AI for an Algerian import and distribution business

Distribution businesses do not have an AI problem. They have a paperwork problem and a "only Karim knows where that is" problem, and AI happens to be good at exactly those two things.

So rather than describing what the technology can do, this is a walk through where the hours actually go in a business that imports and resells, and which of those hours are recoverable.

Where the time goes

In the businesses we have looked at, the same three drains come up.

None of those three is a strategy problem. They are all the same problem: information exists, but not in a form anything can use.

1. Supplier documents

A supplier invoice arrives as a PDF, a phone photograph, a scan with the page at an angle, or an Excel file with the columns in a different order every month. Someone opens each one and types the lines into your stock or accounting system.

This is the single best first AI project in a distribution business, for three reasons: the volume is high, the work is identical every time, and there is a correct answer you can check against. That last one matters more than it sounds. It means the system can be measured rather than believed.

What a working setup does:

• Reads the document whatever shape it arrives in, including a photograph taken on a phone.
• Pulls out supplier, date, reference, line items, quantities, unit prices, taxes and total.
• Handles French and Arabic in the same document, because that is what arrives.
• Checks the extracted total against the sum of the lines, and flags the document when they disagree rather than writing a wrong number into your system.
• Puts anything it is not confident about in front of a person, with the original alongside.

That last point is the difference between a system people keep using and one they quietly abandon. A tool that is right most of the time and silent about the rest creates more work than it saves, because now everything has to be checked. A tool that is right most of the time and tells you which ones it is unsure about means you only check those.

Extraction accuracy is not the hard part. Knowing which extractions to distrust is the hard part.

2. Stock and price questions

The second drain is a person acting as a search engine over information that already exists. A customer asks whether an item is in stock, what the price is at a given quantity, or when the next container lands. The answer exists in a system, a spreadsheet, or a colleague's memory, and it is faster to ask the person than to look.

An assistant built over your own data answers those questions directly, in the language the question was asked in. Not from general knowledge, which would be useless and dangerous here, but by looking up your actual records and answering only from what it found.

Two constraints make or break this one:

• It has to say when it does not know. A confident wrong price quoted to a customer costs you the margin or the customer. We build a refusal path before we build anything else, which is the same discipline described in stopping a model from inventing a citation.
• It is only as fresh as its connection to your data. If stock levels update nightly, the assistant is a day behind, and it should say so rather than pretending otherwise.

3. Orders that arrive on WhatsApp

A large share of Algerian commerce runs through WhatsApp, and orders come in as ordinary sentences: a mix of French, Arabic and darija, product names shortened to whatever the customer calls them, quantities sometimes written in words.

Turning that into a structured order is a genuinely good use of AI, and it is also where we see the most over-ambition. The useful version does not place the order. It reads the message, proposes a structured order with the matched products and quantities, and puts it in front of a person to confirm with one tap. The confirmation step is what keeps a mis-parsed message from becoming a delivery of the wrong thing to the wrong place.

This is covered in more depth in when your business runs on WhatsApp.

What we would not build for you

• Demand forecasting, unless you have several clean years of sales history. Without it, a forecast is a guess with a chart attached.
• Automatic reordering. The volume of purchase decisions is low and the cost of a wrong one is high, which is the exact opposite of a good automation candidate.
• Dynamic pricing. In a market where relationships and negotiation set price, an algorithm setting it will lose you customers faster than it earns you margin.
• A dashboard nobody asked for. If the question is "how much did we sell last month", that is a report, not an AI project.

How a first project runs

1. We look at real documents. Twenty or thirty of your actual invoices, not samples chosen to look good. Data condition is the largest cost driver on this kind of work, and a quote given without seeing it is a guess.
2. We agree what "working" means. A number you would accept, measured on documents the system has not seen. Written down before the build, not after.
3. We build the narrow version. One document type, one supplier group, one output destination. Narrow and reliable beats broad and unpredictable.
4. You run it on real work for two weeks. Alongside the existing process, not instead of it, so nothing is at risk while you find out whether it holds.
5. We widen it or we stop. If it did not hold, widening it will not help.

Cost drivers are the same as on any custom build: the condition of your data, how many systems have to be touched, how low the error rate has to be, and who maintains it afterwards. Those are set out in what an AI project actually costs in Algeria.

The honest summary

If your business moves goods and paper in roughly equal quantities, there is real time to recover, and it is recoverable in weeks rather than quarters. If someone offers you a platform that handles your entire operation, ask them which single document type it reads today and what its measured accuracy is. The answer tells you everything.