AI receptionists for small businesses in the Netherlands

Count the calls before you buy anything

Almost everybody who asks me about this has a feeling rather than a number. The feeling is usually right and the number is usually different from what they guessed, which matters because the two point at different builds.

Two weeks of logs from whatever you already have settles it. Your phone system can tell you how many calls came in outside opening hours, how many rang out during the day, and how long the average caller waited before hanging up. If the answer is a handful a month, buy a better voicemail greeting and spend the money elsewhere. If it is several a day, you have found the thing worth automating and you now know which hours to aim at.

It also tells you something a vendor cannot: whether your problem is coverage or capacity. Calls lost at 21:00 need a line that works when nobody is there. Calls lost at 11:00 on a Tuesday mean the desk is underwater, and an agent that answers the easy half so a person can take the hard half is a different design from one that runs the night shift.

What it actually has to do

A receptionist agent that only speaks is a voicemail with better manners. Five things separate one that earns its keep, and every one of them is a question you can put to a supplier before signing anything.

• Say what it is, in the greeting. The caller cannot see what they have reached, so the first sentence has to tell them.
• Work from your real calendar. Offering a slot it cannot actually hold is worse than offering nothing, because now somebody has to ring back and take it away.
• Write to the system you already use. If the call does not land in the CRM or the practice system with the caller's details attached, your team is retyping it, and you have automated the pleasant part of the job.
• Know what it does not handle. A booking line asked a billing question should stop rather than improvise, and anything outside its job goes to a person by default.
• Hand over with context. The person taking the transfer should arrive knowing who is calling and what has already been said, or the whole thing has cost time instead of saving it.

The Dutch and European part

Three things come up on every Dutch engagement and none of them is exotic, but they do constrain the build, so they belong in the scoping conversation instead of in a surprise six weeks later.

Personal data first. A receptionist agent collects names, numbers and often a reason for calling, which is personal data under the GDPR, in Dutch the AVG, with the Autoriteit Persoonsgegevens as the authority a caller can complain to. That means somebody has to decide how long recordings and transcripts are kept, who inside the business can open them, and what happens when a caller asks for their data. Those are answerable questions, and they are much cheaper answered before the line goes live.

Where the data goes, second. Speech recognition and language models are frequently hosted outside the European Economic Area, and if EEA-only processing is a hard requirement for you it narrows the tool choice considerably. That is a scoping decision, and a supplier who waves it away is telling you they have not had the conversation before.

Language, third. Dutch callers will often open in Dutch even where the business operates in English, and a line that cannot handle the first ten seconds of that gracefully will lose calls it was bought to save. Decide deliberately what it does when the caller switches language, because the default behaviour of most systems is to plough on regardless.

None of this makes me your lawyer, and it is not legal advice. Where the question is genuinely about the law rather than the system, that is what counsel is for.

When the answer is no

There are calls that should not reach an agent at all, and being clear about them early is what makes the rest of the deployment defensible.

Anything where being wrong is unrecoverable in the moment: medical triage, a caller in distress, a safeguarding concern, an emergency. Those need a human on the first ring and a routing rule that gets them there, well before any confidence threshold is consulted. A line that handles bookings brilliantly and hesitates for four seconds over a frightened caller has failed at the only moment anyone will remember.

The other no is quieter. If your booking process is a mess when a person runs it, an agent will run the same mess faster and with less oversight, and you will have bought a more efficient version of the problem. Fix the process, then automate it.

The cost that is not on the invoice

Build cost is the part people compare, and it is the part that varies least. What varies is the integration: connecting a calendar and a system of record properly, deciding what the agent may and may not touch, and covering the cases nobody thought of until the line was live for a fortnight.

After launch there is a second cost that never appears in a proposal, which is that somebody has to own it. Owning the outcome, which is a different thing from maintaining the software: read the escalations, notice when the agent starts stopping on calls it used to handle, and say when the job should change. A line nobody owns drifts quietly out of usefulness, and the first anyone hears about it is a customer complaining.