This article is drawn from The Signal Room conversation with Jakub Grajcar. Listen to the full episode: Outpatient Care Is Broken—Here’s How to Fix It.
Healthcare invests time and money into the resources required to cater to the ongoing needs of their patients. However, they pay very little attention to the resources or systems required to bring a patient into the system in the first place. I find this to be severely unbalanced, particularly, as the phone system (or voice response) is literally most people's first impression of a healthcare service. This system of care access is the most critical component and it literally falls apart. I had the opportunity to discuss this phenomenon in a recent Signal Room interview with Jakub Grajcar, a practitioner who sells and markets voice AI. During the interview, I was able to formulate and sharpen a question that occupied my mind for quite some time. I was able to ask what justification do we have to consider it safe and better care, when the first voice that a patient hears during their interaction with the system of care is in fact a machine. It reminds me that there is a critical and glaring need for health systems to refine their first point of care access.
You Are Unaware of Most Access Failures
Let's start with a reality check that is unrelated to AI. Some businesses may know how many calls they got, but they have no clue how many were missed. The super busy days, the mornings after a holiday, or the thirty-plus people who were answered with a busy signal at the same time, will never show up on a report. As a result of that, the situation will never be discussed. Generally, missed-call rates hit about 50%, and most businesses just accept that is how they have to operate. This is a lesson on visibility. Access failure is the norm, and you cannot fix what you do not know. The most fundamental, basic, most valuable thing you can do is the automation of a process that used to result in zero outcome, to a process where that call will never be lost.
The Importance of the Greeting
The greeting is the first opportunity for a design to prompt instant trust. The greeting's design has implications throughout the entire interaction. The agent must make it known that they are an AI and must transfer a true emergency to 911. The greeting is the first opportunity to build trust. The emergency transfer and the AI agent's disclosure happen first, and only later does the AI agent explain its capabilities. While it may appear courteous to provide an explanation, it is actually providing informed consent to a user for whom the AI is in fact the wrong service. The appointment scheduling system has lost user trust before it has achieved its core function, regardless of how seamlessly it integrates.
What a System Optimizes Shows What It Values
The first test I would put to any vendor, and I took this from this episode, is asking what number they optimize. The most honest answer I heard was that it's better for the phone to ring and for a human to pick it up rather than post a high automation rate, because some of those calls are life or death. Some of those calls are life or death, and it's the person selling the technology who admits to me that the headline metric on the slide has to lose, and that completely changes the buying question. The selling question is better framed as. What does the system do at the edge, at the time when the caller is afraid, angry, panicked, or at the time when the caller is off the script, rather than out of the way when it has a good day? The number that a company decides to go after will tell you whether safety, a reference call, or a demonstration, is running the product.
Least Privilege Moves to the Front Desk
Traditionally, least privilege has been a back-office concept, relating to access for system security. A patient-facing agent pulls it to the waiting room. Not only does it speak. It reads and writes the record. The workable version is intentionally narrow. Consider appointment time, doctor preference (or availability)—the least sensitive points of the continuum, and only for the one interaction at hand. With the least privilege, keep asking what the task is and what is the minimum information necessary for the task at hand. For those who are on the receiving end of such systems, the traditional question of who can touch what becomes relevant again, for the system that interfaces with patients at 2 AM. If an agent can move a chart or a schedule, the scope of what it can also affect is a decision you made, or one that is made for you.
The Two A.M. Line
The design gets honest after hours, because that's when gaps are widest. Clinics (for understandable reasons) try to manage things during the day, and almost always leave the nights and weekends to fate. That is when on-call clinicians often field phone calls from patients who are conveniently trying to book an appointment. This situation is perhaps the most inefficient use of a clinician's time. The important part of this design is the interception. Let the agent hold the routine so the clinician stays free for the genuine emergency. Keep sending genuine emergencies to 911. Give them (the 911 team) the paperwork as a part of the design. A line that encapsulates a summary and a recording that you can review later is a different line than one that just rings into the void. The design that survives the reality is very simple. The most important part is to keep everything that is less important as something that a human can go back and read.
Remove the Work, or Just Move It
One line from this conversation is definitely worth stealing. AI should destroy work, not move it. Most tools fail that test by the book. They create some outputs that a human has to re-enter into some other systems, which has simply moved the work and renamed it automation. The tools that truly earn their keep leave the output where the team works and let the agents finish the work, so no one has to re-enter anything. The only valid claim to automation is the following. Ask for the automation rate one workflow at a time. Appoint managers to cancel a call that ends the appointment, because that approach will always make the numbers look better next to the appointment one, which has to go find an open slot to cancel the call. Blending the figures into one is a marketing artifact. The rate that makes sense is how many of a specific type of call the agents totally finished without a human.
Readiness Is an Organizational Problem
The most challenging idea to sell in any situation is that readiness has virtually nothing to do with the model, and the technology is the least of the problems. What sinks a rollout is dirty data and the rules that live only in the head of one scheduler, the fact that Dr. X never takes Tuesday follow-ups, and nobody takes the time to write it down. An agent joins the team just like a new hire, and the team cannot operate any faster than its ability to articulate its own unwritten rules. The rest is the unromantic side of change management. Give the patients a heads up before it goes live, and ask for their patience. Get the staff and leadership to support it and expect a ninety-day spike with lots of noise, and then mostly quiet for the remainder of the third month. And keep the framing consistent, as it augments the front desk. The human work of it, the insurance card, and the person in the waiting room are still neither close to being automated nor were they the thing that should be replaced.
Related reading from Hutchins Data Strategy
Hutchins Data Strategy Consultants helps health systems put this into practice. See Clinical & Operational Integration.
Authoritative sources
- HHS — HIPAA
- ONC — HealthIT.gov
- NIST AI Risk Management Framework
- World Health Organization — Ethics & governance of AI for health
Frequently asked questions
Why is patient access such an under-discussed failure point in healthcare?
Because the failure is mostly invisible. A practice can see how many calls it answered and has almost no idea how many it missed, so the busy Monday or the morning after a holiday never shows up in a report. Missed-call rates run toward half in this world, and most practices have quietly accepted that as the cost of doing business. You cannot manage what you never measure, which is why access stays a blind spot while everyone focuses on what happens after the patient is already inside the system.
Should an AI voice agent tell patients it is an AI?
Yes, and it should do it in the first breath. When a machine speaks first, trust is settled in the opening seconds, so the agent has to say plainly that it is an AI and route a real emergency to 911 before it does anything else. Only then should it describe what it can help with. That disclosure is closer to informed consent than to courtesy. A front door that hides what it is has already lost the trust it needed, however smoothly it books the appointment.
What does least privilege mean when an agent can read and write the chart?
It means giving the agent only the access the task in front of it actually needs. A scheduling agent mostly reads the doctor availability and writes the appointment preference, the least sensitive corner of the chart, and only for the one conversation at hand. The discipline is to keep asking what the minimum information is for this task, and to grant only the privileges that task requires. Once an agent can move a chart or a schedule, the reach of what it can touch is a decision you make on purpose, or one that gets made for you.
Does deflecting calls to AI give staff time back or just move the work?
The honest test is whether the tool removes work rather than shuffling it somewhere else. Plenty of tools produce an output that a person then has to re-enter into another system, which has moved the labor and relabeled it automation. The version that counts leaves the result where the team already works, clearly marked as done by the agent, so nobody re-keys anything. And the number to ask a vendor for is the automation rate one workflow at a time, since a single blended figure flatters the product and hides where a human still has to step in.
