SENA Insights

Judge Front-Desk AI by the Work It Finishes

Posted on November 12, 2026

← Back to Blog

Judge Front-Desk AI by the Work It Finishes

SENA Health November 12, 2026

By Anthony J. Wehbe, DO, MBA, FACOI | Founder & CEO, SENA Health | Practicing Internist, Senacare | Former Health-System Physician Executive

AI vendors in patient access report plenty of useful numbers: calls answered, hold times, the percentage of conversations completed without a person and, from the stronger vendors, appointments booked.

I would add two more questions.

How much of the work came back to your staff, and who finished it?

And, more importantly:

Did the patient’s need actually get resolved?

Was the referral completed? Did the prior authorization reach a decision? Did the medication get to the pharmacy? Did the patient actually reach the right clinician?

I ask those questions not only as the CEO of a healthcare technology company.

I ask them as a physician who still deals with the consequences when the answer is no.

I Still Practice in the Part of Healthcare Where Handoffs Matter Most

Today, through Senacare, I continue to practice internal medicine throughout the Philadelphia region.

My patients are not limited to people who can easily drive to an office, open an app, navigate a portal, and make three phone calls when something goes wrong.

We care for people in their homes, assisted-living communities, nursing homes and group homes. Many are elderly. Others have intellectual or developmental disabilities. Some depend heavily on family members, nurses, caregivers or residential staff to coordinate their healthcare.

That changes how you experience the “front end” of medicine.

I can finish a visit in a patient’s home and make the clinical decision that a medication needs to change, only to discover that the next challenge is getting the prescription, pharmacy, caregiver and follow-up process aligned.

I can evaluate someone in a nursing facility and decide that a specialist needs to see the patient, but writing the referral is only the beginning. Someone still has to obtain the records, find the appropriate specialist, determine whether that office accepts the insurance, arrange the appointment, coordinate transportation, make sure the facility knows the plan and then get the consultant’s recommendations back to the people caring for the patient.

In one of our group-home programs, a nurse may call with a problem involving a feeding tube or a change in a medically complex resident. The immediate clinical question may be relatively straightforward. What makes the situation difficult is everything around it: who needs to act, what can safely be handled in the home, whether another clinician or service needs to be involved, and who owns the issue until it is actually resolved.

Those are not hypothetical workflow diagrams to me.

They are medicine today.

And one of the lessons I keep relearning is that healthcare is full of tasks that get completed while the patient’s actual need remains unfinished.

A referral can be entered without an appointment ever occurring.

A prescription can be sent without the patient ever receiving the medication.

A message can be routed without anyone taking responsibility for what happens next.

A task can disappear from one person’s inbox while the problem simply appears in somebody else’s.

From the physician’s perspective, all of those failures eventually come back.

They come back as another phone call, another message, another family member asking what happened, or another patient sitting in front of you whose care did not move forward as intended.

That distinction has shaped how I think about AI.

Administrative AI Is Built Mostly to Automate Tasks

In 2026, economists Daron Acemoglu, David Autor and Simon Johnson published a framework for judging whether AI makes human expertise more or less valuable. In September, health economist Blythe Adamson applied it to 132 AI products that were commercially deployed or FDA-authorized in U.S. healthcare.

Of the 40 provider and payer administration products in her scan, she classified 39 as automating existing tasks. All five patient-access and scheduling products she reviewed were coded that way.

Adamson is appropriately careful about what this shows. Her scan represents a sample of products rather than the entire market, it is not weighted by utilization, and the classifications reflect one analyst’s assessment based on publicly available information.

But the pattern resembles what healthcare operators see every day in technology demonstrations.

The headline is usually some version of:

How much of the work can the AI complete without a human?

That is an important question.

A large share of patient-access work is frequent, repetitive and rule-bound. Appointment requests, directions, refill status, reminders and other routine transactions can increasingly be handled by AI at any hour.

Patients absolutely benefit when something that once required waiting on hold can instead be accomplished at 9 p.m. on a Sunday.

But that measures the work the technology handled.

It does not necessarily measure the work the healthcare organization finished.

Every Deployment Leaves a Remainder

Every AI agent eventually encounters something it cannot complete.

A referral needs records from another practice.

A prior authorization is waiting on a payer.

The insurance on file has changed.

The scheduling rules do not fit the patient’s circumstances.

A patient describes symptoms requiring clinical judgment.

The appointment the patient needs is unavailable at the preferred location.

The family does not understand what happens next.

When the AI reaches that point, the work has to go somewhere.

And that is where I think physician organizations need to look much more carefully.

Because the remainder matters.

As routine work moves to AI, what remains for people naturally becomes harder.

Instead of a staff member spending part of the day answering straightforward questions and part of the day solving difficult problems, the remaining human workload can increasingly become a continuous stream of exceptions.

The automation rate rises.

The complexity of the work left behind rises with it.

That can be a very good operating model — if it is intentional.

It is a bad operating model if nobody has designed what happens to the remainder.

I Have Seen This Problem From Both Sides

Long before SENA Health existed, I saw the same operating principle while working as a hospitalist and later as a physician executive at Kennedy Health System and Jefferson Health New Jersey.

We would sometimes look at an overwhelmed service and conclude that we had a staffing problem.

The logic was understandable.

If physicians, nurses or administrative teams were overwhelmed, add people.

But sometimes we would add capacity and discover that the workload barely changed.

Why?

Because the underlying problem was not simply the number of people.

It was the way the work moved.

One department completed its responsibility and handed the issue to another. Then another. Different teams operated from different processes. The patient technically moved forward while the workload multiplied behind the scenes.

We were using highly trained people to absorb the consequences of fragmented workflows.

I learned from those experiences that activity is not the same thing as progress.

Today, practicing through Senacare, I see the same principle on a much more personal scale.

A hospital discharge can be technically complete while a frail patient and family still do not understand the medication plan.

A referral can be documented while a homebound patient still has no realistic path to the specialist.

An order can exist in the EHR while the nurse caring for the patient in a group home still does not know what is supposed to happen.

And a front-desk employee can correctly route a request while the physician eventually receives it again because nobody owned the entire path.

That experience is one reason I am cautious when healthcare organizations talk about AI primarily in terms of labor replacement.

Removing work from one person does not necessarily remove work from the system.

Sometimes it simply moves it.

Physicians Have Seen This Pattern Before

When I speak with physician leaders about front-end AI, I frequently hear the same concern.

They worry about paying a vendor to take away the easiest work and leave their existing employees with everything difficult.

Physicians have already experienced versions of this.

As virtual care, retail clinics and alternative access models expanded, uncomplicated episodic care increasingly moved to convenient settings.

That convenience has real value.

But the composition of traditional medical practice also changed.

The patients sitting in front of physicians increasingly have multiple chronic conditions, unclear symptoms, social barriers, competing priorities and questions that cannot be answered by an algorithm or a five-minute interaction.

That is some of the most meaningful work in medicine.

It is also exhausting when nearly every encounter requires that level of cognitive and emotional effort.

Patient-access teams will experience a similar shift.

The people who remain become more important, not less.

They handle the frightened patient whose symptoms do not fit the workflow.

They figure out why the authorization stalled.

They find the specialist appointment when the obvious options are unavailable.

They reconcile conflicting information.

They understand that a transportation problem may be the actual barrier to completing the visit.

They recognize when an administrative request has become a clinical issue.

Those are not necessarily failures of AI.

They are places where good operating design should intentionally bring humans into the workflow.

The fair test for any partner — software, outsourced service or internal team — is not simply whether it can efficiently perform the easy portion of the work.

It is whether the overall system takes responsibility for the difficult part too.

Count the Needs You Finish

I think healthcare organizations need to change the unit of measurement.

The useful unit of work in patient access is not the call.

It is not the portal message.

It is not the text.

It is not the task marked “complete” in the EHR.

The unit of work is the patient’s need.

A need is finished when the referral results in the patient reaching the appropriate care.

It is finished when the prescription gets to the pharmacy and the patient can obtain it.

It is finished when the prior authorization reaches a decision and everyone knows what happens next.

It is finished when the patient discharged from the hospital has a workable next step.

And when the problem requires clinical judgment, it is finished when the right clinician receives the issue quickly, with enough context to act.

That sounds like a subtle distinction.

Operationally, it is enormous.

Once leaders start measuring finished needs rather than completed transactions, they begin asking very different questions.

Redesign the workflow before automating it

AI layered onto a broken workflow can simply scale the mistake at machine speed.

Before deciding what to automate, understand where work enters, where it moves, where it stalls, where it gets duplicated and where it comes back.

Give every exception an owner

Before go-live, know exactly who owns the work when the AI cannot finish it.

That person or team needs the training, authority and system access necessary to actually solve the problem.

A handoff without ownership is just another queue.

Bring people in intentionally

Human involvement should not automatically be considered an AI failure.

It should be part of the design.

Organizations should identify the moments where empathy, judgment, problem-solving or clinical reasoning add value and bring a person into the interaction before frustration escalates.

Report the remainder

Answer rate, containment rate and automation percentage tell leaders how much work AI handled.

They do not tell leaders how much work AI returned to somebody else.

Every patient-access dashboard should show how much work came back to the practice, where it went, how long it remained open and whether the patient’s need was ultimately resolved.

Otherwise, an organization can celebrate a 70% automation rate while its nurses, medical assistants and physicians quietly absorb the remaining 30%.

Why This Became Personal for Me

SENA Health did not come from a technology thesis alone.

It came from spending years inside healthcare operations and continuing to practice medicine while seeing the same failure from different angles.

At Kennedy and Jefferson, I saw the enterprise version.

At Senacare, I still see the patient version today.

When I walk into an assisted-living community, nursing home, group home or patient’s house, I am reminded that patients do not experience healthcare as departments.

They do not care whether their unresolved problem belongs to scheduling, referrals, nursing, prior authorization, the pharmacy, the payer or the physician’s office.

They simply know whether someone helped them get what they needed.

And the more vulnerable the patient is, the less reasonable it becomes to expect them — or an exhausted family member or caregiver — to serve as the project manager connecting all of those pieces.

That realization heavily influenced how we built SENA.

I did not want to build another call center.

And I did not want to build AI that simply made the first interaction faster before depositing the difficult work back into the medical practice.

I wanted to build an operating model capable of taking responsibility for what happened next.

Why We Built SENA Around People and AI

That is why SENA’s Access Command Center combines people and technology.

AI should handle predictable, repetitive work it can do exceptionally well, around the clock.

Healthcare-trained coordinators should handle exceptions, problem-solving and situations requiring human judgment or empathy.

Clinical issues should reach clinicians.

And work that never needed to exist in the first place should be redesigned out of the process.

The objective is not to maximize automation.

The objective is to maximize resolution.

We use a phrase at SENA that captures that idea:

The hard stuff, resolved.

For me, that is not just a marketing line.

It describes what I want when one of my own Senacare patients needs help.

When a nurse from one of our facilities calls about a patient, I do not particularly care whether AI handled part of the workflow, a coordinator handled another part, or a clinician ultimately stepped in.

I care that the patient received the right care.

I care that the nurse knows what happens next.

I care that the family is not left making five phone calls.

And I care that the unresolved administrative work does not eventually land back in the physician’s inbox simply because nobody else owned it.

That is the standard I believe we should apply to front-desk AI.

Three Questions for Your Team

Before evaluating another patient-access AI tool, I would ask:

  1. What percentage of the requests your AI touches ultimately comes back to your staff — and who measures it?
  2. When a referral, authorization, medication request or scheduling problem stalls, who owns it until it is resolved?
  3. Does your access dashboard measure interactions handled, or patient needs finished?

AI is going to fundamentally change patient access.

That is a good thing.

But the organizations that extract the most value from it will not necessarily be the ones with the highest automation rates.

They will be the ones that redesign the work so AI, healthcare-trained people and clinicians function as one operating system — and take responsibility for the patient’s need all the way to resolution.

Because ultimately, the patient does not care who answered the call.

The patient cares whether somebody finished the work.


Want to talk through your organization’s front end? Schedule a conversation with Dr. Wehbe.