Using a voice assistant to capture missed calls

What shipped

Voice AI is now practical enough for a small business to test on a narrow phone workflow: answering routine calls, taking details, asking agreed questions, and sending a structured note to the right person. This is not the same as replacing reception. The useful version is decision support, with a person reviewing the call note before anything important happens.

For many Irish SMEs, the problem is simple. Calls arrive during lunch, after 5.30, when a van is on the road, or while the receptionist is dealing with someone at the desk. Some callers leave a voicemail. Many do not. If the business depends on bookings, callbacks, quotes or first enquiries, those missed calls have a cost.

A voice enquiry assistant can answer with a short approved script. It can capture the caller's name, phone number, email, location, preferred date, type of request, urgency, and any notes. It can then create a call summary or booking request in the system the team already uses, such as a shared inbox, CRM, booking spreadsheet, practice system, job management tool, or ticket queue.

The key design choice is the handover. In a sensible SME workflow, the AI does not confirm a high value job, make a promise on price, or deal with sensitive issues on its own. The named human step is the Receptionist Review and Confirmation step. A receptionist, sales person or manager checks the transcript and extracted fields, fixes anything unclear, and then confirms the booking or callback.

This is the kind of workflow Artellis can help map, test and improve through workflow design services. The work is less about chasing a shiny phone bot and more about deciding which calls are safe to structure, what data should be captured, where it should go, and when a human must step in.

Who should care

This is worth looking at if your business gets regular inbound calls that follow a pattern. Examples include clinics, trades, training providers, accountancy firms, garages, salons, property services, childcare providers, professional services firms, and B2B sales teams.

It is most useful where staff spend time on the same first five questions. What do you need? Where are you based? When would suit? Have you used us before? Is it urgent? Those questions are necessary, but they do not always need a skilled staff member to type them from scratch.

It is also useful where out of hours calls matter. A plumbing firm may want to separate emergencies from routine quote requests. A clinic may want to capture cancellation requests without confirming anything until staff review them. A training company may want to qualify course enquiries and route them to the right coordinator. A small manufacturer may want to capture spare parts enquiries and create a structured note for the service desk.

Managers should care because the value can be measured without a long project. Useful measures include calls captured outside office hours, percentage of call notes accepted without rework, average time from call to callback, number of enquiries lost to voicemail, and staff time spent writing basic call notes.

There are limits. A voice assistant should not handle complaints, medical or legal advice, complex pricing, HR matters, or vulnerable callers without a quick human handover. If a caller is angry, confused, distressed, or asking for something outside the script, the assistant should stop trying to complete the flow and create an urgent human review task.

For more examples of where AI fits into everyday SME operations, see the Artellis insights library.

A worked example

Take a 35 person home services company in Cork. It handles maintenance, installations and emergency callouts. The office is open from 8.30 to 5.30, Monday to Friday. Calls also come in during evenings and weekends, especially when something breaks.

At the moment, all missed calls go to voicemail. On Monday morning, the office team listens back, writes notes into a spreadsheet, checks the diary, and rings people back. Some callers have already gone elsewhere. Others did not leave enough detail, so the team has to chase basic information before deciding what to do.

A first pilot does not need to cover every call. It can start with five call types:

1. New booking request
2. Emergency callout request
3. Existing booking change
4. Quote request
5. General callback request

For each call type, the team writes a short script. For a booking request, the assistant asks for name, phone number, email, address or Eircode, type of job, preferred date, and whether access is available. For an emergency callout, it asks safety and urgency questions agreed by the manager, then flags the note as urgent.

The assistant answers on a test number first, not the main public number. Staff make trial calls and try normal, messy conversations. They interrupt it, give partial answers, use local place names, and ask questions it should not answer. The goal is to find where the workflow breaks before real customers use it.

When the pilot moves to real calls, each call produces three outputs: the transcript, a short plain English summary, and structured fields for the booking or callback queue. Nothing is treated as confirmed until the Receptionist Review and Confirmation step is complete.

In that review step, the staff member checks the caller details, listens to the audio if needed, corrects any field, and chooses one of four actions: confirm booking, call back, route to a specialist, or mark as not suitable for AI handling. This gives the business control while still reducing the admin load.

After 20 pilot calls, the review should be practical rather than theoretical. How many notes were usable? Which questions caused confusion? Did callers object to the assistant? Were urgent calls flagged correctly? Did staff save time, or did they spend longer correcting poor notes?

If the answer is positive, the business can widen the call types or connect the workflow to the live system. If not, it may still learn enough to improve call scripts, voicemail prompts, website forms, or staff routing.

Where I'd start

Start with the call log, not the technology. Pull a sample of recent enquiries and group them by reason for calling. If you cannot name the top five call types, do that first. A voice assistant needs a clear job.

Next, write the minimum useful script for each call type. Keep it short. Ask only for information the team will actually use. If a field is not needed to decide the next action, leave it out of the first version.

Then define the handover rules. Which calls are safe to summarise? Which need urgent review? Which should never be handled by the assistant? Who reviews the note each morning, and how quickly should callbacks happen?

Run a small pilot on a test number. Aim for 20 real or realistic calls. Review every transcript and every extracted field. Track accepted notes, corrected notes, failed calls, and staff comments.

Only then decide whether to connect it to the main number or booking process. The right outcome is not a fully automated phone line. The right outcome is fewer missed opportunities, cleaner call notes, faster callbacks, and a human still making the operational decision.