Fatora Bot

Invoicing in Morocco costs more than the invoice is worth

Moroccan micro-businesses invoice from carbon-copy books, or not at all. When they do buy software, they buy something designed for an accountant: a per-seat licence, a data-entry screen, a training session, and a salesperson who will never open it from a building site or a shop counter.

Article 145-IX of the General Tax Code established electronic invoicing on a clearance model, where the invoice passes through a government platform before it reaches the customer. The timetable for the smallest businesses still depends on a decree unpublished as of mid-2026: the deadline is coming, but it is not here. That is precisely the window in which a tool has to be useful for something other than compliance — otherwise nobody adopts it until the night before the law bites.

Fatora Bot attacks the problem through habit rather than through law: nothing to install, nothing to learn, and compliance arriving as a by-product of a gesture the team already performs fifty times a day — sending a WhatsApp voice note.

• Nothing to deploy: WhatsApp is already on every salesperson's phone
• No per-seat licence — price follows invoice count, not headcount
• No training: the interface is a conversation
• Tax compliance obtained without having to understand it
• A head start on the mandate rather than a scramble at the deadline

It speaks the language your teams actually work in

Darija is not a standardised written language. It is spoken, written sometimes in Arabic script and sometimes in Latin, and it counts money in riyal or melyoun rather than dirhams. A generic dictation engine fails on all of that.

Fatora Bot handles the voice note end to end: transcription, structured extraction of the customer, items, quantities and prices, then a spoken reply in the same language. "Sayab liya fatora l Ahmed Transport, tlata dyal l-bibane b 1500 dh" becomes a customer, an item, a quantity, a unit price and a total including tax — in a little over a second. Hesitations and mid-sentence corrections are absorbed: "no, two at 200" rewrites the summary without starting over.

Two less visible mechanisms make the difference in practice. A Darija lexicon built from an open corpus anchors the trade vocabulary. And a regression bench replays real cases on every change to the engine: any error a real user hits is added to it BEFORE being fixed, so it cannot come back.

• Voice note or written message, whichever suits the person invoicing
• Darija in Arabic or Latin script, French, Tamazight
• Always answers in the language received — and in audio when the request was audio
• Riyal and melyoun understood without manual conversion
• Mid-sentence corrections absorbed without starting again
• Conversation memory: the bot knows which invoice is being discussed
• Burst buffer: five voice notes in a row form one request, not five invoices
• Regression bench fed by real errors before any prompt is changed

An enforceable invoice, not a receipt

What comes out of the conversation is a PDF an accountant accepts and an auditor can check. Every mandatory field is there: the company's common business identifier, its tax identifier, its professional tax number, its registered name and address, the line items, VAT by rate, totals before and after tax, and the grand total written out in words.

They are entered once, when the company registers, through six questions asked in the conversation. After that nobody re-keys them. The company stamp is photographed once and applied automatically to every invoice issued by anyone on the team.

Numbering runs unbroken through the financial year, with no gaps — a condition improvised tools routinely fail. An error spotted after sending is corrected under the same number, instead of forcing a credit note over a typo.

• Tax identifiers and VAT regime entered once, never asked again
• VAT itemised by rate, totals before and after tax, amount in words
• Unbroken numbering per financial year, specific to each company
• Company stamp photographed once, applied automatically
• Post-issue correction under the same number, no credit note
• PDFs kept and retrievable by voice: "invoice number 8", "this month's invoices"

The invoice reaches the customer without typing their email

An invoice issued but stranded in the salesperson's phone is worth nothing. Three routes carry it to the customer, each chosen with a single word in the conversation.

The QR code is fastest at a counter: the customer scans the salesperson's screen, or a sticker on the till, and the PDF downloads. Nothing to type, nothing to install on their side. The invoice can also go straight to the customer's WhatsApp, or by email for those who archive.

Every outbound delivery requires explicit confirmation. That friction is deliberate: a bot that sends an invoice to the wrong recipient because it misheard one phrase does more damage than it saves.

• QR code: the customer scans the salesperson's screen, the PDF downloads
• Direct delivery to the customer's WhatsApp
• Email delivery for customers who archive
• Explicit confirmation before anything leaves the company
• Customer directory: exact spelling remembered from one invoice to the next
• Difficult name? A photo of a document is enough — the exact spelling is read from it

Every company in a sealed space

The company's WhatsApp number is the technical boundary. Each business owns its customers, its numbering sequence and its stamp; nothing crosses from one to another. Two neighbouring shops using the same bot cannot see each other.

Conversations run through Meta's official WhatsApp Business API rather than a hijacked client — that is what separates an operable service from a rig that breaks on the next WhatsApp update. Data is hosted in the European Union.

The platform switches from simple processing to a job queue on its own once it runs in production, with no configuration: an end-of-month spike does not turn into lost messages.

• Meta's official WhatsApp Business API — no hijacked client
• One number = one company = one sealed data space
• Numbering sequence specific to each business
• Hosted in the European Union
• Job queue enabled automatically in production
• PDFs generated server-side, with no third-party service in the path

An admin that drives subscriptions and supervision

Subscriptions, activations and supervision run from a dedicated back office, separate from the bot. Activating a company, moving it between tiers or checking what happened in a conversation takes no engineering.

The architecture deliberately splits responsibilities: WhatsApp access tokens never leave the invoicing service. The admin talks to it through a protected interface, so it never has to hold the secrets that would allow writing in a company's name.

• Activation and tier changes without engineering involvement
• Supervision of conversations and issued invoices
• WhatsApp tokens never leave the invoicing service
• Protected access, separate from the public site

A site built to be found, and to be quoted

A product nobody searches for by name gets found through its questions. So the Fatora Bot site carries an editorial core: seventeen guides in French and twelve in Arabic answering what people actually ask — the invoicing mandate, what a Moroccan invoice must show, VAT rates, numbering, archiving, payment terms, the difference between the various tax identifiers.

Those pages are prerendered server-side: their content is readable on the first request, without executing JavaScript — the condition for being usable by generative search crawlers, which do not run pages. The site also publishes the files those engines look for, regenerated on every publication, and notifies search engines of each new URL.

Two guardrails hold it together: the build fails if an internal link does not resolve, and an unknown URL returns a real 404 — previously any invented address returned the home page with a 200 status, which engines read as thousands of duplicate pages.

• 41 prerendered pages, up from 4 before the rebuild
• 17 guides in French, 12 written directly in Arabic
• Sector pages: construction and joinery, retail, transport and logistics
• Comparison pages on decision-stage queries
• Sitemap and AI-engine files regenerated on every publication
• Search engines notified automatically of each new URL
• The build fails on a dead internal link
• Real 404 on unknown addresses, not the home page with a 200
• Self-hosted fonts: no third-party domain called

What comes next

The decree that will set the timetable for the smallest businesses has not been published. When it is, connecting to the government platform becomes the subject: the architecture is already organised so that an issued invoice can be submitted for clearance without the user changing anything about the gesture.

The priority channel is not the micro-business directly but the accountant: firms, professional federations and cooperatives equip their members in one move, with integration handled for them. It is also the channel through which public digitalisation support schemes can fund the rollout.

• Connection to the government clearance platform once the decree is published
• Rollout through accounting firms
• Offers for professional federations and cooperatives
• Continuous enrichment of the Darija lexicon from real usage
• Expansion of the Arabic editorial core