How does the AI in SDM work?

The AI operates in two layers:

Understanding Inputs: The first layer interprets your input and translates it into a format the system can process. This layer has been trained on a vast multilingual dataset, enabling it to understand input even in regions where SDM might not yet have active customers.

Producing Outputs: The second layer is specialized for specific tasks, converting the processed input into actionable outputs, such as categories or attribute values. This layer can be trained individually for each customer to ensure data security if a “trained model” is used. Data from one customer is never shared or accessible to another.

How is the accuracy of the AI ensured?

For trained models, to maintain accuracy, we implement a robust auditing process:

• Predictions made with high confidence (automated rows) are sampled and flagged for auditing.
• Auditors validate these samples by comparing the AI’s output with user-approved corrections, ensuring the system performs reliably.
• The sampling rate is customizable and typically ranges between 10% (for new models requiring trust-building) and 1% (for established models with proven performance).

What data is used to train the AI models?

In the case of trained models, the AI learns from user-approved actions:

• Training Data: Rows that are manually validated or corrected by users are used to improve the AI's algorithms.
• Excluded Data: Automated rows are excluded from training to prevent biases, even though auditing mitigates risks of errors in automated outputs.

In the case of LLM models, the AI is trained on huge sets of data that users can benefit from right away.

Can SDM use my existing data for training?

Yes, leveraging your existing data is an integral part of the onboarding process, in the case of trained models. During onboarding:

• We audit your data to verify its alignment with your target taxonomy, correctness, and suitability for supplier or seller requirements.
• This audit ensures that as much of your data as possible is used to enhance the AI’s performance.

Is customer data shared? Can I benefit from others’ data?

No, SDM ensures complete data isolation. Each customer has a dedicated AI model, meaning:

• Your data is never shared with other customers.
• You cannot access or leverage insights from competitors’ data.

What does “Zero-Shot” or AI Classification agent mean?

A pre-trained model, LLM model, which can provide instant results. It is used by default at the classification and extraction steps, without training period required. See related article

How does AI Classification agent understand our data?

LLMs are trained on extensive sets of text, allowing them to effectively interpret and classify product attributes from the beginning. SDM benefits from the best AI models on the market - GPT 4o mini

What if the AI makes a mistake?

LLM models are designed for high accuracy but are not infallible. Misclassifications can be corrected during the step.

Can the AI Classification agent learn from corrections?

No, ZeroShot models don’t rely on feedback loops. As part of future developments, we want to enable adjustments to prompts.