Finance · Inventory · Market intelligence · AI agents
An AI analyst for a multi-branch FMCG retailer
Weekly reports arrived late, incomplete and inconsistent. We built an AI analyst that checks the data first, verifies every figure, and turns it into decisions management can approve the same day.
Shopify Partner · Beirut and Paris
The client
A wholesaler and retailer, mid-way through an ERP rollout
A food and household goods group in Central Africa. It runs a wholesale warehouse, several supermarkets, a hardware store, a bakery and a cold store, plus a head office.
Management relied on weekly reporting from a small accounting team, while the group moved to a new ERP. We cannot name the client, so this page describes the work without it.
The problem
Reports that could not be trusted
- • Up to 22 reports a week were expected. Most weeks, fewer than half arrived.
- • The same report gave different answers depending on its settings: date window, unit, pack size.
- • Transfers between the group's own branches were sitting in customer receivables. At one point, 45 percent of the receivables book was the company billing itself.
- • Old data was still being loaded into the new ERP, so year-to-date profit moved by large amounts from one upload to the next.
- • Management had no single view of cash, credit, suppliers and stock.
What we built
An analyst that works alongside the team, every week
An AI analyst, built on Claude, a model from Anthropic, and run by Byblos Horizon. Seven layers, each with one job.
- Data intake. Reads ERP exports in Excel and PDF, lists what is missing, and checks each report's settings before analysing it.
- Verification. Re-checks every headline figure against the source files with automated tests. Nothing is shared until they pass.
- Analysis. Liquidity, payables and supplier advances, receivables by entity type, margins, landing costs, stock cover and slow stock, loan schedules.
- Action plans. A weekly buy list, stop list, slow-stock actions and commodity plans, in plain language.
- Market intelligence. Satellite-informed site selection for a new store, market share estimated from port discharge lists, and supplier research.
- Team support. Step-by-step ERP guidance for the accounting team, by email, with a person approving every message.
- Independent review. A second AI model audits the work each week.
How it stays reliable
Five rules, each learned from a real mistake
- Establish what a report covers before testing what it says.
- Never guess how two companies are related from their names. Ask.
- Keep measured figures and estimates separate, and label both.
- Re-verify every number before it is published.
- When a finding is wrong, correct it at the top, not in a footnote.
A person approves every email, report and recommendation before it leaves.
What it found
Problems nobody had put a number on
- • Self-billing between the group's own branches, inflating receivables by almost half.
- • A month of payables overstated many times over by a currency posting error.
- • Supplier advances that exactly mirrored old unpaid invoices at seven suppliers.
- • A stock value distorted by a report date setting. Our own earlier finding on this was withdrawn and corrected openly.
- • Slow stock worth around half the inventory value, turned into a stop list and a sell-down plan.
- • A bank loan reconciled to the books exactly.
The result
One verified view, and a plan to act on
- • One verified weekly view of cash, credit, suppliers, stock and sales.
- • Weekly action plans management can approve the same day.
- • A standard data request that tells accounting exactly what to export and how.
Next step: moving the weekly data into a database, with checks built into every load.
What we learned
The analysis is the easy part
In this kind of business, most of the work is establishing what the data really says. An AI analyst is only as good as the checks around it, so we built the checks first.
The figures on this page are our own, taken from the reports of this engagement. They are not published statistics. The client is not named, and staff and amounts are left out on purpose.
The same method sits behind our AI systems and automation work. More engagements are on the case studies page.
Let's Talk
Have messy ERP data and a management team that needs answers?
Book a call. We will tell you what an AI analyst could check first.