BUILD

Kry AI:
Introducing generative AI to healthcare.

How it started

In 2022, generative artificial intelligence models became impressively creative, at times outcompeting humans. They also became widely accessible, affordable, and fast. The excitement was tangible across clinicians and healthcare operators alike. The R&D team at Kry began a thoughtful investigation into problems ripe for scalable, AI-powered solutions: opportunities to improve efficiency and personalization emerged all across our healthcare service. My role was structuring this quest into a design-led project yielding experiences that patients and clinicians love.

What we built

Kry’s digital platform allows clinicians to deliver patient-first, modern care experiences. For our patients, the first step when looking for healthcare or advice is the Kry app. Our clinicians use Kry Pro to communicate with patients and navigate them through their care journey. We built AI-powered tools to enhance these existing products so they provide even smoother access to healthcare.

Digital assistant

We conceived of a Digital Assistant that gathers case information while patients are waiting for a healthcare professional. Not only were 96% of patients wiling to speak to the digital assistant, their reviews remained as positive as before.

AI-assisted nurse chat

We created a better informed and streamlined way for our nurses to help incoming patients through digital chat. During the pilot, the tool helped reduce the number of manually sent questions, and reduced the average chat duration by 66%, which resulted in much shorter waiting times for patients.

Suggestions for clinicians

We integrated suggestions into clinician tools to help them complete their documentation faster, based on recent patient interactions. Large language models author medical note suggestion, letting clinicians focus on the patient during the meeting, instead of having to type. The amount of time saved wasn’t measurable during the pilot.

Design decisions

TONE OF VOICE

The Digital Assistant speaks to patients on behalf of the Kry brand. Chatting with it should always be a consistent and trustworthy experience. Generated language today can be easily mistaken for a human: we chose to keep the assistant truthful about its own roots, therefore we assigned it a limited role and didn’t personify it. To correctly represent our brand voice (accessible, caring, trustworthy), the AI model is prompted to follow a set of tone-of-voice guidelines:

1. The assistant can never talk about itself in first person, such as “I” or “me”.

2. Only one question per message.

3. Write simple sentences.

4. Use everyday words.

5. Do not use slang.

6. Do not use abbreviations.

7. Do not use emojis.

8. No AI-generated empathy.

9. Be kind and respectful.

AI–CLINICIAN HANDOFF

Once the digital assistant collects the patient’s symptoms, it hands the case over to the clinician, such as to a nurse via chat: the patient shows up on the nurse’s dashboard. To get to know the case quickly, nurses can read a brief case summary written by AI, while having access to the full chat history (questions asked by the digital assistant, and the patient’s answers). All patient information is gathered to help them give advice and make care decisions.

CONTEXTUAL SUGGESTIONS

Interactions with the patient, such conversations via chat or video, are captured and used to help clinicians write up documentation faster. Suggestions are made to be autonomous, contextual and specific. They are displayed proactively wherever the clinician is likely to perform a related task. Clinicians accept each suggestion manually, safeguarding the final version based on their expertise and memory.

Examples of suggestion types:

• Diagnosis codes

• Consultation note fields

• Clinical actions, such as referrals and prescriptions

• Pre-filled documents and certificates, such as sick notes

LIMITED MANUAL CONTROL

Our intention is that clinicians have little to no option to adjust the prompts and settings that produce questions and suggestions. AI-powered tools can only make clinical processes efficient if they take no extra human effort and time to control. Therefore all setup and tuning happens under the hood, based on our deep knowledge of what our users expect.

Design principles

I shaped our design principles to help us build healthcare tools powered by AI. I translated our Experience Principles into concrete and actionable steps to take when designing (with) intelligent technologies.

THOUGHTFULLY ASSIST. Help complete specific tasks very well.

The software anticipates clinician needs, their tools combining ease of discovery with the power of execution. AI shall be used to help perform specific tasks in line with what users are expected to produce.

NEED FOR SPEED. Decrease interaction cost.

The goal is to help users do their best work quickly. The interaction model has to be minimal to help people act fast when making decisions.

FLEXIBLE, NOT FIXED. Use AI wherever our clinicians and patients are.

There are no straight paths in patient care, and clinician tools need to adapt flexibly. AI tools have to integrate into the most commonly used tools, without requiring any extra effort to access.

SAFEGUARDED. Let people make the final call.

Designers are motivated by the same principle as our clinicians: help them safeguard patients in their care. AI-powered interfaces shall emphasize and depend on human accountability over the final outcome.

Reflections

Using emerging technologies like AI as design material makes the process unique. I understood early on that these AI technologies are developing too fast for our users’ imagination to keep up with what’s possible. The work was informed predominantly by tech experiments, intuition, and our pre-knowledge about user needs.

What made these concepts work was not original ideas but good execution. Large language models (LLMs) were finally able to produce output that met the high quality bar of our patients and clinicians.

Experimenting got easier. Testing just got harder.

The conversational nature of generative AI models made experimentation possible for people like designers and writers. We could experiment freely and imagine tools for use cases we understood best. But ensuring quality at scale turned out to be far more difficult, because manual testing isn’t enough to reproduce all chance variations.

Work bottom up.

Crafting future visions are tempting, but working with what’s reality today is the only way to build a functional product. Knowing possibilities and limitations intimately comes from hands-on practice through prototyping.

Your AI will need allies.

More so than ever, users have to stay vigilant about what an AI does, never becoming over-reliant on them, especially while the initial excitement it high. With safeguards in place, we can capture a machine’s efficiency and mix it with some good human judgement.

Footnotes

This project was a collaboration with ALEXANDRA HOLM, who wrote the tone-of-voice guidelines and designed the conversational content.

This work is an extension of existing Kry products, designed by and with others. JULIE KONDO wrote the Experience Principles that I reworked for AI. TEDDY DERKERT designed the video module. Kry’s Patient Design team designed the Kry patient app, which we adapted for AI.

I gave a talk at the Project A Knowledge Conference about our learnings and reflections from making large from language models work in healthcare.