The questions people actually ask us.

Can an AI receptionist book appointments?

Yes. It reads live calendar availability, offers real slots, books the caller in, and confirms by SMS or email.

It also reschedules and cancels, and can offer a freed slot to people on a short-notice list. For clinics this is usually the whole reason to buy one: the booking happens on the call rather than becoming a message somebody has to action later.

Where a direct write into your system is not possible, it books into a synced calendar and posts a structured summary for your team to confirm.

Does an AI receptionist replace a human?

No, and we would not sell it that way. It answers every call so nothing is missed, and frees your people for the calls that need judgement.

A skilled receptionist is still better at reading a difficult situation. What an AI agent does is make sure the phone is never unanswered, which is where most businesses quietly lose customers.

You define the escape hatches. It can transfer to a human, take a structured message, or book a callback. It should never guess, and we build explicit fallbacks so an edge case becomes a warm handoff.

Will callers know they are talking to an AI?

Modern voice models sound natural. We still recommend a short, honest disclosure at the start of the call.

It builds trust, it is expected or required in a growing number of jurisdictions, and in practice it does not reduce booking rates.

What happens if the AI cannot answer something?

It escalates rather than guessing. Transfer to a human, a detailed message, an SMS follow-up, or a booked callback.

Which of those happens is your decision, set per topic, per caller or per time of day. The failure mode we design against is an agent inventing an answer, so anything outside its brief becomes a handoff.

How much does an AI voice agent cost?

Two parts: a one-off build fee from $997 and a monthly cost from $199 covering telephony, the model, hosting and support.

The monthly figure moves with call volume and how many systems we connect to. You get a fixed quote after a short discovery call, not hourly billing.

The useful comparison is not against other software, it is against the value of the calls currently going unanswered. Our missed-call calculator does that arithmetic with your own numbers.

How long does a build take?

A single-purpose agent for reception, booking or lead capture is usually live in two to three weeks.

More complex builds with several call paths and deep CRM integration run four to six weeks. We start with a free demo agent so you hear the quality before committing to anything.

When is an AI voice agent not worth it?

When your call volume is low, or when almost every call needs genuine human judgement.

If you take a handful of calls a week and answer nearly all of them, the maths does not work and we will say so. The same is true for work where the first conversation is the sale and rapport decides it.

Our calculator says this out loud: below a certain volume it tells you an AI receptionist would not pay for itself on missed calls alone.

Will it work with my existing phone number?

Yes. We can forward your existing number to the agent, or port the number over entirely.

Most clients start with conditional forwarding, where the agent only picks up if nobody answers within a few rings, then move to full-time once they trust it.

Does it integrate with my CRM?

Yes. HubSpot, Salesforce, GoHighLevel, Pipedrive, Zoho, and most systems with an API.

Industry systems count too: Clio and MyCase for legal, the major dental practice-management platforms. Where a system has no API we usually still get there through webhooks, Zapier or a middleware layer.

Every call produces a structured record rather than a voice memo somebody has to listen to twice.

Can AI handle legal intake without giving legal advice?

Yes, and the boundary is designed in explicitly. It gathers facts and books consultations. It never assesses merits.

It is instructed not to comment on likely outcomes or limitation periods. Anything approaching advice triggers a handoff to a qualified person.

For conflict checks it collects the identifying details you specify and flags anything questionable for a human decision, rather than clearing or declining a matter itself.

Is n8n cheaper than Zapier?

Not always. Below roughly 14,000 operations a month a hosted tool is cheaper and simpler.

Self-hosting has a broadly flat cost, so it wins as volume rises and loses at the bottom end. It also wins where data cannot leave your infrastructure, or where the logic outgrows a visual builder.

Our cost calculator compares all three at your actual volume, and tells you to stay on Zapier when that is the right answer.

Will it understand callers with strong accents or a noisy background?

Usually yes, and it is the first thing we test. Speech recognition handles regional accents well now. Heavy background noise is the harder problem.

The realistic failure is not the accent on its own, it is a bad line plus a building site plus somebody speaking quickly. We test against recordings of your actual callers before launch rather than against a clean demo, because that is where the difference shows.

Where it cannot follow somebody after a couple of attempts, the right behaviour is to stop trying and put them through. An agent that keeps asking a frustrated caller to repeat themselves does more damage than one that gives up early.

How many callers hang up when they realise it is an AI?

Some do, and anyone telling you none do is selling. The number that matters is not hang-ups, it is how many of those callers you were reaching before.

The comparison people make is against a human answering instantly. That is rarely the real alternative. The real one is usually a voicemail nobody returns, or a ring-out while you are on a job, and a caller who hangs up on an AI was already lost in that scenario.

It is still worth measuring rather than assuming. Call outcomes can be reported so you can see hang-up rate against booking rate and decide with numbers.

We tried an AI phone system before and it was awful. Why would this be different?

Usually because the first one was a phone menu with a nicer voice. The test is whether it can finish the job on the call rather than take a message.

The common version answers, collects a name and number, and emails it to you. That is an answering machine with extra steps, and callers work it out within a sentence.

What actually makes the difference is unglamorous: whether it reads your live availability, whether it writes the booking back into your system, whether it hands off cleanly when it is out of its depth, and whether anybody maintains it after launch.

Can a caller skip the AI and get straight to a person?

Yes, and it should be easy. Asking for a human should transfer the call, not start a negotiation.

We set an explicit escape route. A direct request, repeated confusion, or any sign of distress ends the automation and routes the call: to your team in working hours, and to a message or whoever is on call outside them.

A system that traps people is the fastest way to lose the customers you were trying to catch.

What stops it inventing a price or promising something you do not offer?

Constraints, not good intentions. It works from a fixed set of facts you approve, and is instructed to say it will check rather than guess.

Prices, lead times, service areas and policies are supplied as data rather than left to the model. Anything outside that set gets a straight "I will have somebody confirm that" and a task for your team, which is what a sensible new starter would say anyway.

We then try to break it before launch: talk it into quoting, discounting, and committing to work you do not do, and tighten it until it stops.

Can it get an email address or a postcode right over the phone?

Yes, with read-back. It repeats the value and has the caller confirm it before saving.

Getting this wrong quietly is worse than failing loudly. A mistyped email means the confirmation never arrives and nobody finds out until an appointment is missed, so read-back is standard on anything that has to be exact.

Where a value can be checked automatically it is: postcode format, and the obvious mishearings in common email domains.

Which calls should we never let it handle?

Anything turning on clinical, legal or financial judgement, and anything where somebody is distressed. Those should reach a person quickly.

The agent should recognise the situation and route rather than attempt it. Diagnosis, advice on a live matter, a complaint that has escalated, and anybody upset are all straightforward handoffs.

Agreeing that list is part of the build. Being specific about what the agent will not do is more useful than claiming it handles everything.

How do I test it before trusting it with real customers?

Ring it yourself, then have your most impatient colleague ring it. Keep it on overflow or after-hours only until you are satisfied.

A staged rollout is the honest way in. Conditional forwarding means it only picks up calls that were going to voicemail anyway, so the downside is capped at what you were already losing.

We also read real transcripts with you in the first weeks. Nothing exposes a weak agent faster than seeing what it actually said.

How would I know if it is losing bookings a human would have won?

By reading transcripts and comparing outcomes, not by watching a dashboard of answered calls.

Answer rate always looks good, because the agent picks up everything. The number worth watching is calls that ended with no booking and no callback task, and those transcripts tell you whether it was the caller or the agent that gave up.

Better to look at that honestly in the first month than to discover it in the sixth.

Should it answer every call, or only the ones we miss?

Start with only the ones you miss. Move to full-time later if the transcripts justify it.

Overflow-only is the lower-risk setup and captures most of the value, because the missed calls were the leak. Your team keeps the relationships they already handle well.

Some businesses do move to answering everything, usually where the front desk is genuinely overloaded and consistency matters more than familiarity. That is a decision to make on evidence rather than at the start.

How does it tell an emergency from something that can wait?

By rules you approve, not its own judgement. You define what counts as urgent and what happens when it does.

For a plumber that might be water that will not stop; for a clinic, specific symptoms. The agent screens against your list, and anything matching goes straight to your on-call route instead of into a booking flow.

It errs towards escalation. A false alarm costs a phone call. The opposite mistake costs considerably more.

Do I have to tell callers they are talking to an AI?

In some places yes, and it is worth doing everywhere regardless. A short line at the start costs nothing and removes the argument entirely.

A growing number of jurisdictions require disclosure in particular circumstances, California’s bot-disclosure law being the best known. The rules differ by state and are still changing, so building on not saying it is a poor bet.

In practice callers mind far less than owners expect. What annoys people is working it out late, not being told at the start.

Does the call have to say it is being recorded?

It depends where your callers are. US federal law allows one-party consent, but several states require every party to consent, so announcing it on every call is the safe default.

California is the most cited example and it is not the only one. Since you cannot control which state somebody dials from, one announcement applied to all calls is simpler and safer than branching on area code.

In the UK and EU the question is framed differently, as a lawful basis and a retention period under UK GDPR, but the practical answer is the same: tell people, and keep recordings only as long as you can justify.

Is an AI receptionist HIPAA compliant, and will you sign a BAA?

There is no such thing as HIPAA certification. What matters is whether every party touching patient data signs a Business Associate Agreement.

HHS does not run or endorse a certification scheme, so a vendor badge claiming one is marketing. The real question is the chain: your practice, us, and the voice and telephony providers underneath. If any link will not sign a BAA, the arrangement does not work whatever the badge says.

That genuinely constrains which platforms we will build on for a healthcare client. It is a constraint worth accepting rather than working around.

Is it a robocall if the AI rings a lead back after they enquired?

Returning a call to somebody who just asked you to is normally fine. Working through an old or bought list with an AI voice is where you get into trouble.

The FCC ruled in February 2024 that AI-generated voices count as artificial under the TCPA, so the consent rules for prerecorded calls apply to these too. Penalties run per call with no cap, which makes this a poor place to guess.

The practical line is consent and purpose. Somebody who gave you their number asking about a job has consented to be contacted about that job. A list of numbers from two years ago has not, and marketing calls carry a stricter written-consent standard again. If you want outbound campaigns, have it reviewed properly first.

If the AI tells a customer something wrong, are we liable?

Assume yes. It is speaking as your business, and "the system said it" has not worked as a defence for anybody else.

This is the main reason to constrain what the agent can assert, and to keep prices, policies and lead times as approved data rather than something it improvises on the call.

Keeping transcripts helps. If a dispute happens you can see exactly what was said, which is a better position than reconstructing a phone call from memory.

If you build the agent for us, who is the data controller?

You are. Handing the build to an agency does not hand over the accountability.

You decide why the calls happen and what is done with the data, which is what makes you the controller under UK and EU rules. We act as a processor, and the platforms underneath us are sub-processors.

That is not a technicality to skip. It decides who answers a subject access request and who is answerable if something goes wrong, so it belongs in writing before launch rather than after an incident.

If a caller asks us to delete their recording, can we delete it everywhere?

Only if retention was configured deliberately at every layer. A flat yes is usually an overpromise.

A call can leave traces in the telephony provider, the speech and model providers, your CRM and any backup. Deletion is only real if each of those holds nothing or honours the request, and that is a build-time decision rather than a switch found later.

The honest version is to decide up front what is kept, where, and for how long, and to keep as little as the business actually needs.

Will it write the booking into our system, or leave a message for somebody to type in?

It writes in, where the system allows it. Ask any supplier which one they mean, because "integrates with" gets used for both.

There is a real difference between an agent that creates the appointment and gets a confirmed record back, and one that sends your team a tidy note to re-key. The second is still useful, but it is not automation and should not be priced as though it were.

What is possible depends on the platform. Systems with an open API take a genuine two-way booking. Older or closed practice-management software sometimes allows only a one-way write, or nothing, and we will tell you which yours is before you buy.

How would I know if a booking never made it into the calendar?

Because it should alert somebody. A write that fails silently is the failure that actually hurts.

The dangerous case is not an error on a screen, it is a booking the caller was promised that never landed, discovered when they turn up. A failed write should be treated as an incident: the call flagged, somebody notified, the task handed to a human rather than disappearing.

Reconciliation matters too. Comparing what the agent believes it booked against what is actually in the calendar catches the quiet drift that a green dashboard never will.

Will it create duplicate contacts every time somebody calls?

Not if the matching rule is set properly. Matching on phone number alone is what causes the mess.

A household sharing a number, a caller ringing from a different phone, and a mistyped email all defeat naive matching. The rule has to say what counts as the same person, what happens on a partial match, and when to create a new record rather than guess.

Worth deciding explicitly before launch. Cleaning duplicates out afterwards is far more work than agreeing the rule up front.

Does it know it is talking to an existing customer?

Yes, if it can look them up. That is a CRM lookup on the incoming number, not memory inside the AI.

The distinction tells you when it will fail. If the caller withholds their number, rings from a different phone, or is not in the system yet, there is nothing to find, and the agent should handle that gracefully rather than pretend.

When the lookup does work it is the difference between "can I take your name" and "is this about the boiler we looked at in March", which is most of the perceived quality.

We have a CRM nobody uses. Will connecting the phone to it fix that?

Partly. It removes the data-entry excuse, which is real. It does not make anybody open the CRM.

Integration fixes the input problem: calls, enquiries and outcomes land in the system without somebody typing them. That genuinely helps, because manual logging is the step that gets skipped when it is busy.

What it does not fix is a CRM nobody has a reason to look at. If it is not where your team decides what to do next, better data just produces a tidier system nobody opens. That is a process question, and we would rather say so than sell an integration that quietly changes nothing.

Will it book two jobs an hour apart on opposite sides of town?

Not if it books through a calendar that knows about travel time and territory. A plain free-slot lookup will absolutely do this.

Finding a gap in a diary is the easy part and everything can do it. Respecting drive time, which engineer is qualified, what is on the van, and slots deliberately held for a bigger job is what makes a booking useful rather than a problem for whoever runs dispatch.

If your scheduling software already models that, the agent books through it and inherits the rules. If it does not, we constrain the agent instead, which means fewer bookable slots but no impossible days.

How would I know if an automation quietly stopped working?

You would not, unless it is monitored for silence. The dangerous failure is not the one that errors, it is the one that succeeds with nothing in it.

An expired token, a renamed field or a trigger that stops firing produces a green run and no output. Nobody notices until somebody asks why the report stopped arriving, which is usually weeks later.

We alert on absence as well as errors: a workflow that should have run today and produced nothing is itself the alert. It is the single most useful thing to add to an automation somebody else built for you.

What happens to my phone line if the AI provider has an outage?

It should fall through to you rather than fail. That is a routing decision made at setup, and worth asking about before you sign.

Uptime percentages are the wrong question. The useful one is what the phone does in the bad minute: ring your mobile, reach a human service, or take a message. Configured properly the caller notices little more than a different greeting.

Because your number is forwarded rather than replaced, you keep the ability to switch the whole thing off and go back to ringing your own handset.

Who owns the agent and the automations if we stop working together?

You should. Ask before you start, because the usual default is that everything sits in the agency’s accounts.

The test is concrete: whose name is on the platform accounts, who holds the phone number, and can you export the workflows. If the answer to all three is the agency, you are renting something you paid to have built.

We set things up in your accounts where the platform allows it, and hand over credentials, the number and the workflow definitions on request. That is a deliberately low-lock-in position, and it is the question we would ask any supplier.

Can it follow our on-call rota without changing settings every week?

Yes, if the rota lives somewhere it can read. Hard-coding one number is what turns it into a weekly chore.

Demos show a single transfer-to-the-owner number. Real businesses rotate, swap at short notice, and have a different rule for commercial customers and bank holidays.

The workable version points the agent at a shared calendar or rota sheet you already keep, so changing who is on call happens where your team already does it.

Should it just text people back instead of answering the call?

Often, yes. A missed-call text is cheaper and simpler, and it works well for anything that is not urgent.

A text that goes out seconds after a missed call recovers a good deal of the loss for very little money or complexity. For some businesses that is the whole answer, and we will say so.

Voice earns its place when the caller needs an answer now: an emergency, a booking they want confirmed on the call, or somebody ringing three suppliers who will instruct whoever picks up. Those are the calls a text does not save.