The use of AI must be strictly regulated

B2B commerce has additional layers that AI must take into account; otherwise, it immediately comes across as unreliable. Just think about:

• Contract pricing & customer agreements: the right product at the right price for the customer.
• Product access rights: Customer A can view Product X, but Customer B cannot.
• Availability & Delivery Terms: What is available, where, when, and under what conditions?
• Compatibility & safety: An incorrect attribute or the wrong variant can disrupt downstream processes.

‍

A common scenario is when an AI assistant recommends a “matching” part but overlooks a compatibility rule or a variant attribute. The customer places an order and later discovers that it doesn’t fit. This erodes trust in the AI, but more importantly, in the organization.

AI highlights data issues but does not automatically resolve them.

‍

Common causes of unreliable AI output in commerce

In practice, this is where things sometimes go wrong:

• Incomplete or unstructured product data. Important specifications are contained in plain text or PDFs.
• Multiple “sources of truth” (ERP/PLM/PIM/spreadsheets) containing conflicting values.
• No clear ownership. No one is responsible for end-to-end product data quality.
• Governance exists only on paper. Validation rules and workflows are in place but are not consistently applied.
• AI without context rules: the AI is allowed to provide answers outside the scope of contracts, product ranges, or availability, and that immediately feels “wrong.”
  ‍

Organizations sometimes underestimate the difference between traditional data management and AI-ready data, and without that foundation, AI initiatives are bound to fail.

‍

Examples of safeguards for customer-centric AI

If you use AI in a customer-centric way, treat it as an assistant that operates within certain guidelines, not as a standalone chatbot. Here are some practical guidelines:

• Citation per answer: indicate which product data or source it is based on.
• Work within the customer's context: contract prices, product range rights, and availability serve as strict limits.
• Human-in-the-loop for critical information: let AI make a recommendation, which a human then approves (or AI automatically escalates the matter).
• Confidence + “I don’t know”: if data is missing or too uncertain, the AI must explicitly state this and refer the user elsewhere.
• Logging & audit trail: log prompts, sources, product data versions, and output so you can analyze and improve.
• Validation rules in the data layer: ensure that critical attributes (safety/compatibility) cannot be generated “freely” without verification.
• Roles and permissions: Give internal teams more “access” than external buyers (not everything is public)

‍

‍

Where do you start?

Most improvements don't start with technology, but with insight:

• Where does your product information come from?
• Where do differences arise?
• Which data is authoritative and which is not?

‍

From there, you start where the impact is measurable and the risk remains manageable. Consider using AI for internal support—such as accelerating the creation of product content, improving search synonyms, or preparing responses with source references that an employee approves.

Choose one category or product line and make it “AI-ready” by establishing clear attribute standards, ownership, and review workflows. Making data AI-ready is an iterative process that involves metadata and governance, which you need to develop on a case-by-case basis.

By taking a step-by-step approach, you’ll build trust with stakeholders and ensure that AI doesn’t expose the problem, while keeping it under control.