Cuppa Call: a morning phone call from an AI companion, sold as a gift
The question. Will grown up children in the UK pay 99 pounds to give their mum three months of daily phone calls from an AI companion, when they first hear of it in an ad?
what I built
Rosie, a voice agent with a warm British voice. She rings an older person every morning for a short chat and always says she is an AI. The family gets a text afterwards. Around her I built a page with a recorded demo call, a number visitors could ring to try her, a gift card to print, checkout links and a form for the buyer's details. I made two short ad films. The page counted opens and button presses on my own server, with no cookie. The part that places the morning calls and sends the family text was not built yet. The calls were sold to start on 1 January.
ElevenLabs Agents · Gemini 3.5 Flash · Twilio · Stripe Payment Links · Meta Ads · Remotion · Python · SQLite · nginx
what happened
- Rosie passed 6 of 12 quality checks in the first test call and 9 of 12 in the second, with no failures. I played the older person myself. No real older person ever had a call.
- The ads ran from 15 to 17 November. 411.71 kroner spent, about 48 pounds, of a 2,600 kroner budget. 323 link clicks at about 1.27 kroner each. The film "Gift list" got 152 clicks and "Written by hand" got 171.
- The page counted 160 opens from 16 November. 69 stayed ten seconds. 6 scrolled to the prices. 3 pressed "hear a call", and one of those was my own test.
- Nobody pressed buy or ring in those 160 opens. Before that count started, one visitor pressed the 99 pound button twice and did not pay.
- On the last evening I made the offer stronger at the same price: a longer refund, a recorded Christmas hello and 20 founding places. 82 ad visitors saw it. 1 pressed "hear a call". 0 pressed buy.
- 0 sales and 0 email sign-ups.
- The sums: a 99 pound sale leaves about 82 pounds after VAT. Three months of daily six minute calls, with the texts and the card fee, cost about 61 pounds. The calls alone cost about 0.13 dollars a minute. About 21 pounds is left.
Cost. About 48 pounds in ads. The voice agent ran on a 6 dollar a month plan during the test.
screenshots




what broke
- The ads did not run for the first 13 hours. They were set to find people who start a checkout, and the page had never recorded one. I switched the goal to link clicks.
- In the first day and a half, a third of the money went to ads inside other apps and games. Six in ten clicks came from people aged 65 and over. I wanted their children. I set a hard age limit and took those placements out.
- When I took a placement out, the ad tool ticked a box by itself that still let about 5 percent of the budget go there. I found it on the second look.
- The ads promised a 99 pound Christmas gift. The page opened with a general headline, and the price came on screen 9 of 13. I rebuilt the top on day 1.
- On a small phone the cookie question was about 118 pixels high and sat on top of the buy button. I made it one row of 56 pixels.
- Rosie still carried a template call flow from when the agent was created. It hung up the moment the caller said goodbye, with no closing line. A typed test chat showed it in two minutes.
decision
Closed on 17 November 2026, eight days before the planned verdict. 323 clicks gave no sale. At about 21 pounds left per sale, I needed one sale in every 140 clicks just to pay for the ads. The ads are paused and the buy buttons are off the page. Nobody had paid, so nothing is owed to anyone.
- Do the sums before the ads. My target was 80 pounds in ads per customer. A sale left about 21 pounds. The test could not pay even if it hit its target.
- Let a few real users try it first. With three to five real families the page would have had proof to show. With none, 91 of 160 visitors left within ten seconds.
- Read who clicked on day 1. The click price looked great until I saw the ages and where the ads had run.
- Count on your own server. The ad tool only sees visitors who say yes to cookies.
- Type to a voice agent before anyone rings it. Ask for a human, ask something it should not know and say goodbye.
- Decide what counts as a no before the first click. I had written down that 0 sales means stop, so closing was easy when the sums came in.
- My guess, which 323 clicks cannot prove: a gift that starts weeks later is hard to sell to a stranger. Next time I pick a problem the buyer feels today.
The same lessons as a how-to: Ad tests: what to check before you pay for clicks.
Every number on this page comes from the project's own logs, analytics or database. Back to the lab.