Agentic AI UX for chronic care is the design of how an AI system acts autonomously, explains its actions, and keeps people in control. Unlike reminder-based health apps, agentic AI can pursue defined goals across multiple steps and adapt to changing context over weeks or months. That ranges from supporting medication adherence to coordinating approved care tasks.
More than an interface challenge, that makes UX a safety, trust, and autonomy challenge. A reminder can be ignored, and an agent that flags a concerning trend or escalates to a caregiver needs clear boundaries and easy human override.
At Onething Design, we approach this through Agentic Experience Design (AXD), designing agentic systems around human needs, behaviours, and context. Let’s understand the UX practices that matter most when designing agentic AI for chronic care and the principles that can extend to other high-stakes experiences.
What “Agentic AI” Means in Chronic Care and What It Means for UX
Agentic AI UX in chronic care is the design of how an AI system perceives context, takes action toward a defined care goal, communicates what it is doing, and keeps patients and care teams in control. But agentic does not mean unrestricted autonomy. In healthcare, UX has to define the boundaries of that autonomy:
- What the agent can do independently
- When it needs confirmation
- What it should explain
- Who needs to be involved
- How a person can intervene or take control back
A 2026 scoping review found only nine eligible studies of explainable agentic AI in healthcare, with the evidence base still largely experimental or retrospective. Meanwhile, Deloitte’s research shows that 61% of surveyed healthcare leaders were already building, implementing, or budgeting for agentic AI, and 85% expected to increase investment over the next two to three years.
The design problem is therefore a bit different, as we are no longer designing only what the user does. Rather, we are designing what the system is allowed to do.
We approach this through Agentic Experience Design (AXD). This involves designing the behaviours, boundaries, handoffs, and human controls around agentic systems. For chronic care, that means designing experiences that are trustworthy, appropriately autonomous, and human-led while being intelligent.
Levels of Agency in Chronic Care
Agency is best understood task by task. Here’s a practical UX lens for assessing increasing levels of system autonomy and the corresponding level of human oversight required.
| Level |
System agency |
Chronic-care example |
UX must provide |
| 1. Advisory |
Suggests |
Flags a rising BP trend |
Data, rationale, dismiss/review |
| 2. Supervised action |
Prepares |
Drafts a medication query for a nurse |
Preview, edit, approve/decline |
| 3. Bounded autonomy |
Acts within limits |
Adjusts a parameter within a clinician-defined range |
Clear boundaries, action log, override |
| 4. Higher autonomy |
Acts and escalates |
Monitors data and triggers predefined escalation workflows |
Scope, monitoring, escalation, audit trail |
Agentic AI UX vs Chatbots, Rule-based Apps and Portals: What are the Differences?
While portals give users access, rule-based apps follow predefined logic. Chatbots respond to requests, and then there are agentic systems that can proactively plan and execute approved actions.
| Parameter |
Patient portal |
Rule-based app |
Health chatbot |
Agentic system |
| Initiates |
Patient |
System, on schedule |
Usually patient |
Patient or system |
| What it does |
Provides information & services |
Follows predefined rules |
Responds to requests |
Plans & executes approved tasks |
| User role |
Review & act |
Follow & respond |
Ask & evaluate |
Review, correct & intervene |
| UX challenge |
Findability |
Relevance |
Trust |
Autonomy & control |
| Failure mode |
Abandonment |
Ignored prompt |
Wrong answer |
Unintended action |
Why Chronic Care Demands a New Approach to Agentic AI UX
Agentic AI UX for chronic care needs to be designed for continuity, changing context, and human control. Chronic care unfolds over months and years, involving patients, caregivers, clinicians, and multiple systems. So, the fundamental UX challenge is different from designing a one-time digital health interaction.
1. The Design Unit is a Journey
Most digital products optimise for what happens in the next few minutes. Chronic care does not work that way. A patient’s condition, routines, priorities, and ability to manage care can change over time.
This means designing for states:
- Newly diagnosed
- Stable
- Deteriorating
- Recovering
- Disengaged
- Requiring escalation
The agent’s behaviour should adapt accordingly. An intervention that is useful early in a care journey may become intrusive later. Good agentic UX therefore defines when the system should act, when it should wait, and when it should hand control back to a person.
2. The Real Opportunity is Reducing Care Workload
Chronic care creates ongoing management work. This includes tracking symptoms, remembering medications, managing appointments, interpreting information, and coordinating with caregivers or clinicians. Adherence is part of this challenge, but reducing it to a “reminder problem” misses the broader burden.
The UX opportunity here is to reduce the cognitive and coordination work of managing care, and not simply reduce the number of taps, you see. Instead of repeatedly reminding someone to complete a task, the agent can help identify what needs attention, coordinate approved next steps, surface relevant context, or initiate a predefined workflow.
Early healthcare deployments are beginning to move in this direction. For example, WellSpan Health has expanded its AI agent Ana beyond its initial screening use case into post-discharge follow-up and chronic-disease check-ins.
3. Acting Changes the Trust Contract
The moment AI moves from advising to acting, trust alone is no longer enough. Users need control over what happens next.
When an agent can take action, users need to understand:
- What the agent is doing
- Why it is doing it
- What information or context informed the action
- What happens next
- What the user can change, stop, or approve
This is important to make the action and its rationale legible at the moment it matters.
So, the solution lies in appropriate autonomy, that is, the right level of system agency for the task, the user, and the risk involved, with clear boundaries and an easy path back to human control.
What are the Best Practices for Designing Agentic AI UX in Chronic Care?
1. Design the Autonomy Ladder Before the Interface
Before designing the interface, define what the agent can do independently, what requires approval, and what remains off-limits. Map these decisions by task and risk, then align product, clinical, legal, and safety stakeholders before the agent goes live.
Make Autonomy Visible and Reversible
Give users clear control over the agent’s authority. The following patterns can help:
- Permission surfaces: Let users see and set what the agent is allowed to do in plain language.
- Dry-run mode: Show what the agent would do before allowing it to act, particularly during onboarding or testing.
- Action-level reversibility: Make it clear which actions can be cancelled, which can be corrected, and which cannot be undone.
The more consequential the action, the more explicit the user’s control should be.
Design the Override into the Product
Human oversight should be built into the workflow. For high-risk AI systems covered by the EU AI Act, Article 14 explicitly requires mechanisms that enable human oversight, including the ability to disregard, override, reverse, or interrupt the system where appropriate.
Now that is vital because automation bias, the tendency to over-rely on AI output, is a UX problem as much as a training problem. Therefore, for chronic-care products, designing deliberate moments for human judgement, approval, and intervention into the workflow itself is essential.
2. Make the Agent’s Actions Easy to Understand
When an agent takes action, people need to know what it did and what information led to that decision. And they need that context when the action happens.
Explain the Decision
A patient doesn't need to know how a model was built. They need to understand why the agent did something that affects their care.
For example, instead of:
“The model detected a statistically significant deviation in your longitudinal blood-pressure pattern.”
It’s recommended to say:
“Your blood pressure has been above your target for the past 7 days, so I’ve flagged it for your care team.”
Then let the person see the readings behind that decision, what happens next, and what they can do about it.
This is also consistent with current health IT guidance. The ONC’s HTI-1 rule requires certified health IT to make relevant decision-support information available in plain language, including information about intended use, inputs, performance, limitations, and risks.
Be Honest When the Answer isn't Clear
An agent shouldn't sound certain when the information behind its decision isn't. If the data is incomplete, conflicting, or outside the system's validated use, the experience should make that clear and tell the user what that means for the next step.
That doesn't necessarily mean showing a confidence percentage. Sometimes “I don't have enough information to make this call” is more useful than a number.
The FDA's 2026 guidance on clinical decision-support software similarly emphasises giving clinicians enough information to independently assess the basis of a recommendation. This includes relevant inputs, methods, validation, and known or unknown factors.
3. Earn the Right to Interrupt
In chronic care, an agent earns attention by knowing when to speak and when not to. People already deal with a steady stream of reminders, alerts, messages, and health information. Clinical alarm systems offer a useful lesson. When too many signals demand attention without requiring action, people can become desensitised and miss the ones that matter.
Give Every Interruption a Reason
Rather than treating every signal equally, give the agent a simple urgency ladder:
- Log: Nothing needs attention now.
- Digest: Useful information can wait.
- Nudge: There is one action worth taking.
- Escalation: Waiting could create meaningful risk.
The thresholds should be defined with clinical teams and refined using real-world behaviour. You see, more reminders do not automatically mean better adherence. A 2026 systematic review found no statistically significant improvement in medication adherence across 13 digital-health studies involving 1,320 community-based adults. Further, a large pragmatic trial also found that text reminders, behavioural nudges, and a chatbot did not improve 12-month cardiovascular medication refill adherence.
So yes, to sum up, if nothing needs attention, the best intervention may be no intervention at all.
4. Reduce Management Overhead, and Not Just Taps
Good chronic-care UX entails reducing the work of managing a condition, and of course, not just making individual tasks faster. The real measure is how much attention the experience demands from someone trying to get through their day.
Automate the Work Users Shouldn’t Have to Do
Automated insulin delivery offers a useful example here. A 2026 feasibility study tested an experimental Omnipod system that could adjust insulin automatically for adults with type 2 diabetes, without requiring them to tell the system when they had eaten or manually give extra insulin for meals. With the final algorithm, participants spent an average of 68% of time in range – a 24% improvement over standard injection therapy.
It is early evidence, and not a universal rule. But the design lesson is important. When an agent can safely take care of a repetitive task, users shouldn’t have to keep stepping in to manage it.
Design Around the Person
People with multiple conditions shouldn’t have to become the system that connects their care. An agent can bring relevant tasks, conflicts, and priorities together into one coherent experience rather than creating another stream of condition-specific prompts.
This is also where our work with Murphi.ai offered a useful perspective. We worked across its healthcare communication workflows to break down complex stakeholder journeys, prioritise frequent use cases, and reduce cognitive load. In that way, we designed the experience around helping people get to the right action faster.
Chronic-care experiences therefore need to adapt to how much a person can manage at a given moment. When someone is tired, unwell, or overwhelmed, the experience should lighten the load and not ask them to do even more.
5. Design the Handoff as Part of the Experience
When an agent needs to bring a person into the loop, the handoff should feel like a continuation of care. It shouldn’t be perceived as a break in the experience.
A good handoff gives each person the context they need:
- For the patient: Explain why a person is getting involved, what happens next, and what to do in the meantime.
- For the clinician: Surface why the issue was escalated, the relevant patient context, and what the agent has already done.
- For the agent: Make clear when it should stop acting and wait for human direction.
This is vital because poor clinical handoffs can lead to missing information, communication problems, treatment errors, and delays. Research on medical AI similarly argues that human oversight only works when people have enough information, decision-making authority, and a genuine ability to intervene.
The design principle is simple. Instead of just designing the escalation, design what happens before, during, and after the handoff.
6. Design for the Care Circle
Chronic care rarely happens between a patient and a product alone. Family members and caregivers are often part of the day-to-day work of managing care.
Give Caregivers the Right Access
Caregiver involvement should be scoped, visible, and easy to change. Let the patient decide what someone can see or do, make those permissions easy to review, and allow them to be changed as circumstances change.
A caregiver might need access to medications and appointments without seeing every part of a person's health information. Ideally, agentic UX needs to support the care relationship without turning it into surveillance.
Provide Clinicians With the Signal, and Not the Entire History
Clinicians don't need to read everything an agent has done. They need to quickly understand what changed, why it matters, and what needs their attention. Therefore, the agent should surface the important exception while keeping the underlying data available when needed.
The same principle applies across the care circle, you see. Different people need different levels of information and control.
The aim is not to design for one “primary user.” Rather, the objective is to design a shared care experience with the right role, access, and context for each person.
7. Design for Every Kind of User
While designing agentic AI UX for healthcare, we need to take into account factors such as disability, ageing, low digital confidence, and moments when someone is tired, unwell, or overwhelmed.
Make Accessibility Part of the Agent’s Control System
Good accessibility starts with making things easier to understand and easier to control. WCAG 2.2 reinforces this through requirements around predictable help, reducing repeated effort, accessible authentication, and perceivable status changes. For agentic interfaces, these basics become even more important when the system is acting on someone's behalf.
Put simply, that means not asking people to re-enter information the system already has, keeping help easy to find, and making important changes perceivable beyond the visual interface. WCAG 2.2’s Status Messages criterion, for example, requires relevant status changes to be available to assistive technologies without taking focus, so an agent's important updates shouldn't exist only as something a sighted user can see.
Don't Make One Modality Carry the Whole Experience
Voice can make healthcare interactions more accessible, particularly when hands-free or eye-free interaction is useful. But research with older adults also highlights barriers including speech-recognition errors, privacy concerns, and usability challenges.
So don’t make one mode do all the work. Give people more than one way to understand what the agent is doing and respond to it, whether that's through voice, visuals, or another accessible interaction.
Also Read: 10 Best Practices for Conversational UI Design
Build Understanding Before Increasing Autonomy
Agentic onboarding should go beyond showing people where features are. It should establish what the agent can do, when it can act, and how they can intervene or stop it.
Start with lower-risk actions, let people see how the agent behaves, and introduce greater autonomy gradually. By the time it acts independently, users should have a clear understanding of its capabilities and boundaries. While designing, it is therefore important to account for low-capacity moments from the start.
How Should Agentic AI UX Adapt to Different Chronic Conditions?
The right level of autonomy in agentic AI UX in chronic care depends on the condition, the person's context, and what happens when the system gets something wrong.
Type 1 Diabetes
Diabetes offers one of the clearest real-world examples of automation in chronic care. Hybrid closed-loop systems connect continuous glucose monitoring with insulin pumps and use algorithms to automatically adjust insulin delivery, while still requiring some user input.
The results show why this matters. A real-world NHS study across eight paediatric diabetes centres followed 251 children and young people with type 1 diabetes. After 12 months of hybrid closed-loop use, average HbA1c fell by 7 mmol/mol and time in range increased by 13.4%.
It is a useful precedent for agentic UX, you see. Let the system take responsibility for a clearly defined task, within known boundaries, while making its actions visible and giving the person a clear way to step in when needed. The same principle can guide agentic experiences beyond diabetes.
Heart Failure Monitoring
Heart failure monitoring shows where an agent can add value beyond simply collecting data. Changes in weight, breathlessness, swelling, and other symptoms can provide important signals, and the American Heart Association recommends tracking these changes and reporting sudden or worsening symptoms to a healthcare professional.
That makes heart failure a natural setting for an agent to identify a pattern, put it into context, and explain what it might mean. Instead of presenting another graph, the experience could connect a change to information the patient recognises and clearly explain what happens next.
The important part is showing the basis for the alert. If the agent escalates a concern, both the patient and clinician should be able to understand what changed and why it warranted attention.
COPD and Asthma
Respiratory conditions can remain stable for long periods and then worsen quickly. That makes early recognition useful, but it also means the response cannot be improvised at the moment someone is struggling.
Asthma already provides a useful model through written action plans. The Global Initiative for Asthma (GINA) recommends that patients have a written, digital or pictorial asthma action plan appropriate to their treatment, level of asthma control and health literacy, so they know how to recognise and respond to worsening asthma.
An agent can build on that structure:
- Monitor agreed signals
- Recognise changes
- Guide the person through a response that has already been established with their care team
It should not invent a treatment plan or make decisions beyond the authority it has been given.
Mental Health
Mental health needs a more cautious approach because conversational support can quickly begin to resemble clinical care. The question is not simply whether an AI can hold a convincing conversation, but where its role should end and human care should begin.
India already draws an important boundary in its Telemedicine Practice Guidelines. The guidelines state that AI/ML-based technology platforms are not permitted to counsel patients or prescribe medicines. AI can support a registered medical practitioner with evaluation, diagnosis, or management. However, the final counselling or prescription must be delivered directly by the registered medical practitioner.
Other jurisdictions are moving in similar but not identical directions. Illinois’ 2025 law restricts certain AI involvement in therapeutic communication and independent therapeutic decision-making.
The design principle here is to make the AI’s role unmistakable. If it is supporting care, say so. Or, if a situation requires clinical judgement, make the handoff to a qualified professional clear. And when a conversation involves sensitive mental-health information, privacy, disclosure and escalation should be designed into the experience.
Polypharmacy
Polypharmacy is another compelling area for agentic systems because medication management can involve information scattered across prescriptions, clinicians, pharmacies, schedules and refills. An agent could help bring that information together and flag potential duplication, conflicting instructions, missing information, or refill needs.
Truth be told, the scale of the problem is significant. A systematic review and meta-analysis of 27 studies found a pooled prevalence of polypharmacy of 49% among older adults in India.
But this is also where the consequences of an incorrect action can be serious. That makes medication management a strong case for assisted intelligence rather than unsupervised autonomy. The agent can find the issue, explain why it matters, and prepare the information for review. But the final clinical decision should remain with an appropriately authorised human until there is stronger evidence for more autonomous approaches.
Interestingly, across these conditions, the same principle emerges. Agentic UX should find the level of autonomy that makes sense for the task, the person, and the consequences of getting it wrong.
Win Trust With Agentic Healthcare UX
It’s not that agentic AI UX in healthcare is all about making the interfaces smarter. Done well, it can reduce repetitive tasks, surface what matters, coordinate care, and give people more confidence about what happens next. But that value depends on getting the experience right. It’s crucial to give AI the right level of autonomy, make its actions understandable, and keep people in control when it matters.
At Onething Design, we bring this thinking into healthcare through Agentic Experience Design (AXD), helping teams turn complex AI capabilities into experiences that are useful and built around real human needs. Our healthcare work includes Apollo 24/7 and NuvertOS, where we’ve shaped experiences that make complex healthcare interactions clearer and easier to navigate.
If you’re building an agentic healthcare product, let’s make it genuinely useful for the people who rely on it. Get in touch to explore how we can help you shape an experience that feels intuitive and human.