65% Isn’t Good Enough: The AI Gap Healthcare Can’t Ignore

When it comes to AI, large language models (LLMs) are incredible at sounding right. They’re trained on vast amounts of text to generate human-like language, but they’re designed to predict likely answers, not guarantee correct ones.

It’s simple: you ask a question, and they produce fluent, confident answers in seconds. For many use cases, that’s more than enough.

Chatbots, search, summarization, brainstorming, drafting emails — these are all areas where LLMs shine. They’re incredibly useful when speed and fluency matter more than precision. But those same strengths create risk in healthcare, where consistency and correctness matter more than creativity.

In healthcare, “mostly right” isn’t good enough.

When AI is used to analyze clinical conversations, patient journeys, or safety signals, a system that’s right about 65% of the time isn’t progress. It’s a coin flip.

And that’s the uncomfortable truth: in many real-world scenarios, general-purpose language models land in the 60–70% reliability range on complex interpretation tasks.

The problem isn’t that the models are bad. They were never designed for deterministic judgment — to ensure the same input always produces the same output.

Healthcare requires a different approach.

The Problem With Measuring AI

The ground truth is this: before you can improve AI accuracy, you have to measure it correctly. Yet many systems rely on standardized benchmarks to evaluate performance.

These benchmarks are static tests. As models improve, they get better at scoring well on them.

But there’s a catch:

Over time, models learn the benchmark itself, not the real-world problem. They optimize for the test instead of the task.

The result? Scores go up, but real-world reliability doesn’t.

The AI looks like it’s getting smarter, but it’s really just learning the exam.

Healthcare doesn’t operate in benchmarks. It operates in messy, ambiguous human conversations.

At Authenticx, we approach this differently. Instead of relying on static benchmarks, we built our evaluation process around a continuously evolving “golden dataset” of human-labeled healthcare interactions — conversations evaluated by domain experts who understand the nuance of the industry.

Because in healthcare, the hardest question isn’t “what did the AI say?”

It’s “what does correct actually look like?”

The Hardest Problem: Deciding What’s Right

Consider a simple example: a patient says: “I love your product, but I have one complaint.”

Is that positive sentiment or negative sentiment?

“A general-purpose AI might label the same conversation differently every time you ask it. One time it’s positive, the next time it’s negative,” says Authenticx Chief Technology Officer, Michael Armstrong. “That’s not a bug — it’s how large language models work. They’re probabilistic systems designed to generate the most likely answer in the moment. That’s incredibly powerful for language tasks, but it becomes a real problem when you need consistent, defensible answers.”

Because in healthcare analytics, the same input should produce the same judgment every time.

Turning AI Upside Down

Most AI systems work like this: Prompt → Model prediction → Answer

At their core, AI models are designed to make judgments or generate responses at scale.

Traditional AI systems produce the most likely answer. We inverted that process.

Instead of asking AI to guess, we require it to evaluate the problem step by step using structured criteria. Those criteria take the form of rubrics: explicit frameworks that define how a decision should be made.

Take the sentiment example above. Instead of asking, “Is this sentiment positive or negative?”, our system evaluates the conversation against a rubric:

• What signals of satisfaction are present?

• Are there explicit complaints?

• How strong are the positive versus negative indicators?

• How should mixed signals be interpreted?

Our model doesn’t guess. It scores the conversation against defined criteria and produces a deterministic result.

This reduces variability and aligns results more closely with human judgment. In other words, we didn’t try to make AI guess better. We redesigned how it reasons.

Why Rubrics Matter

Rubrics solve two critical problems in healthcare AI.

First, they create deterministic decision-making. The same conversation evaluated under the same rubric  is far more likely to produce consistent results.

Second, they allow organizations to encode domain expertise directly into the model’s reasoning process.

“In healthcare conversations, nuance matters,” says Michael. "The difference between a potential regulatory complaint, a pharmacovigilance signal, or just a casual comment can be incredibly subtle — and getting that wrong can have serious consequences. That’s why the technology alone isn’t enough. The expertise behind the rubrics matters just as much as the model itself.”

Our AI is doing far more than just processing language.

Measuring Beyond Random Chance

Accuracy alone doesn’t tell the full story. What matters is whether AI systems agree with human judgment.

In machine learning, this is often measured using Cohen’s Kappa, a metric that evaluates how much agreement exists between two evaluators beyond what would occur by random chance.

Why does this matter?

Because a model can achieve 65% accuracy on a task where random guessing would produce similar results. In those cases, the system isn’t truly understanding the problem. It’s just approximating it.

Healthcare AI systems need to achieve high agreement with human experts, not just higher benchmark scores. That starts with training models on the right data.

Training AI on Healthcare Reality

To achieve this level of accuracy, we experimented with multiple model architectures.

In one internal test, we compared a large general-purpose language model — a 70-billion-parameter model — against a much smaller open-source model.

Then we added something different.

We trained the smaller model using a LoRA adapter, a technique that adds a small number of specialized parameters to an existing model.

The base model contained about 7 billion parameters. The LoRA layer added roughly 70 million parameters — a fraction of the original model size — but those parameters were trained on our golden dataset of human-labeled healthcare conversations.

The result was dramatic.

In internal evaluations against expert human labeling, we were able to achieve : 99% accuracy after fine-tuning.

The real power of AI comes from combining foundation models with domain-specific training and structured evaluation. It’s not about making the model bigger. It’s about making it smarter about the things that actually matter.

The Last 35% of AI

The gap between 65% and 99% isn’t just incremental. It’s an entirely different way of thinking about AI.

Most systems stop when the model is “good enough," but in healthcare, “good enough” is failure.

Closing that gap requires more than better models. It requires:

• Real-world evaluation

• Human expertise

• Structured reasoning

• Deterministic systems

This is the difference between AI that sounds right and AI that is right.

AI Built for Structured Reasoning, Not Guesswork

We all agree that LLMs are incredibly useful tools. They can summarize, draft, and explore ideas at extraordinary speed. They’re changing how people interact with information.

But when accuracy, compliance, and safety matter, usefulness isn’t enough.

That requires systems designed not just to generate answers, but to prove them.

Because in healthcare, the difference between 65% and 100% isn’t marginal.

It’s everything. Because your patients depend on it.

Download Healthcare AI Oversight Checklist: 5 Questions Every Leader Should Ask to see what it takes to move from “mostly right” to truly reliable.