AI and customer service
A large share of support contacts are order-status, delivery, returns and product-detail questions. Those are answerable from systems you already run, which is what makes them the highest-return starting point.
The differentiator is not the model. It is that the agent can read the order, the fulfillment record, the carrier status, the returns policy and the customer's history, and can take a bounded set of actions against them. Integrated with Gorgias, Zendesk or Kustomer, with escalation rules and full action logging.
AI and product data
This is where AI has changed the economics most clearly. Attribute extraction from unstructured supplier documents, categorization against your taxonomy, tagging, description drafting, translation and duplicate detection are all now feasible at a scale that used to be a headcount decision.
The sequence matters: schema first, then extraction, then generation. Run it in the other order and you industrialize your existing data problems. See PIM and product data systems for the groundwork this depends on.
AI, search and personalization
Semantic search handles the queries keyword matching fails on — descriptive phrases, use cases, and the way people actually type. Recommendations improve when the model can read margin and stock position, not only behavioural similarity.
Both depend on product data quality. A semantic layer over an inconsistent catalog surfaces inconsistency faster; it does not resolve it.
AI in operations
Demand and inventory signals, exception detection, order routing suggestions and anomaly alerts across channels. The value here is time: catching a margin-negative pattern or a stock discrepancy in hours instead of at month end.
- Contribution margin erosion detected against a baseline rather than a target
- Inventory anomalies where system quantity and sales velocity disagree
- Fulfillment exceptions surfaced before the customer notices
- Ranked recommendations with the evidence that produced them attached
How we scope AI work
We start by asking what decision or task the AI is supposed to change, and what data it would need to see to do that. If the answer is data you do not currently have connected, the honest first project is the integration, not the model.
We will say so rather than shipping something demonstrable that quietly fails in production. That conversation is usually shorter and cheaper than the alternative.
Common questions
- What can AI actually do for an ecommerce business today?
- Four things reliably: enrich and normalize product data at catalog scale, improve search and recommendations through semantic understanding, resolve a meaningful share of support contacts that are really order-status questions, and surface operational exceptions earlier than a reporting cycle would. Everything else is still emerging.
- Why do most commerce AI projects underdeliver?
- Because the model cannot see the data that would make it useful. A support agent without live order, stock and shipping context can only paraphrase a help centre. The work that makes AI valuable is almost entirely integration work, which is why it belongs with the systems practice rather than beside it.
- Do we need clean data before starting?
- For product-facing use cases, largely yes — generation on a poorly modelled catalog produces fluent text that is confidently wrong and harder to audit than the gaps it replaced. For support and operational use cases you can start sooner, because the model is reading transactional records that are already structured.
- How do you keep AI output trustworthy?
- Ground it in retrieved records rather than model memory, constrain it to actions with defined side effects, log every action with the evidence that triggered it, and keep a human approval step wherever the action is expensive or irreversible. Confidence thresholds decide what escalates.