Installing an AI that advises on your Shopify store

There are four steps: connect your catalogue, calibrate the advice, test before your customers do, then go live. Each has a precise purpose and an order to respect — skipping calibration or testing to move faster is exactly the mistake that costs money later.

• Step 1: the app. You authorise the Ryma app from inside your store: products, stock and orders follow, live.
• Step 2: calibration. The house tone, the rules, what she never says.
• Step 3: testing. Trick questions and sensitive cases, before the first customer meets them.
• Step 4: going live. With a dashboard where the team keeps the wheel.

Along that path, a general AI of the e-commerce chatbot kind answers differently from a genuine advisor: the difference is decided in calibration and testing, not in the technical connection.

On Shopify, the connection goes through the Ryma app. You authorise it from inside your store: Shopify displays the exact list of access being requested, you approve, and it is connected. Nothing to export, no API key to copy and paste, no intervention in your theme's code.

What that authorisation changes concretely: the advisor reads your products, your prices, your stock and your orders live. When a product goes out of stock or a promotion starts, she knows, without anyone having to update a file.

Control stays with you: the access granted is visible in your Shopify admin, and you can withdraw it at any moment, in one click, without asking us.

The underlying principle doesn't change: the advisor never recommends a product outside your catalogue. That is what prevents invented promises about a product that no longer exists or is out of stock.

Calibration means fixing the house tone, the rules of advice, and above all what the AI never says. It is the step general-purpose tools neglect most, and the one that decides whether the advice sounds like your brand or like a standard robot.

In practice that means defining: the register (formal or familiar), the phrasings that are off limits, and the list of subjects she must never settle alone.

In beauty and skincare, that calibration takes a precise form: a register of guidance reviewed and approved by health professionals (a dermatologist, a paediatrician, a pharmacist). On sensitive cases — pregnancy, allergy, a complaint — the rule is fixed: she hands over to the retailer's team. No improvisation on those subjects, whatever the question.

That point echoes a market reality: the major brands (L'Oréal, La Roche-Posay, Vichy) have each developed their own AI skin-assessment tool for their site. The standard expected of beauty advice is no longer that of a simple support chatbot. To dig into that sector calibration, see the guide to the AI advisor for beauty and skincare.

Before opening to the public, the AI has to be tested on trick questions and sensitive cases, not only on easy ones. The aim: check that she holds her line — her calibration — when the conversation leaves the expected script.

That means simulating deliberately difficult scenarios: a question outside the catalogue, a complaint, a medical request disguised as a product question, a customer who insists. It is in those tests that the rules set at step 2 are checked in real conditions.

A useful order of magnitude for the effort: at Beautymarket, the number one beauty and skincare store in Morocco, the Ryma advisor was tested on 100,000 conversations before going into production. That testing phase also measured that 70% of requests could be handled end to end with no human involvement. It is not a figure to reproduce identically for every store, but an order of magnitude for the testing a broad dermo-cosmetic catalogue (L'Oréal, La Roche-Posay, Vichy, Avène, Bioderma, CeraVe in this case) can demand.

Once live, the AI stays under control through a dashboard where the team keeps the wheel at all times, with every conversation on the record. Going live is not an arrival point: it is the start of ongoing supervision.

Concretely that means being able to take over a conversation in progress, read the history, and adjust the calibration if a new situation appears (a new product, a new recurring question, a new offer). Across channels, the same memory of the catalogue is used on the site, on WhatsApp and on social: no need to recalibrate per channel.

On that ground the gap with a general-purpose support tool is stark: her work doesn't stop at the answer, it ends at the cart. To go further on what the dashboard allows once you are live, particularly on abandoned carts, see the guide to abandoned cart follow-up on WhatsApp.

The first costly mistake is plugging in a general-purpose tool and never feeding it your real catalogue. The result: generic or approximate answers that reflect neither your products nor your stock. It is one of the reasons general-purpose chatbots rank, in France, as the least-liked channel of all, according to an Ipsos BVA study of 5,000 people in 2025.

The second is letting the AI improvise on sensitive cases (pregnancy, allergy, a complaint) for want of a rule set in advance. In beauty and health that is not optional: the rule has to be to hand over to the human team, not to attempt an answer.

The third is neglecting traceability. Without a readable conversation history, it is impossible to check what was said to a customer, or to adjust the calibration afterwards. The risk is not only operational, it is regulatory too. The guide to reliability, GDPR and responsibility sets out the questions to ask before choosing a tool on that point.

The thread running through all three: 32% of customers walk away from a brand they love after a single bad experience, according to a PwC study of 15,000 consumers. Badly calibrated advice is not neutral — it costs you loyal customers.

If your store isn't on Shopify, there is no app to authorise: setup starts from an export of your catalogue, in a spreadsheet for example. The four steps stay the same: catalogue, calibration, testing, going live.

The difference is confined to step 1. Without a connected app, live stock and automated order tracking need a periodic refresh of the catalogue rather than a continuous sync. Calibration, testing and the dashboard work in exactly the same way, whatever the e-commerce platform.

For a wider view of what an AI advisor covers, whatever the platform, see the AI advisor guides hub.