Solving real-time Icelandic voice, then bringing Menni's AI platform in-house
We took Menni from a no-code MVP to a fully owned, multi-tenant AI support platform that answers 100% of website chat and 70%+ of phone calls in over 100 languages, including fluent real-time Icelandic that off-the-shelf voice AI could not handle.
See it live at menni.ai
Challenges and solutions
Making Icelandic voice work on a live call
The challengeSpeech-to-text transcribes Icelandic words well enough on its own. On live, two-way calls it fell apart: duplicated text, only one side of the conversation captured, or the second speaker's words delivered all at once at the very end. A phone agent that mishears the caller while the call is happening is worse than no agent.
What we shippedWe fixed real-time multi-channel Icelandic transcription so both sides of a live call arrive cleanly and in the right order. That reliability is the whole product. It is what let the agent hold a natural conversation, and it drove Menni's rapid local adoption and the funding that followed.
One brain behind chat, phone, and email
The challengeTour operators, car rentals, hot springs and hotels field the same questions thousands of times, across three channels, mostly outside office hours. Answers have to be real, and they have to actually do things. A bot that invents availability or cannot change a booking creates work instead of removing it.
What we shippedWe built three agents that share one knowledge brain, grounded in each client's own content and wired into live systems like the Bókun booking platform. Each shipped as its own product, so a client can start with chat and add phone or email later without changing anything underneath.
Stopping a no-code bill that grew with every conversation
The challengeTo reach the market fast, the first product leaned on a hosted AI service for agent logic and Make.com for the automation flows. Both bill per conversation and per operation. As volume grew, the monthly infrastructure cost climbed with it and became a permanent drag on the business.
What we shippedOnce Menni had traction, we rebuilt the entire AI layer in-house: the agent runtime, knowledge retrieval, agent configuration, and the automations, all moved onto infrastructure Menni owns. The recurring bill collapsed to a small fraction of the hosted stack, and the vendor lock-in went with it.
Keeping every client's knowledge current, on its own
The challengeClients keep the source of truth for their policies, price lists and procedures in Google Drive, and they edit it constantly. Keeping each knowledge base in step meant manual uploads and per-operation Make.com flows, which cost money on every run and quietly went stale between them.
What we shippedWe built a Google Drive integration into the platform. Menni connects a client's Drive folders once, imports them in the background, and keeps them in sync as documents change, with bulk folder sharing so onboarding a client is one step instead of many. The knowledge base now maintains itself, and Make.com came out of the path entirely.
Turning a shared spreadsheet into a product clients pay for
The challengeClients viewed their data through Airtable. To a paying customer it looked and felt like a spreadsheet: rows of raw records, no sense of a conversation, no real analytics, and a clumsy experience for something they were being billed for.
What we shippedWe replaced it with two custom multi-tenant dashboards on Next.js and Supabase with Clerk-enforced role-based access. Clients get readable conversation threads, analytics by channel and language, support tickets and a knowledge-improvement loop. Menni's own team got an admin control plane for provisioning organisations, agents, keys and phone numbers.
Before and after
- BeforeVoice AI that misheard live Icelandic callsAfterBoth sides of a live call transcribed cleanly and in order
- BeforeRepetitive calls and chats absorbed by staff, in office hoursAfter70%+ of calls and 100% of website chat handled by AI, 24/7
- BeforeAgent logic rented from a hosted AI platformAfterAn AI runtime Menni owns, runs, and extends on its own
- BeforeKnowledge bases updated by hand and by per-operation flowsAfterClient Google Drive folders imported and kept in sync automatically
- BeforeClients reading their own data as Airtable rowsAfterA multi-tenant product with sessions, analytics and support built in
- BeforeInfrastructure cost climbing with every conversationAfterA fraction of the previous recurring bill, and no vendor lock-in
Three agents, one brain
Menni ships as three AI agents that share a single knowledge brain: the client's own content, kept continuously up to date and connected to live systems like the Bókun booking platform. Each agent is its own product with its own deep dive further down this page.
- Chatbot AI answers on the website, Messenger, WhatsApp and more, and completes bookings through Bókun inside the chat.
- Phone AI picks up the phone 24/7 in fluent Icelandic, handles the majority of calls end to end, and forwards the rest to staff.
- Email AI triages the inbox, prioritising, categorising and drafting replies, and routes anything sensitive to the right person.
Making Icelandic voice work in real time
This is the piece nobody else had got right, and the reason Menni took off locally.
Icelandic is a low-resource language, but the blocker was never vocabulary. Speech-to-text handles Icelandic words reasonably well. The failure showed up only on live, two-way calls, where the transcript is being produced while the agent has to reason about it. Text duplicated. One channel went missing. The caller's half of the conversation arrived in a single burst after they had hung up. Every one of those breaks the agent's picture of what is happening.
We fixed real-time multi-channel Icelandic transcription, so both sides of a live call are captured cleanly and in the right order. That is the difference between a demo and a phone agent customers actually trust, and it is exactly where off-the-shelf voice AI falls over in Icelandic. It drove Menni's rapid local adoption and the funding that followed.
Knowledge that updates itself, straight from the client's Drive
An AI support agent is only as good as the content behind it, and that content moves. Prices change, tours get added, policies get rewritten. Menni's clients do all of that in Google Drive, not in a CMS.
Early on, keeping each knowledge base current meant manual uploads plus Make.com flows that charged per operation. It was expensive at volume and it drifted out of date between runs.
We built Drive support directly into the platform instead:
- 1Connect once, at the folder level
A client's Drive folders are connected in one step, with bulk sharing so onboarding does not turn into a per-document chore.
- 2Import in the background
Documents are pulled in as a background job, so a large Drive does not block anyone and progress is visible while it runs.
- 3Stay in sync as things change
The platform tracks what it has already ingested and picks up changes incrementally, so the knowledge base follows the client's edits instead of waiting for someone to remember to re-upload.
The practical result: clients keep working the way they already work, the agents stay accurate, and another per-operation vendor came out of the stack.
From Airtable to a real product
The client dashboard turns raw logs into something a business owner actually wants to open:
- Every customer session and interaction as a readable thread instead of spreadsheet rows, with full detail on each one.
- Analytics by channel, language and category, so clients see what their customers actually ask.
- Support tickets clients can raise directly from the dashboard.
- Analytics they can download, or have emailed to them on a schedule.
- A knowledge-base improvement loop: flag a wrong answer, get it fixed, re-tested, and closed.
The admin dashboard is the command centre for Menni's own team: provisioning organisations, spinning up chatbot, phone and email agents per channel, managing API and provider keys, phone numbers, communication channels, and the whole knowledge-base pipeline.
Menni went from an MVP stitched together with no-code tools to a proprietary, multi-tenant platform handling a large share of customer support for service businesses across Iceland: 100% of website chat, 70%+ of phone calls, 100+ languages, 24/7 across chat, phone and email.
Solving real-time Icelandic voice drove viral local adoption and helped Menni raise funding. The infrastructure rebuild turned a runaway monthly cost into a durable margin advantage. And because everything is owned, Menni keeps extending the platform without asking a vendor for permission.
The three solutions we built
Menni.ai runs on three AI agents, each shipped as its own product. Open any one for a self-contained deep dive.
Technology stack
The Devioand team combined deep technical expertise with total ownership of our problem, building creative custom solutions for complex requirements that integrated seamlessly with our existing systems. Their communication was proactive and transparent at every step, and they consistently went above and beyond. We felt like a valued partner, not just a client, and I highly recommend them.
Is your AI stack renting what it should own?
We took Menni off a per-conversation platform and onto infrastructure they control, and made voice work in a language nobody else had solved. Tell us where your product is stuck.