Live demo
WhatsApp Pilot
A WhatsApp assistant wired to your data, in a day.
Answers customers, qualifies requests, triggers your automations and hands over to a human when it should. On the channel your customers already use.
Try it onlineAn assistant that knows your procedures, not generalities found online.
It answers your clients' and your teams' questions from your own documentation, guides them through a procedure, qualifies a request, and hands over to a human when it must — with the context already gathered.
- Answers sourced from your content
- Escalation with context
- Widget, API, WhatsApp or internal
Why chatbots have a bad reputation
It is almost never the model at fault. It is what it is wired into, and what it is allowed to say.
01
It does not know your content
Wired into a twenty-line FAQ written two years ago, it answers beside the point by the third question. The user then hunts for the “talk to a human” link, and support has gained a ticket rather than avoided one.
02
It invents rather than admits
With no explicit rule for what to do when it does not know, an assistant produces a plausible, wrong answer. Once is enough for the team to lose confidence and switch it off.
03
It cannot hand over
When it escalates, it forwards the raw conversation. The human agent starts again from scratch, and the assistant has lengthened the delay instead of shortening it.
From the question to the answer, or to a human
The path is the same for an external customer and an internal colleague: only the corpus and the rights change.
01 / 05
Understanding the request
The assistant identifies the real intent behind the wording, detects the language, and recognises the cases that must go straight to a human — complaints, emergencies, sensitive subjects — without attempting an answer.
02 / 05
Searching your content
It queries your documentation, procedures, product catalogue and, where needed, your systems via API: the status of an order, a case, a contract. The search respects the rights of the person asking.
03 / 05
Answer or explicit refusal
If the information exists, it answers and cites its source. If it does not, it says so — that is a rule defined with you, not a configuration accident. A traceable “I don't know” beats an invented answer.
04 / 05
Qualification and guidance
On requests that call for an action, it asks the missing questions one at a time, checks what it receives, and walks through the official procedure in the order it must be followed.
05 / 05
Informed escalation
When a human takes over, the assistant hands across a case file: the request restated, what it verified, the relevant documents and what remains to be decided. The human picks up where the assistant stopped.
What it takes on
The perimeter is decided with you, subject by subject. Anything outside it goes to a human.
Tier-1 questions
Repetitive requests — hours, procedures, status, conditions, how-to — absorbed without human involvement, with the source cited.
Product and document search
Finds the technical sheet, the contract clause, the current version of a procedure, even in a large and badly organised corpus.
Step-by-step guidance
Walks through a procedure in order, checks each step, and only moves on once the previous one is confirmed.
Querying your systems
Status of an order, ticket, case or contract, fetched live via API, with the user's rights.
Request qualification
Identifies the subject, the urgency and the relevant team, and gathers the missing information before handing over.
Multilingual
Answers in the language of the question, from a corpus that may be written in another.
Wherever your users are
One assistant, one knowledge base, several entry points. You are not maintaining five assistants.
API
Called from your existing application or software.
Widget
Embedded in your site or application, without leaving the page.
Messaging
WhatsApp and compatible messaging environments.
Platform
Dedicated interface, with team and rights management.
Business software
Embedded directly in the tool your teams already use.
Internal or external, the same engine
01 · Context
Support handles two thousand requests a month, a large majority of them already documented questions.
02 · What we put in place
An assistant wired into the knowledge base and the order-tracking API, as a widget on the site and on WhatsApp, with informed escalation to the team.
03 · Result
Tier-1 requests are absorbed. The team handles the cases that deserve a human, with the file already prepared.
01 · Context
HR and IT questions arrive by email, chat and in person. The same answers get rewritten every week.
02 · What we put in place
An assistant wired into internal procedures, the collective agreement and the IT knowledge base, with isolation by department: an employee only sees what concerns them.
03 · Result
Recurring questions get an immediate, sourced answer, available outside office hours.
01 · Context
Website visitors ask precise technical questions the contact form does not answer.
02 · What we put in place
An assistant wired into the product documentation, able to qualify the need and then offer a meeting once the request is mature.
03 · Result
Technical questions get an immediate answer, and the sales rep receives an already-qualified contact.
What we measure at pilot
Set at scoping, measured on your real conversations, not on a hand-picked test set.
- of tier-1 requests handled by a human, on the covered perimeter
- -60 %
of tier-1 requests handled by a human, on the covered perimeter
- availability, including outside your team's working hours
- 24/7
availability, including outside your team's working hours
- of answers carrying their source, verifiable in one click
- 100 %
of answers carrying their source, verifiable in one click
These orders of magnitude come from what we measure at pilot on comparable perimeters. On your corpus they are measured before industrialisation, not promised before.
What you can see running today
These products are already built. Some are available online, the others are shown on request.
Frequently asked questions
Your content: documentation, procedures, catalogue, contracts, and where needed your systems via API. Nothing else. It does not answer from general knowledge about your sector.
The cases where it must say “I don't know” are defined with you, subject by subject. Every answer cites its source. And the error rate is measured at pilot on your real conversations.
It inherits the rights of the person asking. An external customer sees only the public corpus; an employee, what their role allows. The isolation is your systems', not a layer bolted on beside them.
It answers in the language of the question, even if your documentation is written in another. That matters when a French corpus has to serve English- or Arabic-speaking users.
The assistant passes the request restated, what it verified, the relevant documents and what remains to be decided. That is the difference between a useful assistant and one more filter.
It depends mostly on the state of your documentation. On a clean corpus and a well-defined perimeter, the pilot runs within weeks. On a scattered corpus, the cleanup is the real project.
What this solution does not do
An assistant will not compensate for missing or wrong documentation. If your procedures are written down nowhere, there is nothing to query — we then start by building the document layer, and we say so at scoping. Nor do we configure an assistant to handle complaints, disputes or sensitive subjects: those conversations go to a human, by explicit rule.
Which questions do you answer again every week?
Thirty minutes is enough to work out which share of your support can be automated, and above all which share should not be.