Build vs. Buy: What health plans need to know before building AI in-house

Health plan leaders don’t need to be convinced that member retention is hard. You live it every day. Between growing competition, complex benefit structures, and a multilingual, multi-generational member base, keeping members informed and engaged is a high-stakes challenge. And that’s exactly why we built Mia.

Let me tell you, building something similar in-house is far harder than it looks from the outside.  
Mia isn’t just another chatbot. It’s a voice-powered, multilingual, multi-modal, seniors-focused, context-aware, Medicare-smart assistant designed specifically to reduce churn, improve satisfaction, and build lasting trust with your members.

And here’s the intent that guided every decision we made – to simplify Medicare for seniors who are often overwhelmed and underserved, and to help health plans serve them with clarity, empathy, and consistency.

Because when members understand their benefits, they stay. When they get the right answer the first time, they trust. And when they feel supported, your plan stands out for all the right reasons.

Let’s pull back the curtain.

How much work is it to go from PoC to production?

Many teams begin with good intentions – stand up a PoC, use popular LLM, and imagine layering it into existing workflows. But here’s an obvious catch: building a proof of concept ≠ building a system ≠ scaling an enterprise-grade platform.

Building a conversational UX that accurately interprets Medicare jargon, multilingual voice inputs, and complex documents isn’t solved by standard OS libraries out-of-the-box.

Most Medicare data is scattered across documents and formats. The relationships between datasets located in different documents are complex. To give a simple example, the formulary document that lists the drugs covered by the plan typically shows only the tiers’ information while what each individual tier means, and their associated costs are mentioned in the Evidence of Coverage document. There are many such nuances which make retrieving the right details complicated. The challenge is creating parsing and retrieval logic that works across all of them consistently. This required deep engineering + domain expertise, not patchwork solutions that only performed on specific use cases.  

The Medicare space is inherently complex – shaped by decades of regulation, shifting plan structures, and deeply layered documentation. To build a system that works, and evolves responsibly, we had to build modular components and specialized agents, each designed for specific query types. This ensured responses aren’t just technically correct, but context-aware and member-friendly. Medicare documents are full of jargons. Simplifying them with a tone that shows authority as well as empathy is a fine art that blends the technical aspects and goes into human psychology. We closely worked with domain experts to understand the nuances and hit the right balance between accuracy and empathy.  

Will voice actually work for seniors in real conditions?

Most voice bots sound great until real-world users try them.

Real-time voice detection – when to listen, when to talk sounds simple. But latency, interruptions, and device-level permissions make it unstable without careful design.

Add the complexity of older adults speaking in Spanish, Mandarin, or Vietnamese, often with mixed language phrases, and the problem compounds. Then layer in low bandwidth environments, outdated smartphones, and high-friction UIs.

Think about a scenario where a member wants to ask whether a particular drug (example- Mounjaro) is covered by the plan. Drug names are very complex. When we add multiple accents and the ability of seniors to pronounce such complicated scientific terms, it is extremely difficult to predict the correct drug names even when the members themselves might be making mistakes. We spent multiple weeks figuring this out so that you don’t have to. For example, even if the member says Manjaro, Munjuro, and so on, Mia predicts the right drug and confirms the drug name in its response. We had to train Mia on such cases to ensure that it acts like a human and just doesn’t give back a fallback response.  

Same goes with understanding the context of conversation. Answering individual questions was comparatively easier but keeping track of all intents in a single conversation so that the conversation feels human-like is extremely challenging. Healthcare is like an onion. Members start with one question but that leads them to some other area that they are unsure about. A single conversation may span across benefits, general Medicare terms, copayments, reimbursement issues, drug coverages, provider/doctor specific questions, and so on. If the platform cannot hold the conversation across all intents, it fails in real-life situations. We have went through these cycles and built a robust system that works across multiple real-life scenarios in production.

Now, let’s add a multilingual aspect to it. The speech recognition, translations, LLM and text-to-speech, all steps have to work across languages. The answer needs to be making sense even with Medicare and medical terms. That’s hard and takes multiple iterations to get it right. It takes not only deep technical expertise but also a union of linguistics and technology to maintain a tone and persona that feels right. Remember, we’re building for an audience that doesn’t trust technology by-default. They need to feel comforted, secured and cared for to overcome the apprehensions of technology.  

We had to fine-tune flows, rethink interaction patterns, re-engineer our prompts, build Agentic frameworks and stress-test every step to ensure the experience made sense to our audience. And all of this holds up during high-traffic hours and weak connections.‍

‍

Let’s look beyond the system.‍

Let’s talk about an actual moment. Maya, a 76-year-old member in Fresno, recently received a dense plan document in the mail. Her daughter was out of town. She felt overwhelmed. So, she opened the Mia app scanned the letter, and asked:

“Is my plan changing next month?”

Mia read the letter, identified the plan year, cross-referenced her benefits, and replied in Spanish with a clear, personalized answer.

Without any hold-times or confusion. Just clarity.

Then she needed to search for a new dentist since her current one had moved cities. Instead of going through multiple phone calls or websites, she just asked Mia – Are there any dentists around my area, and she got s list of dentists that spoke Spanish. The contact details, phone numbers, hours of operations, everything was right there in a moment. That’s a delight! That’s what converts into retention and true care that seniors deserve.  

This is the result of systemic design, advanced NLP, document parsing, and language layering that works without fail. When your platform becomes the moment of clarity for a member like Maya, you’ve built something worth scaling. Again, we’ve already built it so that you don’t have to.

How does it perform when the network doesn't cooperate?

Health plan members often interact with Mia in environments that are far from controlled. They may be on older devices, speaking with regional accents, or using their phone in areas with limited connectivity.

Mia is engineered to handle these real-world scenarios gracefully. That includes:

• Variability in device hardware and OS behavior
• Background noise and overlapping speech
• Low-bandwidth mobile connections
• Interruptions or mid-sentence corrections
• Diverse accents and multilingual inputs

We didn’t assume the ideal – we built for the actual. Every aspect of Mia’s architecture, from voice detection to response timing, is optimized to ensure reliability, clarity, and speed across diverse environments.

Because for your members, consistency isn't a nice-to-have. It’s the baseline for trust.

‍

‍How do members get timely updates without manual effort?‍

Building AI that can answer questions is one thing. But members don’t just wait to ask, they need to be proactively informed.

That’s why Mia includes voice-based notifications, delivered in the member’s preferred language, through the channel they’re most comfortable with. Whether it’s a reminder about an upcoming preventive visit, a notice about changes in drug coverage, or a nudge to review their benefits, Mia delivers it in a way seniors actually understand and act on.

Designing this required solving multiple layers of complexity:
‍
• Managing consent and opt-in preferences across regulatory frameworks
• Ensuring notifications adapt to regional languages and accents
• Handling edge cases where members use older phones, low bandwidth
• Structuring the message so it feels supportive, not robotic

This isn’t just an “add-on feature.” It’s a system of trust- because when members receive timely, clear, and compassionate reminders, they feel cared for, not left behind.

‍

‍How do plans know what’s working and what’s not?‍

Member interactions aren’t just about resolving queries. They’re a goldmine of insight for plan leaders. What are members most confused about? Which benefits drive the highest call volume? Where are conversations breaking down?

To make this visible, we built a business user portal that shows the exact questions being asked, usage patterns across demographics and languages, and emerging themes that drive member engagement.

Building this isn’t trivial. It requires:
‍
• Designing analytics pipelines that process conversational data in real time without exposing PHI
• Normalizing thousands of interactions across multiple languages, accents, and query types
• Creating dashboards that non-technical users can explore without needing data science teams
• Capturing trends that inform not just call center operations, but benefit design and outreach strategies

Without this layer, AI is just a black box. With it, health plans gain actionable intelligence that improves member experience, reduces confusion, and drives retention. And just like conversational AI, this analytics capability took us months of engineering and compliance work to get right- work that adds even more time and cost if plans attempt to build in-house.

‍

‍Is control worth slower iteration?

Short answer – yes. In regulated environments, stability > speed. The AI ecosystem is moving fast. New releases every week. New features. New APIs. But we’ve stayed clear of shipping anything to production that relies heavily on backward incompatible open-source frameworks. Why? When backward compatibility breaks, Stability falters. Trust erodes.

We pull architectural inspiration from OS tools like LangChain but all business logic is developed in-house. That’s how we stay aligned with our compliance stack, maintain full control over inference behavior, and ensure AI outputs don’t shift unexpectedly after a third-party update.

This approach is slower. It demands a senior team. But it also ensures healthcare-grade resilience – and that’s non-negotiable.

‍

‍How accurate is “accurate enough”?

In a regulated space like Medicare Advantage, accuracy isn’t a performance metric, it’s a trust requirement. It’s one thing for an AI to answer general questions with moderate accuracy. It’s something else entirely to handle plan-specific queries, across multiple languages and formats, with the consistency members and compliance teams expect.  

That’s why we’ve spent months refining:

• Edge cases that damage trust
• Ambiguity in document formats
• Repeat member usage patterns
• Intents that shift mid-conversation
• Guardrails that don’t break the conversation but still avoid compliance issues

Our target wasn’t "good enough." It was sustained, measurable, real-world accuracy- 90% and above, across live member interactions. With the combination of test automation and Human-in-the-loop testing, we have solved the problem of ensuring that each AI change, big or small, goes through a robust test framework that mitigates the risk introduced by predictive nature of Gen-AI. We have built-in guardrails that are battle-tested internally to ensure Mia doesn’t leak any confidential information.

‍

‍What happens when regulations change, or consent must be captured live?

For any health plan, compliance is a moving target that needs to be embedded in every layer of the system.

Mia was built from day one to meet the highest standards in healthcare data governance, not retrofitted after launch.

• Every interaction is logged, encrypted, and traceable.
• Member inputs, AI decisions, and outputs are fully auditable.
• We support HIPAA requirements, PHI handling, and enterprise-grade access controls by design.
• We are SOC 2 Type II certified, with processes reviewed across engineering, operations, and product teams.

Commercial LLMs can drift, guess, or hallucinate. That’s not acceptable in healthcare. Mia includes multiple layers of "trust logic"- checks that validate output against plan documents, reject ambiguous answers, and prioritize factual, Medicare-safe responses.

This isn’t just a tech safeguard. It’s how we ensure accuracy, repeatability, and audit-readiness at every step. When members ask about coverage, medications, or plan comparisons, we handle real-time intent as part of the interaction. Voice or text, English or Spanish, intent is requested, recorded, and respected without derailing the user experience.

Because when members share sensitive health information, they deserve clarity, control, and protection every time.

‍

‍Why health plans trust Mia over building in-house?

When buyers ask, “Why not just build this ourselves?” -here’s what we show them.

Mia isn’t a plug-and-play bot. It’s an enterprise-grade, HIPAA compliant, Medicare-ready platform built with purpose, precision, and patience.

For most health plans, building something similar in-house requires more than just a budget. It requires:

• Deep experience working with large language models, RAG systems, and AI infrastructure
• Multilingual voice UX expertise, especially for older adults in low-bandwidth environments
• Document understanding pipelines that can parse and extract accurate data from inconsistent, merged, or non-standard Medicare plan formats
• A compliance architecture built for HIPAA, PHI, and real-time consent workflows from day one
• Continuous iteration to maintain 90%+ intent accuracy across thousands of interactions
• The engineering maturity to handle edge cases, device-level inconsistencies, and regulatory changes – without breaking core functionality
• Multiple iterations to get the fine-tuning right that doesn’t introduce regression issues
• And the most important resource – time

That’s a multi-year commitment.

We’ve already invested those years. We’ve already solved the hard problems. And we continue to evolve Mia with each deployment, so our partners don’t need to worry about platform stability, compliance updates, or AI model regressions.

In short, buying Mia means:

• Faster time to impact
• Lower total cost of ownership
• Fewer support calls from confused members
• Less strain on internal tech, member services, and compliance teams
• A consistent, scalable, senior-friendly experience – without starting from scratch

When we get it right:

• Members feel confident
• Agents save time
• Plans reduce churn and earn loyalty

That’s why leading health plans don’t try to replicate what we’ve built. They choose to partner with us.

‍

‍Conclusion

We underestimated latency until it broke member conversations, we thought RAG was a solved problem until we didn’t get the right chunks, we built Agentic workflows to handle complex scenarios, we patched multilingual flows until we realized accents and code-switching demanded a different architecture. We saw compliance gaps appear the moment regulations shifted. Every failure taught us that building Conversational UX for Medicare isn’t about “just connecting an LLM and writing some prompts”. It’s about months of engineering discipline layered with healthcare-specific context.

‍