Case study · PALAS
PALAS: making sales agents’ work easier
A trade order often starts with just a few sentences. Someone then has to find the shop, identify the pieces of jewellery, choose the variants and note the quantities. For PALAS, we developed an AI assistant that prepares this work from a message written or dictated on WhatsApp. The request becomes a draft in the application, which the team can pick up and check before committing to the order.
PALAS is a jewellery brand that sells to the public through its Shopify store and to professional retailers. These trade sales used to be handled with Excel files, emails and conversations at trade shows. We built an application for the brand with Medusa, connected to Shopify. It serves both as a management tool for the team and as a portal for the shops. One step remained laborious: turning a freely worded request into an order that could actually be processed. Before it is recorded, a request must be linked to the right shop. Each piece of jewellery mentioned must match a product and a variant in the catalogue, with the right quantity. This re-keying work is what we set out to lighten for the people who collect these requests, including sales agents. The flow presented here covers one specific step: turning a message into a draft order that the team can check.
Write or dictate the request in WhatsApp
We placed the entry point in WhatsApp, a familiar messaging app on the phone. The person preparing the request does not have to fill in a form or navigate the management application. They word the order as they would explain it to a colleague.
The message can be written or spoken. A voice message is first transcribed, then interpreted like a written message. If transcription fails, the same request can be sent in writing.
The conversation is kept together with the state of the request in progress. The order can therefore be built over several messages: first naming the shop, then adding items, then coming back to a quantity, without repeating everything each time.
Finding the shop, even when its name is approximate
The first piece of information to pin down is the customer. The AI assistant looks for the shop among the business customers already registered in the application. This search tolerates approximations: a name that is abbreviated, misspelt or transcribed from its pronunciation can still be matched to the right record. This tolerance is particularly useful for voice messages.
When several shops could match, the flow suggests several of them instead of choosing one blindly. The person points out the right one in the conversation. The order is thus linked to a business that is already known, with its history and details, rather than to a name standing alone in a message.
Matching the jewellery mentioned to the catalogue
In a dictated request, model names may be pronounced approximately or distorted by the transcription. The flow matches the mentions of jewellery to the products and variants of the catalogue kept in the application, and assigns a quantity to each line.
This matching produces a proposal, not a certainty. A vague description may correspond to several pieces of jewellery or several variants, and the interpretation can be wrong. That is why the conversation stays open: the person can specify a model, change a variant or correct a quantity before going any further.
A summary to review before the draft is created
Once the shop and the items have been identified, the flow is designed to present a summary for confirmation: the selected shop, then the products, variants and quantities as understood. This is the moment to spot any gap between what was said and what was understood, and to correct it in the conversation.
This summary serves as a first review, while the request is still fresh in the mind of the person making it. It does not replace checking the draft in the application, described below.
A draft order prepared with data already on record
In the intended flow, confirming the summary in the conversation leads to the creation of a draft order in the application. The draft reuses the information already recorded for the shop: the business customer, the contact email, the known addresses and the currency. It also brings together the products, variants and quantities selected in the conversation. The amounts are based on the prices available in the application. They still need to be checked, especially when a shop benefits from separately agreed terms.
The draft is not a confirmed order: it still has to be checked and finalised in the application. Creating it constitutes neither a payment nor a shipment. The team picks it up in the management area to continue the preparation.
The team checks, completes and finalises in the application
In the application, the team finds a request that is already structured, in the same place as the other orders. They check the company, the contact, the addresses, the products, the quantities and the amounts. They complete or correct whatever needs it, then finalise the order through the usual process. The assistant prepares the data entry, but the relationship with the shop and the decision to commit to the order remain the team’s responsibility.
The value of this flow is to reduce copying work: finding shops and jewellery in the existing data, then gathering the request in a draft rather than keeping it in a separate note. The team can then focus on checking the prepared selection, instead of re-keying everything from the original message.
The flow focuses on preparing orders. It does not cover route planning, scheduling, customer portfolios or commissions.
For PALAS, we developed an assistant that connects order taking on WhatsApp to the management application. Its role is specific: preparing a draft from a freely worded request, using the customers and products already on record. The team retains the checking and finalisation of the order; the AI helps them prepare the data entry.