AI agents in the clinic back office: recalls, collections and stock

AI agents in healthcare are AI systems that carry a task through several steps over time, such as messaging a patient, waiting for a reply and offering a slot, instead of answering one question. In the clinic back office, they are most useful for recalls, collections and stock. Verto offers three ready-made AI roles for these jobs, each working under human approval.

Agent, chatbot or automation rule?

These words are often mixed up. The difference matters when you decide what to trust:

Agents are powerful because they keep going. That is also why they need limits: what they can see, what they can change, and when they must stop and hand over to a person. Our guide to AI governance for clinics sets out those controls.

Job 1: the front desk (recalls and rebooking)

Patients who are due back often slip away. Nobody has time to chase each one. Verto Reception is the AI role for this work.

• Recall chaser. Ask in plain language, for example: “Find the patient I saw last week with diabetic retinopathy and remind them to book.” Verto shows a shortlist and never picks the patient for you. After you approve, it sends a WhatsApp reminder from the clinic's own number, if the patient has consented. It waits for a reply and offers real open slots. If the patient stays silent, it nudges once after 48 hours and hands the task to you after five days.
• Slot rescue. When a cancellation frees a slot, Verto offers it to suitable waitlist patients one at a time, and staff confirm the booking.
• Programs. One instruction can cover many patients. “Dr. Ahmed is sick tomorrow, notify and rebook everyone” becomes one approval card that lists every affected patient and their consent status, then starts one governed recall for each.

Reception can confirm and rebook appointments in your clinic scheduling system. It has no access to invoices, and any change to a patient record always needs a human click.

Job 2: collections (accounts receivable)

Unpaid patient balances and slow payers age quietly. Verto AR Clerk works through them in a set order:

1. A statement to open the conversation.
2. A firmer reminder a few days later.
3. A final notice.
4. If the balance is still unpaid, a hand-off to a person. The chase never runs forever.

If the invoice is paid at any point, the chase stops at once, so no patient is chased after paying. Collection messages go by email and SMS. The AR Clerk can prepare a write-off or payment plan, but it never posts a payment, issues a credit or writes off a balance by itself. Those appear as one-click approval cards, and approved changes can be undone for 24 hours. The same balances sit in Helix billing and accounting, so the chase and the ledger always agree.

Job 3: stock and purchasing

Running out of a fast-moving consumable, or finding a box expired at the back of a cupboard, costs money and patient time. Verto Inventory Controller watches stock levels and expiring or slow-moving lots. When an item drops below its reorder point, it works out the shortfall, net of what is already on order, and drafts a purchase order to the right supplier for your buyer to approve.

If you ask vaguely, such as “we need more of that filler that ran out last week”, it does not guess. It shows a short list of matching products and asks you to confirm the exact one. It never places an order on its own. Stock, lots and expiry live in Helix inventory and purchasing, on the same record that bills consumables from the chart.

What AI agents are not good at

Agents work best on repetitive, well-defined jobs with clear data. They are weaker, and should not be used, where the job depends on judgment or relationships:

• Sensitive conversations, such as a disputed bill after a poor outcome, or a patient who is upset. These belong with a person.
• Clinical decisions about who needs to come back and how soon. The clinician decides; the agent only chases.
• Messy data. If phone numbers, consent flags or reorder points are wrong, an agent will act on the wrong information quickly. Clean the data first.
• Unclear ownership. If nobody picks up what the agent hands back, tasks stall in a new place instead of the old one.

The rules every Verto role follows

• Least privilege. Each role sees only the data and tools its job needs.
• Money and clinical actions are always a human click. No setting, tuning or track record can change this.
• Consent and contact limits. Every patient message is consent-checked. A per-patient cap counts AI messages and marketing campaigns together, and each patient has only one active AI conversation at a time.
• A pause switch per role. Pausing Reception does not affect the AR Clerk. If Verto cannot confirm a role is active, it treats it as paused.
• Honest reporting. Roles report back when a task is finished, failed or handed to you. A program that is only partly done says so.
• Audit and undo. Every action is logged in an immutable audit trail, and approved actions can be undone for 24 hours.

How to start

1. Pick one job. Recalls are usually the easiest win, because the value is visible and the risk is low.
2. Name an escalation owner. Someone must pick up what the AI hands back.
3. Check your consent data. Agents can only message patients who have agreed. Clean consent records make the AI more useful.
4. Review the approval inbox daily for the first few weeks.
5. Watch the role's measures. Reception tracks no-show rate and appointment fill rate. AR Clerk tracks collection rate and outstanding receivables. Inventory Controller tracks stock value and expiring lot value. All are calculated from your own records.
6. Add the next role once the first one is routine.

Collections and stock are two of the places where money leaks out of a clinic without anyone noticing. Our article on revenue leakage in clinics looks at the others.