Agentic AI Accelerator in Action: Taking AI in Healthcare From Workshop to Reality

Unplanned hospital readmissions are more than just an inconvenience; they're a $26 billion annual burden on the U.S. healthcare system. With approximately 20% of Medicare patients readmitted within 30 days of discharge, it's clear that the transition from hospital to home is a critical juncture in patient care.

Our research reveals a startling disconnect:

• 92% of patients seek additional information after discharge, but…
• 70% turn to unverified internet sources rather than doctor-backed resources.

This information gap doesn't just affect patients — it impacts the entire healthcare ecosystem: When satisfactory or personalized answers from “Dr. Google” can't be found, and a provider can't be reached in a timely manner, many simply return to the ER, potentially leading to unnecessary readmissions, complications, and increased costs.

Aside from the direct provider-patient costs of rehospitalization, readmission rates are a key metric examined by healthcare payers and have a substantial impact on insurance reimbursements to the hospital. Given the magnitude of the problem, moving the needle even a few percentage points could have vast implications if implemented at scale.

The problem is clear, and while the perfect solution isn’t obvious, our two-week Agentic AI Accelerator was custom-built to illuminate a path forward using agentic AI and provide actionable next steps for industry and enterprise problems like this. In this article, I’ll reveal how we evaluated these monumental, post-discharge patient care challenges to prototype an innovative AI-powered healthcare companion that merges patient needs with healthcare provider goals, creating aligned incentives between patients, care teams, and payers — all of whom want to ensure access to high-quality, low-cost care.

Our AI-Powered Healthcare Companion

No need to keep you in suspense, and besides that, we’re eager to share how AI can be applied strategically and safely to workflows to change the status quo. Here’s the TL;DR of the tool we prototyped:

• It records and transcribes key conversations between patients and their clinical team.
• It brings these transcripts into a closed AI system along with personal patient records and doctor-curated research.
• It then allows patients and caregivers to ask questions in natural language 24/7 to an AI assistant that references only these data sources, ensuring more accurate, personalized, doctor-vetted responses (i.e., it’s not going to Reddit for medical advice).
• It will hopefully help lighten the burden from an overtaxed healthcare delivery system and, ideally, reduce unnecessary hospital readmissions.

How does it do that? With a secure orchestration of data management, AI agents, and intuitive user interfaces that provide a helpful, seamless experience for physicians, patients, and their caretakers. But more on that in a moment — first, a bit of personal context.

A Patient & Caregiver Perspective

By uncanny coincidence, personal life events crossed over with my involvement in developing the prototype of this AI-powered healthcare companion. As I was directing the latest season of our docuseries, 2 Weeks To Better (2WTB), which chronicles the rapid development, engineering, and design of the aforementioned solution, my father ended up in the ICU.

A relatively routine endoscopy had gone sideways, perforating a hole in his esophagus, which is apparently a very bad thing. Soon, we had visits from gastrointestinal specialists, thoracic surgeons, infectious disease experts, respiratory therapists, hospitalists, case managers, and an endless cadence of nurses who — along with my mother, a pediatric physical therapist for 50 years — I credit with saving his life.

This all started the day after a string of exhausting all-nighters where my family moved out of our house to begin a renovation project, so I was already running on fumes when the call came to drive up to Baltimore. I share this only to underscore one small example of the human realities behind what we often view as faceless medical statistics.

We were one family dealing with a ton of information from many doctors, a barrage of decision-making, and an endless stream of moment-to-moment logistical and emotional and medical needs to navigate. My mother and I were tired. We were overwhelmed. We were scared. And we had questions.

“How do we…” “What happens if…” “Is it normal/expected if…” “Who do we ask about…”

Given my familiarity with AI in healthcare, I’d been primed to understand the problem and prepare for it. So I initially used our Fuel iX enterprise AI platform to develop a makeshift copilot version of the TELUS Digital healthcare companion using these steps:

• I recorded conversations on my phone, transcribed them, and manually downloaded records from MyChart.
• I asked providers to recommend any research studies we should or shouldn’t reference.
• I put all of this into my copilot’s Knowledge Base (basically, a RAG system).

While my father ended up back in the ER about 10 days after his initial discharge (yes, he’s part of that readmission statistic now), I can say with complete confidence that the rudimentary AI tool I created changed everything for my family. It’s improving our ability to get answers quickly, manage our expectations, reduce our need to call overworked doctors and nurses, and understand when something is truly amiss and requires rehospitalization. So, imagine how strong my belief is in the full-fledged prototype I described in the TL;DR.

Responsible AI can drive the change needed in healthcare

One man’s opinion: our healthcare system needs tools like these. Every member of my father’s care team — across multiple hospitals, units, and rehab centers — had impressive credentials, the very best of intentions, and ultimately gave him excellent care. But what became abundantly clear in this kind of complicated case is that our healthcare system is overwhelmed and imperfect.

While it’s essential to have a healthy fear and skepticism of artificial intelligence tools in any healthcare setting — “first do no harm” remains a vital litmus test of any piece of technology — I firmly believe our system can’t afford to wait any longer on implementing these kinds of focused point solutions for patients and caregivers, providing more personalized guidance, helping to prevent provider burnout, and reducing unnecessary hospital visits.

Change is needed — the risk of sustaining the status quo is too great. And with that, let’s dive into how that change can be achieved.

Phase One: Identifying High-ROI Opportunities (AI Use Case Workshop)

Starting with an Agentic AI Use Case Workshop, we brought together a diverse team of technical experts, designers, and healthcare professionals — led by TELUS Digital Partner and VP, Healthcare Industry Commercial Lead Sydnor Gammon — to identify high-impact opportunities for AI implementation in emergency healthcare.

The session was facilitated by 2WTB project lead and Sr. Manager of Data and Delivery Conner Brew, who presented a collaboratively-built patient journey for a patient during/after hospital discharge. The group zeroed in on waypoints when the patient and their caregiver inevitably have overwhelming or clarifying questions for physicians after leaving the round-the-clock care of a hospital — especially emergency questions about whether a particular symptom is expected and warrants a return trip to the ER.

Workshop outcomes

This targeted workshop helped identify key opportunities for innovative solutions that benefit both patients and hospitals. Key outcomes included:

• Identification of post-discharge care as a critical pain point
• Initial concepts for AI-powered patient support
• Alignment on success metrics and ROI potential through reduced readmissions — a key metric for hospital insurance reimbursements

Motivating outcomes like these are the benefit of TELUS Digital’s highly structured Agentic AI Use Case Workshops, in which we guide participants to uncover high-ROI opportunities for data and agentic AI solutions. Workshops foster a shared understanding of agentic AI's potential, identify critical pain points, and prioritize use cases based on measurable metrics. Before moving forward with solutions, teams validate opportunities with quantitative and qualitative data, and that's exactly what the 2WTB team did next.

Phase Two: Validating the Business Case (Research & Strategy)

Following the Agentic AI Use Case Workshop, our research and strategic roadmapping phase, led by Senior UX Researcher Kristen Duke, went beyond surface-level statistics to understand the human experience of post-discharge care. Through a combination of surveys, first-hand accounts, and analysis of existing healthcare data, we uncovered compelling evidence for AI intervention in the discharge process.

"People just basically want to be able to ask questions with answers that are doctor-backed," Kristen explained, focusing on the real human experiences and challenges behind the statistics. "And so there's this big need, this big opportunity to fill that gap."

Take Andrew Deatherage, TELUS Digital Data Advisory Director and a research participant who had recently experienced a hospital stay. His account illuminated the complexity of post-discharge needs: "I was keeping track of medications, my medication schedule, what mixes, what can't mix ... When it comes to follow-up questions, it's more of a what-if analysis rather than just 'this piece of paper says, ‘take this twice a day.''"

Another research participant reinforced Andrew’s point, saying, "People often don't know if they're experiencing symptoms when to call for help. Some people won't seek the care that they really need, and some people will immediately go to the emergency room for anything —– and that's not good for anyone."

Targeting two solutions

Our research revealed three critical gaps in current post-discharge care:

1. Information access: Patients struggle with antiquated paper-based instructions and lack easy, searchable access to their medical information when questions arise.
2. Trust and verification: As Andrew noted, "I don't want to see what Reddit has to say about this concept. I want to see what the Mayo Clinic has to say about it … what the World Health Organization says about it."
3. Continuous support: Patients need ongoing guidance as new questions and concerns emerge during recovery.

The research phase concluded with a clear mandate: develop an "AI Healthcare Assistant" that could provide personalized, doctor-backed information, providing measurable value to healthcare providers through reduced readmission rates and improved patient outcomes.

Key capabilities of this solution identified in our AI use case research included:

1. Live in-room transcripts and AI note-taking
2. Conversational post-visit AI assistant

A third capability, personalized AI medication management, was proposed, but when the team discussed potential dangers — when dealing with dosages and timing, the fear of a small AI mistake or hallucination having a major negative impact is too great — functionality was streamlined. As always: “First, do no harm.”

Phase Three: Leveraging Our AI Accelerator Framework (Technical Architecture)

With a point solution more clearly defined, our technical team — led by Nish Tahir and Chrisopher Frenchi — set out to design a scalable architecture for the AI Healthcare Assistant that could be rapidly prototyped and tested. As leaders of our Data and AI Research Team (DART), Nish and Christopher are constantly developing expertise on the latest tech underlying automation and AI technology advancements —– they’ve published on topics ranging from evaluating conversational AI for politeness to choosing the best speech-to-text (STT) models for voice applications. So when they begin a project like this, it’s typically with a running start.

The solution they proposed for this post-discharge leverages three key components of our AI framework:

1. STT integration: This enables the system to record and transcribe patient consultations with the clinical care team, creating a valuable, searchable resource for both immediate reference and long-term data analysis.
2. Retrieval-augmented generation (RAG): This technique ensures that the AI works within a closed system of accurate medical information, drawing from verified sources rather than broader world knowledge. "We're only going to add in approved documents, and personalized, specific details about the patient themselves," Christopher notes.
3. Natural language processing (NLP): This allows the system to understand and respond to post-discharge patient questions in a conversational manner.

Ensuring trustworthy AI

Speaking from the DART team's deep expertise building AI solutions, Nish notes the unique challenges the healthcare industry poses: "Where things get a lot more complicated is making sure that the system is private, and it's trustworthy."

To address these concerns, the team implemented robust privacy measures and trustworthiness protocols. As Nish explains, "[Clinicians] are still accountable for what is released to the patient. So it's accountable, it's explainable, and most importantly, that's how you de-risk and make sure that the system is safe for everyone."

The team also focused on creating a comprehensive evaluation suite to ensure the reliability of the AI Healthcare Assistant. This meticulous attention to detail and commitment to responsible artificial intelligence development is a hallmark of a mature approach to AI in healthcare. By combining cutting-edge technology with stringent safety and privacy measures, we were able to begin prototyping an AI healthcare assistant that patients can trust and healthcare providers can rely on.

Phase Four: Embracing Human-Centered AI Design (UX Design Prototype)

Our design team, led by TELUS Digital Design Director Ryan Davis, approached the next phase of the challenge with a laser focus on user needs. "How can we really support the patient after they've been released from the hospital? That was really the crux of the product," Ryan explains. This human-centered approach to AI design is a cornerstone of our methodology, and was on display as we developed the core capabilities of the tool:

• ‍Live in-room transcripts and AI note-taking: This feature tackles the problem of information overload during doctor visits.
  
  • "We don't want to place that burden of recall on the patient [who is in the middle of a lot of pain]," Ryan notes. Our goal was to design a solution that makes it as frictionless as possible for the system to record and transcribe consultations. With AI assistance, patients can be provided with a reliable record of their care instructions, so that the precious minutes of interfacing with doctors and nurses aren’t spent furiously scribbling notes.‍
• Conversational post-visit AI assistant: This AI-powered chat function allows patients to ask questions about their condition and treatment in natural language. 
  
  • As Ryan describes, "Someone can ask this chat in a very conversational way, 'Hey, I'm experiencing some pain here. What does this mean? Should I go back to the hospital?'" The AI draws only on trusted, doctor-approved resources, personal patient records, and the recently recorded conversations to provide accurate answers.

Enhancing human-centered design through user testing

The design process was highly iterative, with the team conducting rapid user testing even at early stages. "We love to perform user testing even at that low stage of fidelity because at that point we can still get an idea of how the work is performing," Ryan explains.

The result is a prototype that not only uses advanced AI capabilities but does so in a way that feels intuitive and trustworthy to both patients and healthcare providers. This human-centered design approach, combined with rapid iteration and expert feedback, positions the AI Healthcare Assistant to truly transform the post-discharge patient experience.

Phase Five: Proving the Concept (Solution Validation)

The final phase of our two-week Agentic AI Accelerator process brought us to the University of Virginia (UVA) School of Data Science, where we presented our solution to a distinguished panel of healthcare and business experts. This crucial “Attack the Concept” step in our methodology ensures that innovative solutions are grounded in real-world practicality and ready for implementation (i.e., “Don’t trust an idea from someone who can’t build it.”).

Expert validation and insights

The presentation generated enthusiastic feedback.

Reza Mousavi, a UVA McIntire School of Commerce Professor with expertise in artificial intelligence and business analytics, was particularly struck by the solution's depth: "Given the amount of time that the team just had, I was expecting something more generic, very high-level, not as detailed as we saw today. So that was pretty impressive."

Dr. Matthew Trowbridge, MD, MPH, an emergency medicine physician at UVA Health, highlighted the AI Healthcare Assistant's potential for handling complex cases on the physician side of the equation: "We get what we call ‘zebras,’ someone coming into the ER with a rare variant of something. In that context, you have to make a decision about managing an acute complaint, while trying your best to understand the context of what else is happening in this patient's life."

Artificial intelligence can help here — not going so far as to recommend specific treatments or interventions, but simply riding alongside a provider (especially a less experienced resident) and analyzing and summarizing a complex medical history while the clock is ticking on an emergent medical situation.

Multi-dimensional value proposition

The panel agreed that the AI Healthcare Assistant demonstrated clear value across three key stakeholder groups:

For healthcare providers

• Potential reduced readmission rates through better patient support
• Improved patient compliance with treatment plans
• More efficient use of clinical time and resources
• Better documentation and tracking of patient progress

As Dr. Trowbridge notes: "We could have much better utilization of our existing resources, which include doctors' attention, time, nursing, even just the physical space of emergency departments. Everything could run a lot better."

For patients

• 24/7 access to verified medical information
• Reduced anxiety about post-discharge care
• Better medication adherence
• Seamless communication with healthcare providers

The impact on patient experience was particularly noteworthy. As Professor Mousavi observed, "The moment a chatbot can answer ‘Why am I breathless today?’ by surfacing last week’s abnormal spirometry, we shift from generic advice to precision triage.”

For healthcare systems

• Lower costs and higher insurance reimbursements through reduced readmission potential
• Improved patient satisfaction scores
• Better allocation of resources
• Valuable data collection for system improvement

Professor Mousavi emphasized the broader implications: "When chatbots securely tap electronic-health-record pipelines, they not only raise patient-satisfaction scores and trim clinical workload, they also generate high-fidelity datasets the FDA now recognizes as decision-grade for trials and post-market studies."

Implementation roadmap

Based on this expert feedback, we next developed a clear path to implementation:

1. Initial workflow integration: As Sydnor Gammon notes, "For purposes of this exercise, we're going to create something that's likely to be a standalone app. But realistically, we know this would need to be embedded into existing workflows for a clinician or how a patient would want to navigate their own care experience." (i.e., a feature of Epic or MyChart)
2. Continuous improvement: Conner Brew outlined the next steps: "We would set up a cross-functional team consisting of software engineers, AI engineers, and data scientists who could begin laying the groundwork for that agentic AI system."
3. User experience refinement: "We would take a researcher and a team of designers and continue conducting rapid iterations to improve that user experience," Conner adds.

After in-depth feedback, the validation phase confirmed not just the technical feasibility of our solution, but its potential to transform healthcare delivery. Professor Mousavi framed the AI Healthcare Assistant as the successful, real-world application of academic best practices: “In our AI & analytics courses, we preach four fundamental components: the right governed data, the right AI model, iterative experimentation, and empathetic user experience. TELUS Digital turns that syllabus into practice, monetizing de-identified datasets for R&D while never losing sight of the patient voice.”

Ready to Transform Your Healthcare Technology?

This AI Healthcare Assistant is just one example of how our Agentic AI services can drive innovation in healthcare. Whether you're looking to improve patient outcomes, reduce costs, or enhance operational efficiency, our proven Agentic AI Accelerator process can help you:

• Identify high-impact AI use cases through our structured workshops
• Rapidly prototype and validate solutions
• Build alignment and scale successful AI/ML innovations across your organization

Ready to explore how our Agentic AI services can transform your healthcare technology strategy? Schedule an AI Use Case Workshop or learn more about our Agentic AI Accelerator program.