Algeria AgencyArtificial intelligence
AI chatbots & virtual assistants
An assistant on WhatsApp, Messenger or your site that answers from your real content, and hands over the moment it leaves its subject.
A demonstration
An assistant answering, on a real site.
An imaginary dental practice, six pages. Its assistant answers in derja, Arabic, French and English — even typed in Latin letters.
Try the demo (opens in a new tab)The practice is fictional: nothing you book there exists.
In Algeria, messaging is the counter. A large share of commercial questions arrives as a private message on Facebook, Instagram or WhatsApp, usually outside opening hours, and usually as the same five questions: your prices, your hours, your address, your lead times, and whether you deliver to Oran. An assistant that handles those five well gives you back your evenings and your Fridays. It does nothing more than that, and that is already most of the gain.
What we build answers from your actual content — pages, catalogue, prices, internal procedures, past conversations — rather than from whatever the model believes about your company. The difference shows in week one: a generic assistant invents a plausible returns policy, while a grounded one says it cannot find that information and passes the conversation on.
The most important part of the work is not making it answer. It is deciding what it must never answer. A quote, a firm delivery date, a discount, medical or legal or financial advice, a complaint: these are commitments made in your company name, and a machine that is right nine times out of ten gets them wrong the tenth. That list is written with you before launch, and the assistant transfers instead of guessing.
Finally, we will sometimes tell you that you do not need us. A business fielding the same question twenty times has a missing page first; writing it costs a day and fixes search engines as well as private messages. We check that before anything else.
What we usually find
- Private messages pile up in the evening and at weekends, and some never get an answer at all.
- One person answers the same five questions twenty times a day instead of selling.
- You tried a consumer tool and it invented a delivery policy you have never had.
- Nobody internally can say what the assistant would really cost per month in peak season.
What changes
A written list of what it will not do
Quotes, firm deadlines, discounts and complaints go to a named person and are never handled alone. Written before launch and visible in the admin panel.
Answers that point at their source
Every answer cites the page or record it came from. A mistake is therefore fixed by correcting the content once, not by rewriting instructions to the model.
A monthly ceiling that holds
Cost per conversation measured against your real volumes, with a limit past which the assistant drops to taking a message rather than carrying on billing.
What you get
Grounded in your content
Your pages, catalogue and procedures indexed, with an explicit refusal when a question falls outside that corpus. Refusing is a feature, not a coverage gap.
WhatsApp, Messenger, Instagram and the site
The channels your customers already write on, sharing one history so that somebody who switches channel does not have to explain everything twice.
Handover without a restart
A transfer carries the whole conversation and states why it happened. A handover that opens with "could you repeat that" cancels the benefit of having an assistant.
Darja, Arabic and French in one thread
Your customers write in three languages, often in a single sentence and often in Latin script. The assistant is tested on your own past conversations, not on clean examples.
A log of unanswered questions
Every question the assistant could not handle is recorded. That is the list of pages you are missing, and after a month it is the most useful output of the project.
Honest opening hours
Outside hours it says so and states when a person will reply. Promising an instant answer at 11pm on a Friday is a way to lose the customer twice.
How we work
Reading your real conversations
We read several hundred of your past messages and classify the questions. That classification sets the scope; without it you build for imaginary questions.
Writing the limit
The list of what the assistant does not handle, approved by you, and the name of the person each case goes to. That document is the commitment, not the demo.
Building and testing on your data
An assistant grounded in your content, then replayed against your old conversations to compare its answers with the ones your team actually gave.
A supervised launch
Two weeks where every conversation is read back, then a sampled review rhythm. The refusal log is reviewed with you monthly.
Is this the right fit for you?
This is for you if
- You receive more private messages than your team can clear the same day.
- Your answers repeat and already exist in writing somewhere.
- Somebody can be named to pick up transferred conversations.
This is not for you if
- You want an assistant that negotiates prices or commits to a delivery date.
- Your product information changes daily and is written down nowhere.
- You are trying to replace a support team rather than give it its evenings back.
What we commit to
No invented answers in production
If the assistant cannot find something, it says so and transfers. We would rather have a high handover rate in month one than a plausible wrong answer.
Your data stays yours
We tell you which provider processes the messages, where, and what is retained. If that does not suit certain conversations, those are excluded from scope.
The cost ceiling is contractual
A monthly threshold set with you, an alert before it is reached, and an automatic fallback to message-taking rather than an invoice that doubles unannounced.
What changes for an Algerian business
The channel is not the website, it is messaging. An Algerian SME commonly receives more qualified enquiries through Facebook and WhatsApp messages than through its contact form, and those enquiries arrive outside working hours more often than inside them. An assistant that only covers the site misses the conversation that matters.
Language is the real technical obstacle, not the model. Your customers write darja in Latin script, switch to French mid-sentence, and return to Arabic for numbers. The demos you will be shown run in clean French; we test against your real conversations before telling you whether this works.
The cost is in foreign currency. Billing is per request to an overseas provider, which makes the monthly ceiling more important here than elsewhere — which is why we measure cost per conversation against your peak-season volumes rather than an annual average.
Frequently asked questions
How long before it is usable?
Expect a few weeks, half of that spent reading your conversations and writing the limit. The technical part is the shortest stretch of the project.
What happens when it does not know?
It says so and transfers to the named person with the whole conversation. Each case is logged and used to write the page that was missing.
Can it take an order?
It can prepare an order and present it for human confirmation. It does not confirm alone: a wrong quantity or address only surfaces at delivery.
Does it understand darja?
Partly, and unevenly by subject. We measure that on your own messages and show you the failure cases before launch rather than after.
What does it cost per month?
It depends on conversation volume, and we price it against your real volumes before starting. A ceiling is set so the invoice stays predictable.
What if we want to switch it off?
The indexed content is yours and stays usable elsewhere. The assistant is switched off in one action and messages go back to your team.
How to start
Send us an export of last month’s private messages, unsorted.
We return the question breakdown, an estimate of how much of it is handleable, and the list of what must not be handed over.
What we have written on this subject
Two quotes for the same chatbot: what actually moves the price
Two suppliers price the same need at one and at five. The gap almost never comes from technology: it comes from five scope decisions.A chatbot that understands darija: what it adds to the work
Darija is not expensive because it is difficult. It is expensive because it is never written the same way twice, and that is settled on your own messages.Chatbot: what it must refuse to answer
An assistant is judged on its refusals, not its answers. What the list has to contain, and what happens when nobody wrote it.Which messaging channel, and what each one imposes
The channel decides the rules, the deadlines and what you are allowed to send. What to know before choosing where to reply.Automation: the queue nobody opens
An automation that works produces cases it cannot handle. They pile up somewhere, and the report goes on showing a comfortable success rate.What you are allowed to send to a model hosted abroad
A chatbot connected to your mailbox exports your customer file, one line at a time. Nobody around the table described it that way.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.
- No invented answers in production
- Your data stays yours
- The cost ceiling is contractual