Odysi builds WhatsApp AI assistants for customer service, what many call a WhatsApp AI agent. The assistant answers your customers' repeated questions from your own rules and documents, collects and checks the photos and documents needed to open a request in your system, and hands the conversation to a person when a case needs judgement. We build and test it for you: it is not a bot builder you configure yourself. It runs on your WhatsApp Business number through Meta's official API, where customer service assistants are allowed; nothing is installed on your website. We work fully remote, in Spanish and English, with businesses wherever they are. It is live in production at Fixlab, a vehicle certification lab in Spain.
- Vehicle homologation labs and engineering firms: complete files before anyone signs
- Vehicle testing stations and test centres: what to bring, booking, retests and results, on the number you already publish
- Driving schools in Spain: enrolment, lessons, DGT exams and "when is my test?", without promising dates
- River outfitters and rafting companies: guest questions answered by text, WhatsApp, web and email
- Gyms and CrossFit boxes: member messages, freezes, cancellations, failed payments and coach cover
- Ski resorts: lift status, tickets, rentals and lessons, by text, WhatsApp, web and email
The problem is volume, not difficulty.
Most inbound support comes in across scattered channels (WhatsApp, email, phone, your platform), by text, voice note and photo. The questions look involved, but most follow a pattern and could be answered in seconds. Instead they sit in a queue.
The volume pulls skilled people away from the work only they can do. First replies take a day or more, worse in peak season. Customers chase across channels and get asked for the same thing twice. The slowest businesses simply lose them to whoever answered first.
A reply inside 5 minutes reaches and qualifies a lead at a rate that collapses within the hour.
Sources: MIT / InsideSales 15,000-lead study; Lead Connect, 2020.
Four things it does.
None of it is magic. Each is a specific task your team does by hand today, done the same way, at any hour, and handed straight back the moment a case needs a person.
Answers from your real knowledge base
It replies using your own criteria, not generic guesses. The source of truth stays in the model's working context, so answers are consistent instead of drifting over time.
Starts the paperwork
When a customer is ready, it identifies them against your platform and opens the request directly through your API, capturing the details you need to act. No re-keying, no dropped context.
Understands more than text
Customers send photos and voice notes, not tidy paragraphs. It reads images, transcribes audio, and groups rapid-fire messages into one coherent turn instead of replying to fragments.
Knows when to call a human
When a case needs a person, it escalates into your team's tools with the full conversation attached. The human takes over, the customer never notices a seam, and the assistant resumes with context intact.
It runs in two layers.
A messaging middleware sits at the front and handles everything that isn't conversation. Behind it, an autonomous agent runs the dialogue and decides what each message needs.
Messaging middleware
Layer 01 · the frontConversational agent · multi-model
Layer 02 · the dialogueThe route a message takes
Seven stationsWe're not tied to one model. We use Claude, Gemini and ChatGPT, and pick the right one for each job. When a conversation is dense with conditions and exceptions, choosing the right model is what keeps answers correct instead of plausible-sounding. We use several because it keeps the answers right, not because a longer list of model names looks impressive.
If the agent doesn't respond in time, the customer gets an apology and the case is escalated to a person automatically. Network calls retry on transient failures.
The point was never a clever demo. It's something that holds up on a Monday morning, with real customers waiting.
What it handles, and what goes to a person.
The important property isn't that it answers everything. It's that it knows the difference and escalates rather than guessing.
Try it on your own WhatsApp.
Pick your kind of business, then write in the phone the way your customers write to you. See how an assistant built for you would answer, what it would do next, and what it would need from you.
This demo needs JavaScript. Book 30 minutes and we will run it with your real messages.
We agree the test before we ship.
A conversational system is easy to demo and hard to trust. So we don't ship and hope. Before launch, we agree a written acceptance test with you: a golden dataset of real scenarios covering every conversation path: identification, knowledge questions, request creation, escalation, and resolution.
The bar is explicit: zero critical failures and a pass rate you sign off on. It goes to production only after it clears that bar, and after the full cycle, escalation to a human and back, is validated end to end. That's how you turn “it seems to work” into “it passed.”
We built this for Fixlab, and it's live with their customers.
Fixlab is a vehicle-certification business in Spain whose support was scattered across WhatsApp, email, phone and their platform. The assistant answers from their real knowledge base, files reform requests through their platform, and hands off to a person when a case needs judgment. It cleared the acceptance test we agreed up front, and it's in production.
“We recommended you for your work, but above all for the kind of people you are.”
Translated from Spanish
“We know Fixlab’s system because we work with them. Congratulations on the work: the way it handles WhatsApp enquiries is impressive.”
Translated from Spanish