Can an AI agent schedule appointments over the phone?

Yes, as long as it is connected to the real schedule and verifies identity before booking. The agent answers, understands the reason in natural language, finds slots, offers options and books, and if something is unclear it transfers the call to a person with a summary. In Omcall, booking requires verification with a one-time code or a higher level, and the appointment stores the data Resolution 1552 requires: request date, desired date, assigned date and the site's REPS code (its entry in Colombia's registry of health service providers).

Does the patient need to know they are talking to an AI?

Yes, they should know from the first message and always have the option to talk to a person. It is a matter of trust and also of data protection: Law 1581 requires telling data subjects what their data will be used for and that answering questions about sensitive data, such as health data, is optional. In Omcall, the greeting that says "virtual assistant" is a fixed rule your IPS cannot remove, and the patient can ask for a person at any time.

What happens if a patient tells the AI agent about an emergency?

The agent stops the conversation, gives the 123 emergency line and immediately hands the case to a person. It gives no clinical guidance, not even to reassure the patient. In Omcall, detection runs before the model responds, using phrases your IPS can configure, and the handoff enters the emergency queue with higher priority. The default phrases are a draft for your clinical team to review and complete with the terms your patients actually use.

How does the AI verify the patient's identity before sharing information?

It asks for the ID document and an additional detail only the patient knows, or sends a one-time code, and that raises the verification level of the conversation. Each action requires a level: searching for slots needs no verification, but viewing appointments or booking does. If the conversation lacks the required level, the system rejects the action, and the AI cannot bypass that rule. This matters because Law 1581 classifies health data as sensitive.

How do you measure whether an AI agent is working well?

With the same metrics as your team, calculated with the same formula, plus a few specific to AI. The shared ones are service level, ASA, AHT and FCR; in Omcall, a case counts as resolved on first contact if the person does not come back within 7 days (adjustable). The AI-specific ones are resolution without a person, escalation rate, cost per interaction and tool errors. In addition, each version must pass test scenarios such as an emergency or a manipulation attempt before it is published.

Does AI replace the people in the call center?

No. AI takes over repetitive tasks with clear rules, and people remain for what requires judgment: emergencies, complaints, cases that do not fit and patients who prefer to talk to someone. In Omcall, people and AI share the same queues, and the staffing forecast separates the volume the AI will handle from the volume that will reach your team, so you can size shifts with Erlang C. That way you decide how many people you need based on data, not assumptions.

How much does Omcall cost?

Omcall is billed as a license per agent per month plus AI and channel usage, such as voice minutes and messages. The price depends on how many people handle contacts, how many sites and channels you connect and which AI agents you turn on, so it is sized in the demo with your real operation, and that produces a specific figure, not a range. AI usage has a monthly cap set by your IPS, with an alert and a stop. IPS that join as design partners get a 30-day pilot at no cost. You can book the demo.

AI agents for an IPS call center

Before you put an AI in charge of answering your patients, decide what it can do on its own, what should go to a person and how you will measure it.

Updated October 15, 202610 min read

This guide is for you if you run an appointment center, coordinate its quality or are responsible for data protection at an IPS (instituciones prestadoras de salud, Colombia's healthcare provider institutions), and you are evaluating AI agents over voice or WhatsApp. It explains what an AI agent does compared with an IVR or a chatbot, which tasks are worth automating and which are not, and what controls to require before you let it talk to patients.

In short

What it is: an AI agent converses in natural language and carries out tasks in your systems, such as finding open slots or booking. An IVR or a button-based chatbot only follows a menu.
What to automate: administrative tasks with clear rules, such as scheduling, confirming, rescheduling, slot refilling and giving the status of authorizations or PQRS (petitions, complaints, claims and suggestions).
What not to automate: emergencies, clinical advice and complaints with legal risk. Those go to a person.
Fixed rules: it introduces itself as a virtual assistant, offers a person at any time and verifies identity before sharing data.
How to measure it: with the same metrics as your team, plus resolution without a person, escalation and cost per interaction.
Health data: it is sensitive under Ley 1581 de 2012 (Law 1581 of 2012), and the Superintendencia de Industria y Comercio (SIC, Colombia's Superintendence of Industry and Commerce) asks for a privacy impact assessment when the risk is high.

What an AI agent does in an appointment center

A classic IVR asks the caller to press options and routes the call; it does not resolve anything on its own. A rule-based chatbot follows a tree of buttons and gets lost when the patient goes off script. An AI agent understands a free-form sentence, such as "I need a pediatrics appointment for my daughter sometime next week in the afternoon", checks the schedule, proposes options, asks for verification and books. When it cannot, it hands the case to a person with a summary, so the patient does not have to repeat everything.

The underlying difference is that the agent acts. So the question is not how natural it sounds, but which actions you allow and who reviews what it did.

On voice the pattern is the same: the patient can interrupt and, while the agent checks the schedule, hears a holding phrase instead of silence. On WhatsApp it answers at any hour. The guide to WhatsApp appointment scheduling shows that flow step by step.

Tasks worth automating:

Scheduling, rescheduling and canceling appointments, with prior verification.
Confirming appointments ahead of the date and recording the response.
Offering freed slots to the waitlist.
Answering administrative questions from a knowledge base: sites, hours, requirements and documents.
Receiving a photo of a medical order and giving the status of the authorization.
Filing a PQRS and giving its status.

Reminders have evidence behind them: a Cochrane review found that text message reminders increase attendance compared with sending none, with an effect similar to that of phone call reminders. There is more on this in the guide to reducing no-shows for medical appointments.

Tasks that should not be automated:

Emergencies: when a patient says "my chest hurts" or shows any sign of a life-threatening risk, the agent gives the 123 emergency line and hands off to a person.
Clinical advice: symptoms, medications or interpretation of results stay out of scope, even if the patient insists.
Complaints with risk: any mention of a tutela (a constitutional rights action), a lawsuit or the Superintendencia Nacional de Salud (Supersalud, Colombia's National Health Superintendency) goes to patient services.
Anything your IPS has not delegated: if an action is not in the agent's rules, the agent does not do it.

What an AI agent for an IPS should have

This list works for evaluating any vendor, Omcall included. Ask to see it working with difficult cases.

Transparency that cannot be removed: the greeting says it is a virtual assistant, regardless of configuration.
A person one step away: the patient can ask for one at any time, and the handoff carries the case summary.
Emergencies before the model: rules that detect risk phrases before the AI responds, with a fixed message and an immediate handoff to a person.
Tiered verification: searching for slots can be open; viewing appointments or booking should require a code or a detail only the patient knows.
Autonomy levels per action: act (the AI does it), approve (it does it once a person approves) and suggest (it leaves the proposal to whoever is handling the contact). Canceling an appointment or generating a charge should not start at act.
A single gateway to your systems: every action goes through one point that validates permissions, prevents duplicates and keeps a log.
Testing before publishing: scenarios such as a routine appointment, an emergency, an upset patient or a manipulation attempt, with versioning and rollback.
The same metrics as your team: service level, AHT and FCR with the same formula.
Minimal data to the model: names and ID numbers replaced before they are sent, and text from documents or messages treated as data, never as instructions.
Spending cap: a monthly budget that, once used up, hands conversations over to people.
Supervision with control: the ability to view an AI conversation and take it over.

How Omcall handles it

Omcall's agents run on its own engine, with rules your IPS cannot delete: introduce itself as a virtual assistant, always offer a person, give no clinical advice and, in an emergency, give the 123 line and transfer to the emergency queue. On voice, a risk phrase triggers the fixed message without the model processing it. Every handoff carries the open case.

Omcall includes ten agent templates that share one case file per case:

Receptionist: identifies the patient, understands the reason for contact, answers from the knowledge base and hands off to the right agent or person.
Scheduler: finds slots, books, reschedules, cancels and adds patients to the waitlist.
Confirmer: sends the reminders at 72 and 24 hours and records the response.
Slot filler: offers freed slots to the waitlist.
PyM campaigns: contacts health promotion and maintenance (PyM, promoción y mantenimiento) cohorts from imported lists.
Authorizations: receives the order, files it and gives its status.
PQRS: files, classifies, gives the status and hands risky cases to patient services.
Payments: reminds patients of pending payments; a new payment link requires human approval, and disputes go to a person.
Supervisor: reviews ongoing conversations and alerts supervisors or asks for a person to take over.
Quality: evaluates interactions and proposes knowledge base changes as drafts.

The templates arrive turned off and in draft. Your team adjusts instructions, knowledge and tools in the studio, tests them in a simulator and can only publish a version that passes its evaluation suite (by default, 90% of checks and 4 out of 5 on tone and clarity) and, if it handles sensitive data, has an approved privacy impact assessment. Every published version can be rolled back.

Every action goes through a single gateway, the tool gateway, which enforces the verification level the action requires (ID document plus an additional detail, a one-time code, or both) and applies the autonomy you set per tool: act, approve or suggest. In the Scheduler, for example, canceling an appointment comes set to approve. Every tool call is recorded in the session trace.

Before anything reaches the model, names, ID numbers and phone numbers are replaced with placeholders, and whatever is stored is encrypted. The IPS sets a monthly AI budget with an alert and a stop. Reports compare people and AI with the same formula, plus cost per interaction, and supervisors can take over any AI conversation.

Today the agents handle inbound calls, WhatsApp, web chat and email, and outbound AI campaigns go out over WhatsApp. The context of the appointment center is in the guide to call centers for IPS, everything else is on the platform page, and to see it with your own cases you can book a demo.

Applicable regulations

Law 1581 of 2012: health data is sensitive (art. 5) and processing it requires, as a general rule, explicit authorization (art. 6). When requesting authorization, you must state the purpose and that answering questions about sensitive data is optional (art. 12). Omcall gives the data processing notice before recording a call and in the first written reply; its privacy policy explains the rest.
Circular Externa 002 de 2024 (External Circular 002 of 2024) of the SIC: when an AI system may pose a high risk to data subjects, it requires conducting and documenting a privacy impact assessment before designing the system, with a description of the processing, the risks and the measures.
Resolución 1552 de 2013 (Resolution 1552 of 2013): an appointment booked by the AI must record the same dates as one booked by a person.
Ley 2300 de 2023 (Law 2300 of 2023): collection or commercial campaigns run by the AI follow the same hours and frequencies as those run by your team.

The legal texts, emergency phrases and deadlines that Omcall ships with by default are drafts to review with your legal and clinical advisors: Omcall provides the mechanisms, it does not certify compliance.

Frequently asked questions

Yes, as long as it is connected to the real schedule and verifies identity before booking. The agent answers, understands the reason in natural language, finds slots, offers options and books, and if something is unclear it transfers the call to a person with a summary. In Omcall, booking requires verification with a one-time code or a higher level, and the appointment stores the data Resolution 1552 requires: request date, desired date, assigned date and the site's REPS code (its entry in Colombia's registry of health service providers).
Yes, they should know from the first message and always have the option to talk to a person. It is a matter of trust and also of data protection: Law 1581 requires telling data subjects what their data will be used for and that answering questions about sensitive data, such as health data, is optional. In Omcall, the greeting that says "virtual assistant" is a fixed rule your IPS cannot remove, and the patient can ask for a person at any time.
The agent stops the conversation, gives the 123 emergency line and immediately hands the case to a person. It gives no clinical guidance, not even to reassure the patient. In Omcall, detection runs before the model responds, using phrases your IPS can configure, and the handoff enters the emergency queue with higher priority. The default phrases are a draft for your clinical team to review and complete with the terms your patients actually use.
It asks for the ID document and an additional detail only the patient knows, or sends a one-time code, and that raises the verification level of the conversation. Each action requires a level: searching for slots needs no verification, but viewing appointments or booking does. If the conversation lacks the required level, the system rejects the action, and the AI cannot bypass that rule. This matters because Law 1581 classifies health data as sensitive.
With the same metrics as your team, calculated with the same formula, plus a few specific to AI. The shared ones are service level, ASA, AHT and FCR; in Omcall, a case counts as resolved on first contact if the person does not come back within 7 days (adjustable). The AI-specific ones are resolution without a person, escalation rate, cost per interaction and tool errors. In addition, each version must pass test scenarios such as an emergency or a manipulation attempt before it is published.
No. AI takes over repetitive tasks with clear rules, and people remain for what requires judgment: emergencies, complaints, cases that do not fit and patients who prefer to talk to someone. In Omcall, people and AI share the same queues, and the staffing forecast separates the volume the AI will handle from the volume that will reach your team, so you can size shifts with Erlang C. That way you decide how many people you need based on data, not assumptions.
Omcall is billed as a license per agent per month plus AI and channel usage, such as voice minutes and messages. The price depends on how many people handle contacts, how many sites and channels you connect and which AI agents you turn on, so it is sized in the demo with your real operation, and that produces a specific figure, not a range. AI usage has a monthly cap set by your IPS, with an alert and a stop. IPS that join as design partners get a 30-day pilot at no cost. You can book the demo.