How generative AI in healthcare is transforming care and product development

In the AMA’s 2026 survey of nearly 1,700 US doctors, 81% said they use AI professionally, up from 38% in 2023. Most of the tasks they named are generative ones:

• summarizing research and standards of care (39%);
• drafting discharge instructions, care plans or progress notes (30%);
• preparing replies to patient portal messages (19%).

For care, this changes how the workday unfolds. Documentation that physicians often finished after hours now starts as a draft waiting for review. Instead of writing from scratch, the doctor edits a patient’s portal question with a suggested reply.

For the teams building healthcare software, the shift goes just as deep. AI-generated drafts need an interface that shows where the information came from, a workflow that requires clinicians to review and approve the text, and an integration that saves the approved version to the EHR. At the same time, GenAI is changing how these products are researched, designed, coded, and tested.

Below, we look at both sides of that shift: what generative AI already does for care, and what it takes to build products around it that clinicians trust.

What is generative AI in healthcare?

Generative AI in healthcare is a type of artificial intelligence that creates new content, such as clinical notes, patient messages, record summaries, medical images, synthetic data, or candidate molecules, by learning patterns from large medical and general datasets. Where traditional healthcare AI classifies or predicts, generative AI produces drafts that people review and use. Its output can be far more useful, and it can also be fluent and wrong, which is why high-risk healthcare applications typically keep a human review step before generated content affects care or the patient record.

Healthcare products today operate on a few model families:

• Large language models (LLMs) handle text, including powering ambient scribes, chart summaries, and patient-facing assistants.
• Diffusion models generate and enhance images. They are mainly used in imaging research and for producing synthetic scans.
• Generative adversarial networks (GANs) and variational autoencoders (VAEs) are older generative methods, commonly used to create synthetic data and in early-stage drug design.
• Multimodal models combine text, images, and other signals in one system, so a single model can read a scan alongside the patient’s history.

{{banner}}

How generative AI is used in healthcare

Generative AI is opening up new possibilities across healthcare, from streamlining documentation and patient communication to supporting medical imaging, drug discovery, and synthetic data generation. Below, we explore some of the most relevant generative AI use cases in healthcare.

Clinical documentation and ambient AI scribes

Ambient AI scribes listen to a visit, with the patient’s consent, and turn the conversation into a structured clinical note for the physician to review and sign. Documentation is one of the biggest pain points in healthcare, so GenAI adoption starts here.

First, a phone or workstation documents the conversation between a patient and a doctor. The AI model separates clinical content from small talk, and a draft note then appears in the EHR tailored to the specialty’s documentation requirements. Instead of rebuilding the visit from memory in the evening, the physician corrects a draft while the details are still fresh.

The effect shows up at scale. Over 63 weeks, 7,260 physicians at The Permanente Medical Group used an ambient scribe in more than 2.5 million encounters, according to an analysis published in NEJM Catalyst. Together they saved the equivalent of 1,794 working days in a year. Nearly half of their patients said the doctor spent less time looking at the screen.

Medical coding, claims, and prior authorization

After every patient visit, diagnoses and procedures must be translated into standardized medical codes that match the clinical documentation. Insurance claims must follow each payer’s rules to avoid payment denials. And when a treatment, test, or medication requires prior authorization, staff need to gather supporting evidence from different parts of the patient’s medical record. Generative AI healthcare tools can help teams handle these tasks by reading the record and preparing:

• a draft authorization request or letter of medical necessity, with the supporting findings attached;
• suggested diagnosis and procedure codes for a certified coder to confirm;
• a list of gaps in the documentation that would likely lead to a denial.

Staff then review and submit instead of compiling everything from scratch. Faster submission alone doesn’t guarantee lower costs, though. The products that hold up in audits keep every suggestion traceable to the line in the chart it came from.

Patient communication and virtual assistants

Generative AI can also be used in patient communication. For care teams, it drafts replies to portal messages from the patient’s question and relevant chart context, so a nurse or physician doesn't write a response anymore, just edits it. On a patient's side, virtual assistants answer routine questions about medications, appointments, and preparation 24/7.

The hard part of AI chatbot development in healthcare is deciding when the answer is outside the assistant’s competency. A useful one recognizes symptoms that need a human and passes the conversation to staff with enough context to take over.

Medical imaging and diagnostic support

Generative AI can help radiologists analyze medical images, such as X-rays, by preparing draft reports and flagging potentially serious findings. Northwestern Medicine offers one of the clearest results so far. Its in-house generative model drafted reports for nearly 24,000 radiographs across 12 hospitals, according to a study in JAMA Network Open. Radiologists who started from those drafts documented studies 15.5% faster, and peer review found no difference in clinical accuracy. The system could also flag unexpected collapsed lungs so they were read sooner.

That pattern holds well beyond one health system. In imaging, the model prepares an initial report and highlights findings that may need urgent attention. The specialist reviews the images, checks the AI-generated report, and makes the final clinical assessment. This reduces the time spent writing reports from scratch while keeping medical decisions in the hands of clinicians.

Clinicians don’t want AI to read the scan for them. They want it to take the typing off their plate.

Personalized treatment plans

A patient’s history may span years of clinical notes, lab results, imaging, and genetic data, while relevant research and guidelines continue to change. Generative AI can bring this information together by summarizing the patient’s record and comparing it with current guidelines and research.

In oncology, for example, a model can prepare a case for a tumor board by pulling together the relevant history, genetic findings, and potentially relevant therapies or clinical trials, with sources for each recommendation. Specialists still make the final treatment decision, though.

Drug discovery and clinical research

Traditional drug discovery proceeds as follows: researchers find a biological target, design a molecule that acts on it, test it in the lab, and then move to human trials. Generative models can speed up the first two stages. They propose targets from biological data and generate candidate molecules with the desired properties, which narrows what chemists need to synthesize and test.

Rentosertib shows how far that path now reaches. Generative AI identified both its target and the molecule. In a phase 2a trial for idiopathic pulmonary fibrosis published in Nature Medicine in 2025, patients on the highest dose saw their lung function improve while the placebo group’s declined. Larger trials will decide its future, but it’s one of the first AI-designed drugs with results from a randomized trial.

Synthetic data generation

Synthetic data is artificially generated patient data (records, images, or signals) designed to reflect the statistical patterns of real data without directly copying individual patient records. Teams commonly use synthetic data for three purposes:

1. Building and testing software without exposing real patient information.
2. Filling gaps for rare conditions or underrepresented groups, so models see enough examples to learn from.
3. Sharing datasets between institutions when real records can’t legally leave the building.

One caution applies to all three. A generative model can memorize parts of what it was trained on, so synthetic datasets still need a privacy review before anyone treats them as safe.

Benefits of generative AI in healthcare

The value looks different for different members of the process: patients, clinicians, healthcare organizations, and product teams.

For patients

Value for patients comes up in three things that have little to do with the technology itself:

• Attention. When a visit no longer revolves around the keyboard, the conversation gets the doctor’s eyes back, and patients notice it.
• Clarity. Discharge instructions, test results, and care plans can be rewritten at the reading level and in the language each person needs.
• Access. Even if a patient asks a question late at night or between appointments, they can get an answer from an AI assistant.

For clinicians

Burnout in medicine has many causes, but clerical work is one of the few that technology can remove directly. A multi-site study published in JAMA Network Open in 2025 followed clinicians using ambient documentation. At Mass General Brigham, burnout fell by 21.2 percentage points after 84 days. Notably, clinicians who spend less mental energy on notes have more of it for diagnostic reasoning, difficult conversations, and complex cases.

For healthcare organizations

Generative AI can reduce the time teams spend on documentation, coding, prior authorization, and other repetitive tasks. That leaves more time for patient appointments, shortens waits, and lowers the number of claims returned for corrections.

When clinicians spend less time on paperwork, they have more time for patient care and less administrative work after hours. For organizations already struggling to hire and retain staff, lifting that burden can be part of the solution.

For healthcare product teams

For the teams building generative AI for healthcare, research synthesis, interface prototypes, test data, and code drafts that once took weeks can now take days. That means more ideas can be tested with clinicians before the budget is committed to one direction.

But faster development doesn’t automatically mean a better product. Clinical workflows, compliance requirements, and usability still need attention throughout healthcare software design and development.

How GenAI is changing healthcare product design

Building digital health interfaces has always meant working around the same challenges: clinicians have little time for interviews; every screen carries some clinical risk; the same content has to make sense to a specialist and to a patient reading it on a phone at night. Generative tools can help tackle some of them. Others remain a human responsibility.

Ways GenAI speeds up healthcare user research

To conduct reliable user research, researchers need to interview healthcare professionals, observe how they work in hospitals, and review support tickets. In the past, analyzing all this information took a lot of time and had to be done manually. Today, GenAI helps healthcare product teams process interview recordings, observation notes, and support tickets faster. It can transcribe conversations, group recurring problems, and draft profiles of typical users, such as ICU nurses or outpatient oncologists.

The output is a starting point, since models tend to amplify what is said often and miss what is said once. For example, ten comments about interface colors could outweigh one comment about a workaround nurses use on the night shift. Researchers therefore check the original recordings and notes behind each AI-generated theme before using it to make design decisions.

From ideas to prototypes with generative AI

With generative AI, teams can explore several layout directions, turn a promising concept into an interactive prototype, and populate it with realistic synthetic patient data much faster than before.

This speed changes how ideas get validated. Instead of debating one direction in meetings, teams put two or three versions in front of users and let their reactions decide. The designer’s role shifts toward checking that:

• layouts meet accessibility requirements (contrast, text size, touch targets);
• the sample data is clinically plausible;
• the flow matches how care actually happens on the floor.

Designing AI features clinicians can trust

A generative feature is only as valuable as clinicians’ willingness to rely on it. In healthcare AI UX design, that trust comes from a few concrete decisions:

• the source behind every suggestion is one click away;
• uncertainty is visible, so a weak answer looks weak;
• editing, accepting, or rejecting the output takes a single action;
• AI-generated content is clearly labeled as such everywhere it appears.

Keeping these interactions simple also matters for reducing cognitive load in ICU interface design, where clinicians need to make decisions quickly.

Making healthcare content clearer and more accessible

A discharge summary full of medical terminology can leave a patient unsure what to do next. With generative AI, that summary is tailored to the reader: a plain-language version for patients, a translated one for families, a concise one for a busy specialist.

Rewriting medical content still requires care. A simplified explanation might leave out an important dosage instruction, while a translation could change the meaning of a warning. Teams need to check these versions for clinical accuracy and test whether patients can follow the instructions.

Making complex medical information easier to navigate for different users was a key consideration in our work on an AI-assisted oncology platform. For a demo version presented to investors, the team mapped user journeys and designed a mobile experience for patients and a desktop workspace for clinicians. It then delivered a complete UI/UX system ready for further product development. The new structure made patient information easier for providers to see and remote guidance easier for patients to follow.

{{banner-2}}

Generative AI in healthcare software development

When it comes to engineering, digital health teams are expected to ship features quickly while ensuring every release meets the requirements of security reviews, audits, and clinical scrutiny. Today, code assistants and automated testing tools take on part of that load, but engineers still own what reaches production.

AI-assisted coding and code review

The role of AI in healthcare software development is mainly to take care of repetitive coding tasks. Developers can use it to generate boilerplate, map data between systems, write unit tests, and draft routine functions. This frees them up to focus on the harder parts of the work, such as system architecture, integrations, and making different healthcare systems work together.

How much faster this makes a team is still an open question. Independent trials have produced mixed results, and the gains depend heavily on the codebase and the developer’s experience. That uncertainty is one more reason for caution in generative AI software development in healthcare: AI-written code goes through the same human review as any other code. Reviewers pay extra attention to how it handles patient data and standards like HL7 and FHIR, which general-purpose models know less well.

Test generation and quality assurance

Medical software can fail in situations users rarely notice: an unusual date format in a lab result, a claim with a missing field, or a timeout during an EHR sync. Writing tests for every possible case by hand is slow, which means some scenarios may go untested.

Generative tools can speed up this work. They can read requirements and existing code, suggest test cases for edge conditions, and update test suites as features change. Engineers still decide what needs to be covered and review the generated checks, because automation cannot catch problems it was never designed to look for.

Synthetic data for healthcare software testing

Healthcare software teams need realistic data to test their products, but using real patient records during development can put sensitive information at risk. Testing with empty or overly simple records can also hide problems that might arise in real clinical settings.

Synthetic data offers an alternative: artificially generated records that mimic real-world data. Teams create and use these records in three steps:

1. They define the production data structure and the clinical patterns that need to be tested.
2. A generative model creates records that follow those rules, including rare combinations of clinical details.
3. The team uses the synthetic records in test environments.

This approach lets teams test performance, integrations, and interfaces under realistic conditions while reducing their reliance on real patient data during development.

Documentation and compliance artifacts

In regulated software, documentation is part of the product. Requirement specifications, traceability between requirements and tests, risk files, and release notes all have to stay in sync with the code, and they drift the moment a sprint gets busy. AI in healthcare software development drafts and updates these documents from commits and tickets, which keeps them current with far less manual effort. An engineer still signs off on every artifact before it goes to an auditor. Teams seeking this approach from day one can explore Halo Lab’s healthcare software development services.

Challenges and risks to plan for

Every strength of generative models has a shadow in clinical settings. Just like the fluency that makes a draft useful can also make an error convincing.

Why hallucinations are a clinical risk

Language models predict the most plausible next words, and plausible isn’t always accurate. When the prediction is accurate, everything goes smoothly; when it isn’t, the model fills a gap with a dose, a lab value, or a citation that sounds right but doesn’t exist. Such hallucinations can lead to something minor, like an embarrassing marketing email, or something as serious as the wrong treatment in a discharge summary.

A safe clinical AI system has several layers of control:

• Verified clinical sources: The model works from the patient’s chart, approved guidelines, and other trusted sources.
• Traceable clinical statements: AI-generated claims are tied to their sources, making it possible to check where the information came from.
• Clinician review: AI-generated content remains subject to clinician review before it is added to a patient record or used in a clinical decision.
• Defined limits: The system does not generate unsupported clinical recommendations when the necessary information is missing or uncertain.

Data privacy, security, and HIPAA compliance

Healthcare AI tools must protect patient information, especially when building a HIPAA-compliant AI portal. In practice, it looks like this:

• Managing third-party vendors: Putting a business associate agreement (BAA) in place with each AI vendor that handles PHI, where HIPAA requires it.
• Limiting data exposure: Sending only the minimum necessary information in each prompt and de-identifying data when the task allows it.
• Maintaining audit trails: Logging who sent data to the model, what information was submitted, and what the model returned.
• Preventing prompt injection: Protecting against malicious instructions designed to make the model reveal sensitive information or bypass its safeguards.

Bias in generative AI systems

Bias isn’t limited to humans; generative AI can reproduce or amplify biases present in its training data. In dermatology, for example, studies have found that some AI models are less accurate when assessing skin conditions in people with darker skin tones. Similar biases can occur when training data underrepresents certain age groups, languages, or communities.

These disparities may not be obvious in overall accuracy scores. To identify them, teams need to test performance across different patient groups, monitor results after launch, and include people from diverse backgrounds in design research.

FDA guidance and the EU AI Act issues

The regulatory risk rarely comes from the model itself. It comes from discovering late that the product counts as a medical device, which can mean months of extra validation, a redesign, or a delayed launch. In the US, software that helps diagnose or treat patients can require FDA clearance. Tools that let clinicians independently review the basis of a recommendation may fall outside that definition, a line the FDA restated in its clinical decision support guidance updated in January 2026. The regulatory requirements also depend on the product’s intended use and the type of data it analyzes.

Models that keep learning bring a second risk: every significant update could trigger a new review. The FDA’s final guidance on predetermined change control plans, issued in December 2024, addresses this. Manufacturers can describe planned updates in their submission and roll them out without starting over.

Products sold in Europe face an additional layer. The AI Act entered into force in August 2024 and classifies AI used in regulated medical devices as high-risk. Its requirements for risk management, data governance, human oversight, and technical documentation phase in over the coming years. Teams that account for these requirements early can reduce the risk of costly redesigns and delays further into development.

Integrating GenAI into EHRs and clinical workflows

A model that works perfectly in a demo can still fail on day one in a hospital. The reason is usually integration:

• the tool can’t read the right data from the EHR;
• it writes back in the wrong place;
• it adds a separate window clinicians have to switch to.

Standard interfaces such as HL7 FHIR make the data connection more predictable, but workflow fit takes observation.

How to implement generative AI in a healthcare product

Plenty of promising pilots stall somewhere between a convincing demo and daily use. The sections below illustrate the ways to incorporate GenAI into a healthcare product.

Finding the right problem to solve

The best first use case is rarely the most impressive one. Teams that get to production tend to pick a task with three qualities, in this order of importance:

1. It happens many times a day.
2. It frustrates people in a way they can describe.
3. A mistake can be caught before it reaches a patient.

Drafting notes, answering routine portal messages, and preparing prior authorization packets all fit that profile. Autonomous diagnosis does not.

If a pilot is gone, and nobody notices it, it’s not worth the investment.

Getting healthcare data ready

A model works with whatever it’s given to read. Medication lists may live in one system, allergies in another, and half the history in scanned PDFs. In that case, the output will be incomplete no matter which model runs underneath. Preparation comes first:

• mapping where the needed data lives;
• normalizing it to shared terminology;
• stripping identifiers from anything used for development or evaluation.

That groundwork often takes longer than building the AI feature itself.

Choosing the right model for your product

Among GenAI models, three options cover most cases:

• Large general-purpose models offered through enterprise agreements handle language well and adapt quickly. When handling patient data, they require appropriate security controls and a BAA with the vendor, where applicable.
• Medical models tuned on clinical text can perform better on specialized vocabulary.
• Smaller models hosted inside your own cloud trade some capability for full control over data and cost.

The choice depends on the task’s risk, the data’s sensitivity, and the volume you expect, and many products end up combining two of them. Whichever mix you choose, most of the effort goes into connecting the model to records, permissions, and audit trails, which is the core of our AI integration services.

Keeping clinicians in the loop

Human review is the core safety mechanism of almost every clinical AI feature. So, it has to be designed as carefully as the model. Reviewing should take less effort than writing from scratch; otherwise people skip it. What’s more, clinicians’ edits become valuable feedback, and the record should always show who approved the final version.

Testing before real-world deployment

Testing runs in widening circles:

1. Offline: the team tests the system on past cases with known outcomes, measures how often it gets them right, and analyzes recurring errors.
2. Shadow mode: the system runs on live data, but nobody sees its output except the evaluators.
3. Pilot: a small group of clinicians uses it in real work, with a clear way to report problems and switch the feature off.

Scaling and monitoring in production

After the launch, the team needs to ask different questions. Patient populations shift, documentation habits change, and vendors update their models, so performance that looked stable in the pilot can quietly drift. Teams that scale well do two things:

• they track output quality, edit rates, and clinician usage continuously;
• they review those numbers with clinical leads on a set schedule.

Even a promising pilot can run into problems with integration, adoption, and performance, which makes the transition from healthcare AI prototype to production particularly challenging.

Where GenAI in healthcare is heading

A year ago, most of the AI conversation focused on single-task assistants, like ambient scribes, inbox drafts, or summaries. The future of GenAI in healthcare centers around agents that can complete a workflow from start to finish.

Take prior authorization: instead of drafting a request and stopping, an agent can gather records, submit the request, and track the payer’s response. It only returns to staff when a decision is needed. That also changes the design challenge: teams need to decide where human checkpoints belong and how staff can see what the agent has already done.

Patients are also bringing AI into their care. A KFF poll from March 2026 found that 32% of US adults had used AI for health information or advice in the past year. Yet only about six in ten who used it for physical health questions later followed up with a doctor. Patient-facing products will increasingly compete with chatbots patients already use. They need to offer something those tools can’t: a clear, quick path to the actual care team.

Models are also becoming multimodal, combining text, images, and other signals. A single feature can weigh a scan, lab trend, and clinician note together. This makes it even more important for interfaces to show which inputs shaped an answer.

Finally, governance is becoming part of the product itself. Health systems are formalizing AI oversight, so buyers increasingly look for monitoring dashboards, audit trails, and clear accountability for AI outputs. Products that include these from the start will be easier to evaluate and approve.

Building healthcare products with generative AI

GenAI can already change familiar healthcare workflows. But faster note-taking or automatically generated responses don’t necessarily make a clinician’s job easier. If the output is difficult to verify or the new tool doesn’t fit into existing workflows, some of the time saved simply gets lost.

That’s why developing healthcare products with GenAI takes much more than configuring a model. Thoughtful design helps clinicians work with AI-generated outputs, reliable engineering ensures the tool integrates with other systems, and clinical validation shows whether it performs well outside a test environment. The goal is to create a tool that makes everyday work easier, not one that adds another complicated process to it.

{{banner-3}}