A voice agent that recommends beard oil is a sales channel wearing a support badge
Ten employees, no live phone line, and a working voice agent assembled in roughly a day. That is the setup behind Beardbrand’s experiment, and the interesting part is not the build time. It is what the bot actually does once a caller is on the line. Read the published transcript again and notice that nobody asks for a refund, a tracking code or a returns label. The caller asks which product suits a two month old beard, and the bot runs a five question discovery conversation before naming two scents and recommending beard oil over balm for that length. That is not customer service. That is a sales assistant with a phone number.
The distinction matters because it changes how you measure the thing. Support bots are judged on deflection rate and cost per ticket. An agent that steers a hesitant shopper towards a specific SKU belongs in the funnel, measured on assisted conversion, average order value and the share of callers who place an order within forty eight hours. If you deploy one and only track how many calls your team no longer has to take, you will undercount it badly and probably kill it during the next budget round.
What the transcript reveals about the questions shoppers cannot ask your product page
Look at the sequence the bot follows. Length of the beard. Texture, dry or coarse. What the caller currently uses. Scent preference, framed as a closed choice between fresh, woody, spicy and sweet. The caller answers “resin, dark, mellow” and gets two matching products with the actual notes named, Palo Santo and frankincense for one, fir balsam for the other.
Every one of those questions is a gap in the product detail page. A filter for scent family would answer the fourth one. A guided selector would answer the first three. Most webshops have neither, which is why the product page for a sixteen euro oil reads like a specification sheet and the shopper leaves to read reviews elsewhere. The voice agent is doing the work that a well built product finder should already be doing on the site, and it is doing it for a population of shoppers who picked up the phone precisely because the site did not get them there.
So the first thing to take from this is diagnostic rather than technical. Log the calls, transcribe them, and cluster the questions. If forty per cent of callers want to know which variant suits them, you have a merchandising problem, not a staffing problem. The voice agent becomes a research instrument that tells you what to fix on the page, and the transcripts are far richer than a session recording because the shopper says the quiet part out loud.
Where these agents break, and the two guardrails you need before launch
The published conversation ends mid sentence with the bot saying it will check which product is most popular. That cut off is a useful place to pause. The moment an agent reaches into live data, stock levels, bestseller rankings, order status, delivery promises, it stops improvising charmingly and starts making claims you are liable for. A recommendation of a scent is subjective. A statement that an item ships tomorrow is a commercial promise, and a wrong one produces a refund, a chargeback dispute and a bad review.
Guardrail one is therefore a hard boundary on what the agent may assert. Product attributes, usage advice and routing are safe. Stock, pricing, delivery dates and order specifics should come from a live lookup or not be answered at all, with a graceful handover rather than a guess. Guardrail two is disclosure and logging. Beardbrand’s bot opens with “Thanks for calling the Beardbrand VoiceBot on a recorded line”, which handles two obligations in one breath: the caller knows it is a machine and knows the call is recorded. Under the EU AI Act’s transparency rules, telling someone they are interacting with an AI system is not a courtesy, and the recording consent sits under the GDPR. Get both into the first six seconds of the greeting, the way that transcript does.
There is a third consideration that gets overlooked. A phone channel is an accessibility channel. Shoppers who struggle with a dense product grid, who use a screen reader on a page that was never tested with one, or who simply prefer speech, now have a route to a purchase. That cuts both ways under the European Accessibility Act. A voice agent does not excuse an inaccessible site, but it does serve a group your checkout may currently be losing silently.
What to run first if you want to test this without burning a quarter
Start narrow. Pick the single highest volume pre purchase question you receive by email and give the agent that one job, with a phone number placed on the relevant category page rather than buried in a footer. Set a measurement window of four weeks and compare three numbers against the previous period: orders attributed to callers, average order value for that group, and the return rate on items the agent recommended. That last one is the honest test. An agent that pushes the wrong variant inflates revenue for a month and then gives it back in returns and in trust.
Then decide what the agent is allowed to do after the recommendation. Reading a basket link by voice is useless. Sending an SMS with a prefilled cart, or a short link to a checkout with the recommended items already in it, closes the loop and removes the retyping step where most of these conversations leak. The handover from voice to checkout is the weakest joint in the whole design, and it is the part that the demo transcripts almost never show.
One more point on skill level. The author rates himself a four out of ten as a developer, familiar with APIs and DNS, unable to write code. That is the real signal here. The build is no longer the barrier. The barrier is the merchandising logic you feed the thing, the boundaries you draw around what it may claim, and the discipline to measure it as a revenue channel rather than a cost saving.
Source: practicalecommerce.com
