The short version: I put an AI voice agent on my business line for 30 days, it handled 71% of calls without me touching them, and it saved me roughly five hours a week. It also lost me a client, and the reason it lost me that client is the bit nobody selling these tools will tell you upfront.
Why I even tried this
I was answering my own phone between client calls, school pickup, and writing proposals, and I hated it. Not the talking to people part, the interruption part. Someone would ring about a speaking enquiry while I was mid-sentence on a strategy call, I'd miss it, they'd try once, maybe leave a voicemail, and half the time I never heard back. I lost count of how many enquiries just went cold because I didn't answer fast enough.
So in January 2026 I set up a voice AI agent built on a real-time conversational model, the kind of thing platforms like Bland AI and Synthflow are built on, and pointed my business number at it for a full month. Not a test with a burner number. My actual line, my actual clients, my actual reputation.
What the AI did
The setup took about four hours spread over two evenings. I wrote a script, not a rigid one, more a set of rules: greet callers, find out if they're an existing client or new enquiry, book a call in my diary if it's the latter, take a message with urgency level if it's the former, and hand off to my mobile if someone says a specific trigger phrase like "I need to speak to Lilach directly."
Over 30 days:
- 214 calls came through the line
- 152 of those (71%) were handled start to finish by the AI with no human involvement
- Average AI call length was 2 minutes 40 seconds versus my usual 6 minutes on the same type of enquiry
- The tool cost me £340 for the month including the phone number routing and per-minute usage
- It booked 19 calls straight into my calendar without me lifting a finger
On the numbers alone, this was a win. I got back close to five hours a week, which over a month is a full working day I didn't have to spend on the phone.
Where it started to fall apart
Here's the bit that people selling these tools skip. The AI wasn't caught out by accents, or bad phone lines, or people talking fast. It handled all of that better than I expected. It fell apart on nuance, specifically on the calls where the caller wasn't being straightforward about what they wanted.
In week three, a long-standing client rang about what she described as "a quick question on the invoice." The AI, doing exactly what it was told, treated this as routine and offered to log a message for me to call back within 24 hours. What she meant was that she'd been double-charged, was annoyed, and wanted to hear from me that day. The AI had no way of picking up the tension in her voice, because it doesn't listen for tension, it listens for keywords and intent categories. By the time I called her back the next morning, she'd already emailed to cancel her retainer, worth £4,200 a year. She told me later the thing that tipped her over wasn't the billing mistake, it was that a machine had decided her problem could wait.
That's the uncomfortable truth about voice AI in customer-facing small business right now: it's brilliant at volume and terrible at emotional weight, and most small business owners set it up to handle both because they don't realise there's a difference until they've lost someone over it.
The disclosure problem
I told the AI to identify itself as an AI assistant at the start of every call, because from November 2026 several US states require disclosure by law and I'd rather build the habit now than retrofit it later. About one in six callers commented on it, usually with mild surprise, sometimes with a flat "oh, I'll just email then." A small number hung up. I don't know how many of those were lost enquiries versus people who'd have hung up on a human too, and neither does anyone selling you this software, because that number is hard to track and conveniently absent from every case study I've read.
If you're building anything similar, I'd rather you knew that going in than found out from a complaint.
What I changed after the first month
I didn't rip it out. I narrowed its job. Now it only handles new enquiries and appointment booking, both low-stakes, transactional, first-touch interactions where a caller has no existing emotional history with my business. Anything from an existing client gets routed straight to my mobile with a text alert, no AI layer at all. That one change, splitting new business from existing relationships, is the single most useful decision I made in the whole experiment, and it's the thing most guides on AI phone agents don't mention because they're written by the software companies selling the tool for every use case at once.
Since narrowing its scope, I've had zero client complaints and the same 70%-ish call handling rate on new enquiries. The AI is doing less, and doing it better, which is a pattern I keep seeing across AI tools generally, not just phones.
Where an interactive tool beats a voice agent entirely
For a chunk of what I originally wanted the AI to do, a phone call was the wrong channel altogether. Quote requests and "how much does this cost" enquiries don't need a conversation, they need a self-serve answer. I've since moved a good part of that traffic to a pricing tool on my site instead, and it converts better than the phone ever did, partly because people can play with it at 11pm without feeling like they're bothering anyone. If you're weighing up voice AI against something simpler, it's worth reading about how interactive calculators convert visitors before you spend money on a phone bot for the same job.
What it costs, broken down
People ask me "how much does an AI voice agent cost" as if there's one number. There isn't. Here's what I paid across the month:
- Platform and per-minute usage: £340
- Setup time, valued at my own consulting rate of roughly £150/hour, four hours: £600
- A second setup pass after the client complaint, tightening the scope: two hours, £300
- Total real cost for month one: roughly £1,240, dropping to around £340/month from month two onward with no more setup needed
Compare that to what I was doing before, which was nothing, just me answering the phone badly between meetings and losing enquiries I never even knew about. There's no clean control group for "enquiries I lost because I was too slow," which is exactly why so many small business owners underestimate how much a bad phone process is already costing them before they ever touch AI.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
The tone question nobody tests
Voice is brand. I spent longer than I'd like to admit picking the AI's voice and pacing, because a flat, over-formal tone made my business sound like a call centre, and an overly chatty one felt fake given what I do. Brands that get this right treat voice and tone as a design decision, not an afterthought, in the same way Duolingo's marketing built a whole personality around one owl and Calm built theirs around slowing everything down. If you're curious how deliberate that kind of tone-setting looks when it's done well, both the Duolingo marketing strategy and the Calm marketing strategy are worth a look, because the discipline is the same whether it's an app or an AI answering your phone: know exactly who you sound like before you let a machine speak for you.
Watching what happens on the calls
I only caught the invoice call problem because I review call transcripts weekly, not monthly. That habit, watching real behaviour rather than trusting a dashboard summary, is the same discipline behind tools like Hotjar, which built its entire product on the idea that summary metrics hide the moments that matter. The Hotjar marketing strategy is worth reading if you want to understand why watching raw sessions, or in my case raw transcripts, beats trusting a percentage on a screen.
I also connected the AI's booking output into my calendar and CRM using simple automation rules, the kind of if-this-then-that logic that made IFTTT's whole marketing strategy work for years before AI made "automation" a buzzword. New booking triggers a confirmation text, a calendar hold, and a task for me to prep. None of that needed to be complicated, and I'd caution against overbuilding this part before you've even tested whether callers trust the AI in the first place.
Would I recommend it
For high volume, low emotional stakes, first-contact enquiries, yes, without much hesitation. Booking calls, answering opening hours questions, taking basic details, all of that is well within what a voice AI handles cleanly in 2026. For anything touching an existing relationship, a complaint, or money someone's upset about, keep a human on it, because the cost of getting that wrong isn't a bad review, it's a client walking with no warning at all.
If you're weighing this up for your own business and want a second opinion before you spend anything, that's exactly the kind of decision an AI consultant for small business should be helping you scope, not the software vendor's sales page.
Related reading: Bad days and how to cope with them.
Frequently asked questions
Do AI voice agents sound robotic in 2026?
No, and that's part of the problem, not a solved one. The best ones sound close enough to human that callers only realise it's AI when told, which is why disclosure rules and honest expectations matter more than voice quality does.
How much does an AI voice agent cost for a small business?
Expect £300 to £400 a month in usage fees for moderate call volume, plus setup time that typically runs four to eight hours if you're doing it yourself, or a one-off fee if you pay someone to build it.
Will an AI answering the phone put customers off?
Some will hang up once told they're speaking to a machine, and there's no reliable way to measure how many enquiries that costs you. It matters far more on emotionally loaded calls than on simple booking or information requests, which is why splitting new enquiries from existing client calls is worth doing from day one.
Should I tell callers they're speaking to an AI?
Yes. Several US states now require it by law, and trust erodes fast if a caller works it out themselves partway through a call. Building disclosure in from the start costs you almost nothing and avoids a much bigger problem later.
Related reading: The Best Free Social Media Scheduler in 2026 (And What "Free" Costs You) and How to Get Your Small Business Quoted by ChatGPT (Not Just Ranked by Google).
I go much deeper on this in the AI marketing guide.
