Guides
What Happens When the AI Cannot Handle the Call
The handoff is the real product. How an AI phone agent should escalate to staff, what context it must carry, and the transfer rules practices actually ask for.
Every vendor demo shows the call that works. The patient asks for a Tuesday afternoon, the agent books a Tuesday afternoon, everybody nods. That call was never the risk. The call that decides whether this works in your practice is the one the agent cannot finish, because what happens in the next fifteen seconds is either a clean handoff or a patient repeating themselves to a human who has no idea what was just said.
Practice administrators evaluating AI phone agents should spend most of their evaluation on that second call. The handoff is not an edge case or a failure mode to be minimized in a slide. It is the product.
Why the handoff carries so much weight
A front desk already runs on transfers. The difference with an automated first layer is volume and consistency: the agent will hand off the same categories of call the same way, hundreds of times a week, and whatever it does badly it will do badly at scale.
Two things go wrong, and they are opposites. An agent that transfers too eagerly is an expensive phone menu, and staff stop trusting it because every call lands on them anyway. An agent that transfers too reluctantly traps callers who have already decided they want a person, which is the single fastest way to make patients hate a phone system.
Practices are specific about this. Two practices in the same week asked for the agent to add deliberate friction before transferring: acknowledge the wait, restate what the agent can actually do, and only then transfer. That is not a vendor trying to deflect calls. That is a practice that knows a transfer costs a staff interruption and wants the caller to get an answer first when there is one available.
The four things a handoff has to carry
A transfer that drops the caller into a cold queue has destroyed the value of everything that came before it. A usable handoff carries four things, and a buyer should check for all four.
Who the caller is, already verified. The staff member should not restart identity verification. If the agent confirmed the patient against the chart, that confirmation travels with the call.
What they actually want, in their words. Not a category code. The reason for the call, captured as stated, so the staff member opens with context instead of “how can I help you.”
What the agent already did. If it checked eligibility, looked at open slots, or found that the requested provider is not accepting new patients, that work should not be repeated. Repeating it in front of the patient is worse than not doing it.
Where it landed in athenaOne. The call should leave a record attached to the right patient, so the next person to touch the account sees it, including when the transfer itself fails and the patient hangs up.
That last one matters more than it sounds. An abandoned transfer with a written record is a callback list. An abandoned transfer without one is a lost patient nobody knows about.
The rules practices actually write
Escalation design gets specific fast, and the specifics are where generic automation falls over. A few real examples, all of them from practices running live call volume.
An OB/GYN practice’s rule for “I want to speak to a doctor” is a question first, then a branch: urgent or pregnancy-related concerns go to nurse triage during hours or the on-call provider after hours, while non-urgent requests get an appointment offer or a message to the care team. The same sentence from a caller routes three different ways depending on the answer to one follow-up question.
A psychiatry group’s medication rules include a hard escape hatch. Patients must have been seen within 90 days and have a scheduled appointment before a refill request is processed. But for a named list of medications where abrupt discontinuation is dangerous, if the patient refuses to schedule, the agent must not hold the line. It transfers to a live medical assistant immediately. The clinical reasoning behind that list came from the practice’s medication-management director, which is exactly where it should come from.
A primary care practice wanted callback details captured for third-party service companies calling on behalf of patients, and an OB/GYN practice needed authorized third parties explicitly permitted to book. Both are exceptions to identity verification rather than features of it, and both have to be designed deliberately or the agent either blocks legitimate callers or verifies nobody.
Bilingual calls add their own handoff. Auto-detection works, but bilingual callers intermix languages mid-sentence, and practices want certainty before switching rather than a guess. Some want a given language path on for one queue and off for another.
None of these rules can be inferred from a specialty name. They come out of a conversation with the people who answer the phone today.
Where the line sits, and why it does not move
Pretty Good AI runs administrative front-office work inside athenahealth. The agent books, confirms, routes, captures and chases paperwork. It does not decide what care anyone needs. {{SYMPTOM_SCREENING_LINE}}
That boundary is also an escalation rule, and it is the firmest one in the system. When a caller describes a clinical concern, clinical decisions remain with the care team. The call follows the screening and routing workflow the practice approved, including its designated staff or clinical queue. The agent’s job at that moment is to be fast and to arrive with context, not to be clever.
The same principle holds for anything urgent. Emergencies route to the practice’s own emergency instruction and to a human, immediately, with no attempt at resolution. Practice-approved screening and urgency-based routing are not independent clinical decision-making.
What to ask in the evaluation
Five questions, and ask for a live demonstration of each rather than a slide.
- Play me a call the agent cannot complete. What exactly does the staff member see and hear at the moment it arrives?
- What happens when the caller says “let me talk to a person” in the first five seconds?
- Where does an abandoned or failed transfer get recorded, and who works that list tomorrow morning?
- Who writes the routing rules for clinical concerns, and how long does a change to them take?
- What happens after hours when the destination queue is closed?
A vendor who can answer all five concretely has run real call volume. A vendor who reframes the question toward containment rate has not.
Containment rate deserves one direct comment. It measures the share of calls handled without a human, and optimizing for it in isolation rewards exactly the behavior patients hate. Read it beside transfer quality and callback volume, or do not read it at all.
Key Takeaways
- The call the agent cannot finish is the one that determines whether staff and patients trust the system, so evaluate the handoff rather than the happy path.
- A usable transfer carries verified identity, the caller’s own words, the work already done, and a record written back into athenaOne.
- An abandoned transfer with a record is a callback list; without one it is a lost patient nobody knows about.
- Escalation rules are practice-specific and come from the people answering the phone today, including deliberate friction before a transfer and explicit exceptions for authorized third parties.
- Clinical decisions remain with the care team; screening and urgency-based routing follow the practice-approved workflow. Treat containment rate as a secondary metric, never a goal.
Ask any vendor to walk one failed call all the way to the booked visit or the callback. What the staff member sees at the moment of transfer tells you more about the product than any demo of a call that went perfectly.
Sources
Ready to See It in Action?
See how Pretty Good AI escalates to your staff with the full call context inside athenaOne
Schedule a Demo →Written by Kevin Henrikson