Automation moves data, orchestration moves care

Show me a healthcare decision maker who’s not dreaming about AI as a hammer for their most painful nails. Who wouldn't want a team of tireless AI agents that solves staffing shortages, maximizes patient engagement, prevents readmissions, diagnoses more accurately and knows how to hit every one of CMS' latest menu of quality metrics?

But after buying dozens of point solutions in the last decade, each one creating an additional data silo and failing to deliver the promised value for the organization, care team or patients, history looks like it’s about to repeat itself. The last thing anyone wants is the Blood Pressure Agent not being aware of what the Post-Discharge Agent is up to.

Healthcare is sick with incrementalism, an unfortunate outcome of the hippocratic oath that has burdened us with a risk averse crowd that prefers to rearrange deck chairs on the Titanic.

AI is bringing us to a crossroads in history—we have a chance to shift from a scarcity mindset to one of abundance. To reap its value we need to stop automating workflows and start orchestrating care—exactly what the Awell platform is built for.

First - why now?

Post-pandemic, the hospital is unbundling. Care is increasingly delivered in multiple settings from patients' homes to the clinic to their smartphones, and across modalities, from synchronous in-person visits to asynchronous virtual care. The patient-doctor dyad has exploded into a team-based care approach where your annual check-up involves more people than the average wedding party.

In other words, care delivery has been rapidly evolving from a single-setting, two-person, one-point-in-time activity to a multi-setting, multi-channel, multi-stakeholder series of activities. How does a care delivery organization tie all of those activities together, making sure everyone works at the top of their license, not spending a minute too long on a given task and seamlessly handing over to the next stakeholder while keeping patients engaged?

This unbundling is happening against a backdrop of chronic margin pressure for care delivery organizations, where they face the triple challenge of managing rising operational costs while dealing with reduced reimbursements and increased competition from new care models.

The treatment historically prescribed for these ailments is called automation.

The different flavors of automation

The ultimate promise of automation is end-to-end reduction of manual tasks across time, settings, stakeholders and modalities as described above.

But automation as an umbrella term is about as useful as saying preventive health is going to solve healthcare. To make sense of this, we’ll look through the lens of how automation is enabled by software and define three big buckets. You might find conflicting definitions out there but we’re only slightly bending reality here - stay with us, it should make sense.

Below we describe these three buckets. Here’s a visual overview of the main differences

The first bucket is Workflow Automation which focuses on automating machine-to-machine tasks. Zapier is the generic poster child of Workflow Automation: a specific event in one application (e.g., an appointment booking) triggers an action in another (e.g., sending a confirmation SMS). Workflow automation focuses on connecting data between systems to save time on repetitive, manual tasks. Many software applications, including EMRs, also have baked-in Workflow Automation where an action in one part of the product (place lab order) triggers an automated action in another (send lab order to lab).

Process Automation is a step up from Workflow Automation in terms of complexity. It involves human tasks and more complex, multi-step activities that require deterministic logic and a sequence. It goes beyond machine-to-machine connections and introduces conditional steps that need humans to review or intervene. For instance, insulin adjustment based on continuous glucose monitoring (CGM) readings is a good example where process automation steps in to reduce manual error. Process automation often involves more endpoints than workflow automation: systems, applications, humans. The more endpoints involved, the more sophisticated the automation needs to be.

The last bucket is Care Flow Orchestration. Anyone who has ever tried to map out a flowchart to represent a process in healthcare understands that process automation based on the Business Process Modeling (BPM) paradigm is just not a good fit for healthcare. There is a good reason for this. Care processes are knowledge- and data-intensive, with many exceptions and possible variations. Trying to capture this in a simple boxes-and-arrows visual is like trying to fit a three-dimensional puzzle in a two-dimensional frame. Healthcare is more like jazz improvisation than techno - it’s futile to think all of the beats and melodies can be mapped out meticulously upfront. This is why Awell has been built as a care flow orchestration platform.

Increasing sophistication as more endpoints (human stakeholders, systems, applications, AI agents) are involved and as process complexity increases.

Pitfalls of workflow and process automation

So why not automate a bunch of workflows and rely on the organization to close the gaps? That’s what humans are for, right? To keep the oversight and tie it all together. Or even a job for our new AI overlords?

Fact is, improving a subprocess doesn’t necessarily mean improving the results of the overall process. In other words, if you’re digging patients’ graves and you automate things, you’re only digging graves faster.

Although this truth is as early as the field of process improvement itself, it is not widely understood and due to constraints (budget, time, proper understanding of the true process to begin with), the default modus operandi is to identify a sub-process, spend time and money optimizing it only to realize the problem isn’t solved or simply displaced. Like blowing out a candle when the house is on fire. 

How has healthcare evolved on a macro level after decades of point solutions? The needle has hardly moved at all because all point solutions optimize subprocesses and take care of parts of the workflow, leaving a trail of fragmented semi-automated workflows with gaps in between. These gaps are then covered by humans or rigid, hardcoded software.

With AI we’re about to experience the same fallacy. Because of the black box nature of current (mostly generative) AI solutions, a misunderstanding has crept up that AI solutions won’t be held back by the same gravitational force as non-AI point solutions. However, nothing could be further from the truth.

Consider the example below. 

• Prior to engaging with Awell, a customer had implemented an AI agent to automate post-discharge outbound calls to patients. Before the AI agent implementation, this task was fulfilled by call center operators.

• Before the AI agent implementation, the process started with the call center operators receiving a daily list of patients to call. They then scheduled and executed multiple attempts over multiple days to reach the patient, tracking each outreach attempt and outcome in the EMR. The objective of the AI agent was to reduce these manual actions to zero.

• The evaluation of the AI agent was done in the usual way, through a pilot. Isolated, with a non-representative volume of patients, not integrated with existing systems. The goal of the pilot was to understand whether the call quality was good enough to interact with real patients. The results were encouraging and the implementation of the AI agent process got a green light. But during the pilot no attention was given to how the AI agent would collaborate with humans in the overall process.

• Lo and behold, after the implementation it turns out that the call center team has gained somewhat in efficiency, but another team was burdened with new work to support unforeseen activities upstream and downstream of the AI agent such as identifying which patients would be good fits to be called by the agent, as well as dealing with exceptions that arose during the agent calls but couldn’t be resolved. Problem not solved but merely displaced.

• Ultimately the solution was to orchestrate the activities in between the different human stakeholders and the AI agent with Awell.

Care flow orchestration and AI agents

It’s plausible and for some use cases inevitable that AI agents will become integral members of the care team over time, but they will require orchestration to achieve meaningful outcomes and ensure proper collaboration. 

If the human nurse is not aware what the blood pressure agent is planning on doing, we have a problem. The nurse can’t just pick up the phone or send a Teams message to the AI. Care flow orchestration acts as the glue between these AI agents, care team, patients across the various care settings and modalities involved in patient care as the hospital keeps unbundling. The questions who is supposed to do what, when and what’s next? have become only more relevant with AI.

Orchestrating AI agents in care flows means implementing them not as the next generation of siloed tools but as true collaborators within a dynamic care process. As an integral part of the care team. The true value of AI emerges when these agents are coordinated effectively, ensuring data flows seamlessly between stakeholders, systems, and environments. When orchestrated properly, AI agents can help ensure care is delivered efficiently, minimizing manual burden without introducing new pain points elsewhere.

The CareOps Platform

Ultimately, for AI agents to be fully embraced by care teams, they need to earn trust - from both patients and care teams. Trust is built on consistent, transparent, and effective performance. It’s not enough for AI agents to perform a task well in isolation; they must do so in a way that visibly integrates into the overall care flow.

Starting at entry-level tasks such as documentation or triaging, these agents can progress through more complex roles like clinical decision support and patient monitoring. Their autonomy grows with experience and data, ensuring they fit seamlessly into the care environment. Much like a doctor in residency, the goal is to throttle the exposure to more complex cases and more responsibilities over time, supported by continued medical education to stay abreast of the standard of care. 

Using Awell’s CareOps Platform, care organizations can onboard and oversee AI agents safely and grow their agency and decision making capabilities incrementally, similar to clinical trial phases. This approach ensures AI agents are rigorously validated and aligned with patient safety standards before scaling.

At the same time, we recognize the care flow paradigm itself is shifting. Before AI agents, care flows were fully deterministic. We are currently already seeing implementations of care flows that have a deterministic backbone and where certain actions are handed over to AI agents. In the near future, we’ll see more and bigger parts of previously deterministic flows being handled by agents (the “Post-Discharge agent”, the “Chronic Care Management” agent). An iteration beyond that will see a flip where agents drive the overall care flow and invoke deterministic parts of processes like evidence-based protocols when needed.

1. Deterministic flows. If-this-then that logic drives flows with zero black box but higher rigidity.

2. Deterministic backbone. Agents perform parts of care flows but the process flow remains deterministic

3. Agentic backbone. Deterministic patterns remain (e.g. for compliance, medico-legal reasons) but agents drive the flow

As the leading CareOps platform, Awell not only enables care delivery organizations to capitalize on this evolution from deterministic to generative care flows. By providing the care team with capabilities to deploy agents in care flows for specific tasks, monitor their performance on those tasks, increase their knowledge and skills and expand their agency in a controlled way, Awell is creating the infrastructure to let humans collaborate with AI and finally hammer a good amount of those nails into the ground.