AI Voice Phone Ordering — Why Bulgarian and Greek Are the Wedge

Phone orders are still 20–40% of restaurant revenue across most of Europe. The volume is concentrated in the same hours your kitchen is busiest. The result is a familiar pattern: Friday at 19:30, three cooks, the phone rings, nobody picks up. That call was a €40 order. Multiply by every Friday for a year and the missed-revenue number gets uncomfortable.

AI voice ordering exists to close that gap. The pitch is simple: an AI agent answers every call, takes the order against your menu, sends a payment link by SMS, drops the order on your KDS. Your staff never stopped what they were doing. The catch — until now — was that the agents available on the market only spoke English.

Why most US-built voice agents fail in Europe

Maple, Slang.ai, Toast Voice, Popmenu, HungerRush — all marketed for American restaurants, all assume the caller speaks English, all start at $399/month and require a US-based POS. For a restaurant in Sofia or Athens, none of that fits. Bland AI gates multilingual to enterprise contracts. The market gap for Bulgarian and Greek phone ordering was wide open.

Most pure-play voice ordering vendors charge $399+/month with English-only agents. Ordering.Tools voice ordering ships native Bulgarian and Greek at every plan tier from day one.

How we picked the voice stack

We evaluated four serious bundled options: Google Gemini Live, OpenAI Realtime, ElevenLabs Conversational AI, and AWS Nova Sonic. Three criteria mattered: native BG and EL voice quality, end-to-end latency under 1.2 seconds, and per-minute cost that lets us hit ~50% gross margin at the entry tier.

Gemini Live with Chirp 3 HD voices won on BG and EL quality. The product owner blind-tested both Google and ElevenLabs on Bulgarian — Google sounded native, ElevenLabs sounded close. Cost is competitive at ~€0.10–0.15/min. Function calling is first-class so our existing tool endpoints — search-menu, add-to-cart, confirm-order — work without modification.

But locking ourselves to one provider was a non-starter. So we built the agent provider as an interface from day one. Default is Gemini Live; ElevenLabs and OpenAI Realtime are one config flip away — per venue, no redeploy. If Google raises Gemini prices, we move a single venue to ElevenLabs Agents and the rest of the system does not notice.

The telephony question — your number stays yours

A common worry: do we have to give you a new phone number? No. The telephony layer is also pluggable. Three ways the call can reach the AI:

• •Twilio DID — we provision a fresh number in your country, you publish it where you like.
• •SIP trunk — your existing PBX (Asterisk, 3CX, FreePBX, Cisco) points a SIP trunk at us. You keep your number and operator.
• •Conditional forward — your existing number stays exactly as it is; only calls nobody picks up forward to our AI.

Conditional forward is what most established restaurants want. Their staff still picks up first. Only overflow goes to AI. The €49/month entry tier becomes pure additional revenue, not a process change.

Server-side reprice and the safety cap

Two things that have to be right or the agent becomes a liability: pricing and order limits. We solved both server-side, never trusting the LLM.

When the agent calls confirm-order, the server re-fetches every product from the database, re-applies variants and modifiers, and recomputes the total in cents. If the LLM read out €19.50 but the menu says €19.90, the server uses €19.90. The customer never gets a wrong bill from a hallucinated price.

The safety cap (maxOrderValue per venue) is the second guard. If the AI ever tries to confirm an order above this number — say €200 for a small pizzeria — the server rejects and the agent transfers to a human. This catches both prank orders and edge cases where the LLM misunderstood a quantity.

EU AI Act compliance is mandatory, not optional

Article 50 of the EU AI Act becomes effective 2 August 2026. Any AI system that interacts with humans by audio must clearly disclose that it is AI. We hardcoded the disclosure as the agent's first utterance, per locale, with no toggle to disable it. GDPR call recording consent is announced in the same opening line.

This is a feature, not a constraint. Restaurants in EU markets will increasingly be audited for compliance. Slang.ai, Maple, and Toast — even when they add multilingual support — will need to retrofit this. We shipped it from day one.

Why we ship five free calls per plan

Voice ordering is a leap of faith. Restaurant owners want to hear the agent on their own menu, in their own language, before paying. So every plan tier — Free, Starter, Pro — gets 5 AI voice calls per month at no extra charge. Pure experimentation allowance. Plenty for a restaurant to test the pickup flow, hear the BG voice, see the bill land in admin.

When billing arrives in Phase 2, you keep the same config and pick a tier. Until then, a superadmin can lift the limit per venue without changing the plan — useful for pilots where you want to give a friendly venue a few hundred calls without billing complexity.

How to try it

In admin, open Settings → Voice Agent. Pick a provider (default Gemini Live), pick a connector (default Twilio DID — but conditional forward is the easier path if you already have a number), set a max order value and handoff number, click Save and provision. Place a test call. Listen to the agent on your own menu.

If something does not sound right, every call has a full transcript and every tool call logged. The debug story is built in. We want to hear what works, what does not, and which tools you wish existed. Each missing tool is a clear signal for what to add next.