Context
A small business misses a call. Thirty seconds later that customer is talking to a competitor.
That is the entire business case for voice AI, and it is also the reason most voicebot projects fail: they optimise for sounding human instead of for never losing the call.
AgenciGlosowi.pl is my implementation service for exactly this problem. It is not a self-serve product you sign up for: every agent is designed, built, tested and launched for one specific business, inside their own phone line, calendar and CRM. The clients are companies that live and die by the phone - clinics, workshops, salons, real estate offices, restaurants, hotels.
The offer is deliberately done-for-you, because the businesses that need this hardest are the ones with no IT department to hand a platform to. This case study covers what the engagement looks like: the problem as owners actually experience it, the handful of decisions that decide whether a deployment works, and what generalises to any AI implementation.
The problem: two pains, one invoice
Talk to any owner of a service business and you will hear the same complaint about the phone. But it is really two problems wearing one coat.
Unanswered calls are lost revenue. The phone rings hardest exactly when nobody can pick it up: peak hours, when staff are with customers on site. A small team will never staff for the peak. Outside office hours everything goes to voicemail, and almost nobody leaves one.
Unconfirmed bookings are abandoned bookings. A surprising share of people book online and then call to check the booking is real. If nobody answers to reassure them, they quietly cancel. This one is invisible in every dashboard, because the booking simply never converts.
Then there is the cost nobody counts: no record of who called. No contact, no reason for the call, no way to follow up. The customer relationship never starts.
The market context is not subtle either:
- 66% of consumers say they express emotion most through the phone
- 46% of callers hang up after four IVR menu options
- 89% of customers want the option to reach a person
- Gartner expects 25% of support interactions to start with an AI agent by 2027
- The voicebot market: $8.7B in 2025, projected at $66B by 2035
People still call. They just refuse to be put through a phone tree.
What I build for each client
Narrow and reliable beats broad and impressive. Every deployment is scoped to the part of the call that never needed a human:
- Answers 100% of calls immediately. Hundreds of concurrent conversations, no queue, no voicemail, any hour of the day.
- Books appointments. Checks real availability, reserves the slot in the calendar, sends an SMS confirmation during the call.
- Qualifies leads. Asks the right questions for the industry, scores the intent, routes hot leads to a salesperson.
- Writes everything to the CRM. Number, name, reason for the call, a summary and an intent tag: booking, FAQ, sales lead.
- Hands off to a human, with context. When the case goes off-script, the call transfers with the full conversation history attached.
The scope is the same everywhere; the scenarios, the knowledge base and the integrations are rebuilt per client. A real example of what “done” looks like, from a real estate deployment: a caller asks about a two-room flat from a listing, gets availability, picks Saturday at 12, gives their name and number, and hangs up with a confirmed viewing and a financing detail already noted for the agent. Total time: 1 minute 12 seconds.
The four decisions that actually matter
Every engagement comes down to these four. Everything else is implementation detail.
1. Latency is the whole illusion
Under ~500 ms average response time or the conversation stops feeling like a conversation. Voice gives you none of the buffer that chat does: there is no “typing” indicator, no moment to think. A pause reads as a broken system.
That constraint shapes everything upstream. Tool calls - checking a calendar, looking up a listing - have to happen mid-sentence without a dead pause. Best-in-class voices (ElevenLabs) matter, but they are worth nothing behind a two-second gap.
2. The LLM handles language, not facts
This is the decision I would defend hardest.
Prices, available slots and booking confirmations are computed deterministically from the client’s own systems. The LLM is responsible for one thing: turning that into natural Polish. It never invents a number.
A voicebot that quotes a price it made up is not a rough edge, it is a liability - and when you are the one who deployed it, it is your liability. Splitting the system this way keeps the critical data as correct as the source system, and confines the failure modes of the language model to phrasing.
3. 500+ scenarios before a single real customer
The engagement is roughly 25% building and 75% testing.
Wiring telephony, calendar, CRM and knowledge base is the fast part. What takes the time is bad connections, background noise, regional accents, people interrupting mid-sentence, and the genuinely strange things callers say. Each one gets run, watched, fixed, and run again, against the client’s actual scenarios rather than a generic script.
This is the step most voicebot projects skip, and it is exactly the difference between “works in the demo” and “still works on the fiftieth call.” It is also the main reason this is a service and not a template: the testing is where the value is, and it cannot be self-served.
4. There is always a plan B
89% of customers want a route to a human. Fighting that number is a losing strategy.
So the agent is built to recognise when a case is off-script and transfer it, with the whole conversation visible to whoever picks up. The customer never repeats themselves. The bot starts, the human continues.
The feature people actually want is not a bot indistinguishable from a person. It is the certainty that they can reach a person when it matters.
Positioning: situations, not features
The insight that shaped the whole offer: nobody buys a voicebot. They hire it for a situation where they are currently losing time, money, or patience.
So the service is sold as situations, in the end customer’s own words:
- “I’m calling after hours and I want to book, not talk to voicemail.”
- “I’m calling at peak time and I don’t want to sit in a queue.”
- “I booked online and I want to be sure it wasn’t a mistake.”
- “I’m annoyed and I want to be understood, not transferred in circles.”
- “My case is unusual and I want a human without repeating everything.”
Each one maps to concrete agent behaviour, and to a scenario that gets built and tested during the engagement. This is a better way to sell AI in general: describe the moment the customer is in, then show what changes.
The Polish angle is a real moat
Most competitors in this space are a translated landing page over an English-first platform. Being a genuinely local implementation partner is hard to fake:
- Voice and language understanding built for Polish, not translated into it
- EU hosting, minimal data collected (name, phone, appointment time), a GDPR data processing agreement included in the package
- Polish-language support, a Polish VAT invoice, and a Polish counterparty on the contract
For an SMB owner deciding whether to put a machine in front of their customers, that last list is not a footnote. It is often the deciding factor - and it is a reason to buy an implementation rather than a licence from abroad.
The engagement: no IT department required
The client has no engineering team, so the deployment cannot assume one. Everything happens on my side. Four steps:
- Map the process. Analyse real calls, decide what the agent handles alone.
- Train the agent. Voice, tone, knowledge base for the industry. Test on real scenarios.
- Connect the systems. Calendar, CRM, telephony. The client keeps their existing phone number.
- Go live. Monitor quality from day one and keep optimising for conversion.
The client keeps working in the tools they already have: Google Calendar, HubSpot, Salesforce, Pipedrive, Booksy, Twilio, Slack, Zapier, Make, WooCommerce, Notion, Microsoft 365, plus an open API for anything else. Nothing gets migrated, nothing gets replaced.
Commercially it is priced as a service with an ongoing component, because an agent that is never retuned degrades: a one-time implementation fee, a monthly subscription covering support and optimisation, and call minutes by usage. No hidden costs, three tiers, from a single location up to multi-site networks with their own API integrations.
One thing that stopped being a problem
Since 2 August 2026 the EU AI Act requires callers to be told they are speaking with AI. Hiding it is not an option.
It turns out not to matter. Once people know it is AI, they stop expecting a human and start expecting speed. That changes what you optimise for: no fake breathing, no acting, just an agent that understands intent, never loops the same unhelpful line, and always captures the request even when it cannot handle it.
The result
A deployed agent answers every call, including after hours. No lead is lost, because each one lands in the client’s CRM with a name, a number and a reason. And clients see +34% more booked appointments.
That last one is the only metric worth reporting back to a client. “Calls handled” is a vanity number. Booked appointments is the business.
What it proves
Availability beats quality more often than anyone admits. The competitor winning your customers is usually not better at the job. They just picked up.
Split the deterministic from the generative. Let the model own language and let your systems own facts. Almost every embarrassing AI failure in production comes from blurring that line.
Narrow ships. An agent that answers, qualifies, books and hands off does more real work than a general-purpose assistant that dazzles in a demo. It also ships in weeks, because most six-month AI projects spend five of those months deciding what to build.
The hard part is not the technology, it is the implementation. Voice models are a commodity now. What a business actually cannot do itself is map its own call scenarios, wire the agent into systems it barely understands, test it against 500 ways a real conversation goes wrong, and keep tuning it after launch. That is the work, and that is what the service sells.