire AI spend going up month-over-month.”

User adoption

Days after implementing Barndoor, MealSuite deployed Salesforce, the MCP most requested across departments. Today, the team has onboarded 32 departments on Barndoor and is now able to retire direct MCP connections entirely. About two-thirds of the MealSuite workforce uses AI tools day-to-day, with training during the rollout so employees understand why governance decisions were made. The team expects to reach nearly 100% AI adoption by the end of the year.

Advice to others on governance: “Don’t wait”

Wincey’s advice to other companies considering AI governance is clear. “Don’t wait,” he said. “The industry is moving too fast for any company not to be governing how their users apply AI. When it goes wrong, they’re going to be a front-page story,” said Wincey. “Barndoor has enabled us to have the visibility that allows us to utilize our data as best we can to continue to advance our business.”

For Xu, her advice is to make sure your people are bought in. “If you want to take the people with you at the right pace, in the right way, investing in this kind of platform is absolutely foundational,” she said. “You have to give every employee, not just the early adopters, a safe way to build AI skills instead of being left on their own to catch up. Barndoor allows us to feel we can sleep better.”

What’s next: Governed AI automations

With AI governance established, MealSuite is now looking toward deterministic data retrieval, structured outputs, and governance through Barndoor. As a first use case, the team is exploring a weekly cost and operations dashboard, which will point toward a larger opportunity in preparing enterprise data for model tuning.

Watch the full interview