AI Copilot: query and enrich your catalogue in natural language

What if your teams could query their product catalogue just as easily as sending a message? That’s exactly what Pixee PIM’s AI Copilot does — no SQL queries, no technical training, no middlemen.

The problem: the data is there, but nobody can access it

In most companies, product data is stored in the PIM. But to extract anything useful from it — how many products lack a description, which items have a completeness score below 60%, which suppliers have delivered errors this month — you either need to go through a developer, be proficient in SQL, or wait for a weekly report.

The result: business teams (buyers, catalogue managers, product managers) rarely consult the data, and make decisions without having all the facts.

How the AI Copilot works

The AI Copilot translates your questions in natural language into SQL queries, runs them on your database, and responds in English with the exact figures. The entire process takes just a few seconds:

1. You ask your questionin French, just as you would ask a colleague: “How many products in the Electronics category don’t have a main image?”
2. The LLM analyses your questionand understands the context (your catalogue, your data tables, your custom fields)
3. An SQL query is generatedand executed on your database — you can view it if you wish
4. The answer is formulatedin natural language with the data: "247 products in the Electronics category do not have a main image, including 89 with stock levels above zero."

Practical examples of use

For the catalogue manager

• "What is my progress in terms of completeness by category this week?"
• "Which products have been modified by more than one supplier this month?"
• "Show me the 30 SKUs with the most missing fields."

For the purchasing manager

• "Which suppliers have an import error rate of over 5%?"
• "How many SKUs have had their prices changed by supplier X this quarter?"
• "Which products are sourced from a single supplier (risk of stock-out)?"

For the compliance manager

• "List non-compliant REACH products with stock levels exceeding 10 units."
• "Which items in the Electronics category do not have a RoHS declaration?"
• "How many products fall within the DPP scope according to their categories?"

For the sales director

• "How many products are listed on Shopify but not on Amazon?"
• "Which items are priced below the configured MAP?"
• "What is the breakdown of my catalogue by level of completeness?"

Asking a good question: three reflexes

Natural language removes the technical barrier, not the need for precision. A vague question produces a precise answer… to a different question. Three reflexes avoid most misunderstandings.

Name the scope. "How many products have no description?" does not say whether archived products count, nor whether an empty description and a missing description are the same thing. "How many active products in the Electronics category have no description in French?" leaves no room for doubt.

Name the period. This month, since the last import, for the current quarter: without a time boundary, the Copilot picks one, and it is not necessarily yours.

Name the counting unit. Are you counting SKUs, variants, or stock lines? On a catalogue with variants, the gap between those three answers is easily an order of magnitude — and it is the most common misunderstanding between two departments looking at the same dashboard. The subject is covered in detail in our article on managing product variants.

Why it’s different from a simple chatbot

AI Copilot isn’t a generic chatbot that makes up answers. Itactuallyqueries your database and gives you exact figures. If the answer is “247 products”, it really is 247 products — you can check it.

Two key differences from a traditional chatbot:

• Real-time data— the Copilot queries your catalogue as it stands right now, not a static snapshot
• Traceability— the generated SQL query is visible. You can check what was requested from the database

What the Copilot does not do

A tool is judged as much on its limits as on its promises. Four things not to expect from the AI Copilot:

• It only sees what is in the PIM. Your sales, your real margin and your customer returns live in the ERP or the BI tool. The Copilot answers on the product repository, not on revenue.
• It forecasts nothing. It counts, filters, aggregates and compares. "How many incomplete records?" is a good question; "which records will sell?" is not.
• It does not fix the data for you. It tells you 247 images are missing; the 247 images still have to be produced.
• It does not guess your in-house vocabulary. If "pending" means a precise status in your company, say so explicitly rather than relying on interpretation.

There is also a point of method. A generated query can be syntactically correct and semantically wrong: the right field is not always the one you assume, especially on a data model carrying custom attributes. That is exactly why the query stays visible. On a question that will inform a decision, read it once; on a daily check, repetition is enough to make an anomaly stand out.

Copilot or semantic search: two different tools

The two are often confused, because both understand text. They do not, however, answer the same questions.

The Copilot translates a question into a query over structured data: counts, filters, aggregates. It shines when the answer is a number, or a list defined by explicit criteria.

Semantic search relies on vector embeddings stored in Qdrant and searches by meaning: "waterproof sports shoes" also returns "running trail waterproof". It is useful when you are looking for similar products without being able to describe them through criteria — reconciling two supplier catalogues, spotting duplicates written differently, suggesting related products.

In Pixee PIM, Qdrant semantic search belongs to the Scale plan, whereas the Copilot is available on every paid plan. In practice you use both in the same workflow: semantic search to find, the Copilot to count.

BYOK: which AI model powers the Copilot?

Copilot runs on BYOK — your own API keys, like every AI task in Pixee PIM. It uses the AI provider you have configured for the "copilot" context in AI Services. By default, OpenAI GPT-4o — but you can use Anthropic Claude, Google Gemini, Mistral, or even Ollama for a 100% on-premises deployment if your privacy requirements demand it.

Your product data never leaves your database — only the query structure is sent to the LLM, not the data itself.

Access rights: the question worth asking

A tool that queries the database in natural language raises a simple governance question: can a user obtain, by asking, a piece of data the interface denies them? That is the point to check with any vendor before rolling this kind of feature out to a whole team.

Pixee PIM spreads rights across 62 permissions and 4 roles defined at repository level. That grid must remain the reference: a copilot is one more access route, not a back door. In practice, frame its rollout the way you would frame direct access to the data — starting with the roles that already have the right to read what will be queried, then widening.

On which plans is AI Copilot available?

AI Copilot is available on all paid plans (Starter, Growth, Scale, Enterprise) in BYOK mode — you provide your own OpenAI API key or similar, and pay AI costs directly to the provider, with no Pixee PIM markup.

Frequently asked questions

Can the Copilot modify my data?

The use described here is a read-only one: the Copilot counts, filters, aggregates and compares. That is also the right starting frame. Any write action triggered from a sentence should be treated like an import: a preview of what will change, explicit user rights, and a log allowing you to roll back. A conversation does not exempt you from the safeguards of a mass update.

How much does one question cost?

The cost of a question is the cost of one call to the model assigned to the copilot context, on a short context: the structure of your data model and the question, not the catalogue. Since everything runs in BYOK, the amount appears directly in your provider’s dashboard — the only honest way to know what a team’s daily usage actually costs.

Do you need to read SQL to check an answer?

No. Two checks are enough without reading a line of the query: cross-check the figure with the same filter applied in the catalogue interface, and rephrase the question to see whether the answer holds. If both converge, the doubt is settled. If they diverge, it is usually the question that was ambiguous, not the tool that was wrong.

Is my product data sent to the model?

No: the query structure goes to the LLM, not the contents of your database. If your privacy constraints require going further, the copilot context can be assigned to a model running locally through Ollama — processing then stays entirely on your infrastructure.