AI Automation

Orders and documents that AI takes into the ERP without retyping

Many customer orders and almost all supplier delivery notes and invoices arrive as PDFs or as text in the body of an email, each in its own format. An AI automation can read those documents, extract the lines, SKUs, quantities and dates, and translate the customer's or supplier's references into your ERP codes.

Before saving anything, the system checks what it extracted against your master data: that the SKU exists, that the price matches the price list, that the minimum order is met. What fits goes to a validation queue; anything uncertain or inconsistent is flagged for a person to review alongside the original document.

The most laborious part of the project is usually the mapping table between the references customers and suppliers use and your own. Once built, every correction the team makes extends it, and fewer and fewer references go unrecognised.

AI-assisted replies in customer service

Enquiries arriving by email or through the store's contact form are classified by type (order status, return, delivery problem, invoice) and routed to the right person. For the most frequent ones, the automation checks the actual status in the store, the ERP or with the carrier and prepares a draft reply with that data.

The agent reviews the draft, corrects it if needed and sends it. The automation does not promise refunds, compensation or deadlines on its own: those decisions remain with a person, and sensitive or unusual enquiries go straight to the team without a draft.

Drafts follow the tone and templates your team already uses, and each reply sent is linked to the order or customer in your service desk tool or CRM, just as if it had been written by hand.

AI product record enrichment with approval

When a supplier sends their catalogue, the data rarely arrives the way your store needs it: attributes with different names, mixed units, terse descriptions or descriptions in another language. The automation normalises attributes to your structure, suggests descriptions based on the technical data and prepares the translations.

Nothing is published without review. The suggestions arrive in the PIM or the eCommerce platform as drafts, and the product manager approves, edits or rejects them. That way the team spends its time reviewing and deciding, not copying and pasting.

The rules for each product family, such as which attributes are mandatory, which units are used or which expressions must not appear in a product record, are defined with the product team, and the automation applies them to every suggestion.

Translations are reviewed with the same criteria, paying particular attention to technical terminology and to the units used in each market.

Human oversight and control of AI automations

Every automation we design has a confidence threshold below which the case goes to a person, a review queue visible to the team, a log of what it did with each document or enquiry, and a metric for how many results have had to be corrected.

We always start in suggestion mode: the automation proposes and a person decides. Only when the correction data show that a specific type of case is consistently handled well do we consider letting it go straight through, and there is always someone on your team who can stop it or switch it back to suggestion mode.

These controls are not an optional extra: they are part of the design from day one and are reviewed regularly with your team, along with the usage cost of each automation. If what it takes to maintain or correct it outweighs the time it saves, it is redesigned or withdrawn.