AI has moved rapidly from experimentation into the commerce roadmap.
Retailers and B2B organisations are exploring conversational discovery, automated content, recommendation support, service assistants, intelligent operations and increasingly agentic workflows that can do more than simply answer a question.
The opportunity is real. So is the risk of starting with AI rather than the commerce problem.
The most useful question is not, “Where can we add AI?” It is, “Which decision, interaction or workflow would materially improve if the system could understand more context, generate useful output or take a controlled action?”
That changes the conversation from novelty to value.
Discovery can become more conversational
Traditional commerce discovery depends heavily on navigation, filters and keyword search. Those mechanisms remain useful, but they do not always match the way a customer expresses intent.
AI can help interpret natural-language needs, translate them into product attributes and constraints, and support a more conversational path through a catalogue.
The difficult part is not producing a fluent answer. It is grounding that answer in the correct assortment, availability, pricing, customer eligibility and product information.
In B2B commerce this becomes even more important. The correct product or price may depend on account, market, contract or business rules. A persuasive response that ignores those constraints is worse than a conventional search result.
Content is a strong use case—when the source is trustworthy
Commerce teams manage large volumes of product descriptions, attributes, translations, campaign copy and channel-specific content.
AI can accelerate drafting, transformation and enrichment, particularly when it works from governed product information and clear brand or regulatory rules.
But generation should not become a substitute for source-data quality. If product attributes are incomplete or inconsistent, AI may simply make the inconsistency sound more convincing.
A stronger pattern is to use AI around authoritative product data: identify gaps, suggest enrichment, adapt approved information for different channels and keep human review where commercial or compliance risk requires it.
Operations may create less visible but significant value
Some of the strongest AI opportunities are behind the storefront.
Commerce operations generate repetitive decisions: categorising issues, summarising incidents, identifying anomalies, preparing release information, supporting customer-service agents, analysing search behaviour or helping teams navigate operational knowledge.
These use cases may not produce a dramatic customer-facing demo, but they can reduce friction for the teams operating the platform.
The same principle applies to software engineering and quality. AI can assist developers, accelerate repetitive implementation work, support test creation and help teams investigate problems. The benefit depends on engineering controls, review and the quality of the surrounding delivery process.
Agentic commerce changes the authority question
As AI systems move from recommending to acting, architecture becomes more important rather than less.
An assistant that suggests a product is different from an agent that changes a cart, applies an offer, initiates a return or triggers an order-related workflow.
Every action raises questions about authority and accountability.
What is the agent allowed to do? Which system validates the action? What customer or business context is required? Which decisions need confirmation? What is logged? How can an action be reversed? What happens when an upstream service is unavailable?
These are not model-selection questions. They are commerce architecture and operating-model questions.
AI needs the enterprise around it
A production AI capability often depends on systems that already exist: commerce platforms, PIM, ERP, CRM, search, order management, identity, APIs, data products and operational monitoring.
The model can interpret or generate, but the enterprise still provides truth and execution.
This is why disconnected pilots frequently struggle to become production capabilities. A prototype can work with a small curated dataset and manual supervision. Production needs permissions, integrations, observability, security, error handling, cost controls and clear ownership.
The gap between those two states is where architecture and engineering matter.
Where AI may not be the answer
Not every commerce problem needs intelligence.
If a workflow is deterministic, stable and easily expressed as a rule, conventional automation may be cheaper and more predictable. If source data is unreliable, fixing the data may create more value than adding a model. If an integration is missing, AI does not remove the need to connect the systems. If the organisation cannot define who owns a decision, autonomous execution may increase risk rather than reduce effort.
AI should therefore compete with simpler solutions, not automatically replace them.
Start with readiness and measurable value
A practical commerce AI initiative can begin with a narrow use case and a few disciplined questions:
- What business or customer outcome are we trying to improve?
- What context and data does the capability need?
- Which systems provide authoritative information?
- Is the AI recommending, generating or acting?
- What authority can safely be delegated?
- Where is human review required?
- How will quality, cost and operational behaviour be measured?
- What happens when the AI is uncertain or wrong?
These questions make it easier to distinguish a useful production capability from an attractive demonstration.
AI is likely to become a normal part of the commerce stack. The organisations that benefit most will not necessarily be those that add it everywhere first. They will be the ones that connect it deliberately to real customer needs, governed enterprise context and workflows where intelligence genuinely improves the outcome.