The AI agent in Supplier Data Manager (SDM) uses pre-trained large language models (LLMs) to automate three key steps in your supplier data workflow: Classification, Extraction, and Mapping. Unlike trained AI models, the SDM AI agent requires no additional training or user feedback loops, enabling faster onboarding and immediate value from day one.

How the SDM AI agent works

The SDM AI agent uses pre-trained natural language understanding to interpret labels, descriptions, and other text-based sources from your input files. It operates across three workflow steps:

• Classification: Automatically assigns products to categories and families based on your predefined taxonomy.
• Extraction: Fills in missing product attributes by analyzing available text fields such as product names and descriptions.
• Mapping: Suggests field mappings between your source file columns and your internal SDM taxonomy fields.

No training datasets or ongoing model updates are required. The AI agent performs accurately out-of-the-box.

Key benefits

• Reduced time-to-value: Save up to 75% of the effort previously required for training-related tasks.
• Multilingual support: Supports classification and extraction in multiple languages, including Japanese, German, and Arabic.
• Improved accuracy: Tested across diverse datasets to ensure robust performance in real-world scenarios.

AI Classification step

The SDM AI Classification step uses a pre-trained LLM to analyze product data and assign each product to the most appropriate category or family in your taxonomy.

Automated mode vs. review mode

You can configure the Classification step to run in one of two modes:

• Automated mode: The AI classifies products automatically. Only rows the AI could not classify are surfaced for manual review. Successfully classified rows appear in the checked by AI tab (blue).
• To check mode: All product rows require manual verification, even when the AI has found a match. AI-suggested rows appear in the to check tab (red).

Preserving existing classifications

If your source file already contains valid family or category information, you can configure the Classification step to skip AI predictions for those products and use the existing data instead. This avoids unnecessary processing and preserves correct classifications already provided by your supplier.

To enable this:

1. Ensure the family and category fields are mapped in the Mapping step and contain valid codes.
2. Set the replace_existing option to false in the Classification step configuration via the SDM admin panel. See the Admin guide for details.

Classification limitations

• Maximum products per job: 30,000
• Recommended maximum source attributes for AI: 50 — using more than 50 attributes may reduce accuracy.
• For best results, test with fewer than 200 products and fewer than 20 source fields before scaling up.

AI Extraction step

The SDM AI Extraction step automatically fills in missing product attributes by analyzing text sources — such as product names, descriptions, and other fields — from your supplier's input file.

How extraction works

The AI scans the configured text sources and fills in empty attribute values, for example dimensions, materials, or technical specifications. Attributes filled with low-confidence predictions are flagged in orange for review. Images from the input file can also be displayed to help you confirm or edit attributes.

If an attribute already has a value, the Extraction step can be configured either to skip it or to re-evaluate it. When re-evaluating all attributes, the step does not notify you when its output differs from the supplier-provided values.

Extraction limitations

• Maximum products per job: 30,000
• Recommended maximum source attributes for AI: 50 — using more than 50 attributes may reduce accuracy.
• For best results, test with fewer than 200 products and fewer than 20 source fields before scaling up.

AI Mapping suggestions

The SDM Mapping step includes an AI-powered suggestion feature. When you click Generate AI mapping, the AI analyzes your source file's column names and sample values, then proposes the most likely match for each unmapped field.

• Only matches with a confidence score above 70% are surfaced.
• Fields that are already mapped are left untouched.
• AI mapping supports source files with up to 2,000 columns. Files exceeding this limit require manual mapping.

You can accept, adjust, or ignore each suggestion. See Mapping: Structure Supplier Data for full details.

Custom prompts for better accuracy

You can improve the AI Classification and Extraction steps by writing custom prompts. Custom prompts provide specific instructions that tailor the AI's behavior to your taxonomy and company context.

Adding a custom prompt to the Classification step

To add a custom prompt to the AI Classification step in Supplier Data Manager:

1. Go to Workflow > Settings > Steps.
2. Select the classification step you want to customize.
3. Enter your instructions in the text box provided.

Writing an effective prompt

• Include sufficient source information. Fields such as product names and descriptions contain the most useful context for the AI.
• Provide examples. The more examples you include, the more reliably the AI follows your instructions.
• Use labels, not codes. Codes must be explicitly introduced for the model to interpret them correctly. Prefer full category labels over internal codes.
• Provide full category paths. For example, write "Kitchen > Large appliances > Refrigerators > Wine fridge" instead of just "Wine fridge".
• Avoid contradicting the AI's general knowledge. If a product description clearly identifies a product type, prompting the AI to classify it differently will reduce accuracy rather than improve it.

We recommend taking the Akademy course on prompting for AI to learn how to create effective prompts for your context.

Changing the source attributes used by the AI

For both the Classification and Extraction steps, you can control which attributes the AI uses to generate predictions. This is configured in the Sources section, located alongside the custom prompt in the step settings.

The screenshot above shows the Sources configuration panel in an SDM Classification or Extraction step, where you select which input attributes the AI uses to generate predictions.

Common questions

For answers to frequently asked questions about AI in Supplier Data Manager — including how the AI handles your data, security, multilingual support, and model accuracy — see AI-related questions.