In e-commerce, the reliability of AI agents depends above all on the quality of the data provided to them. Written in collaboration with Frédéric Godefroy, co-founder of Djiin by Sereneo.
AI agents are progressing quickly in e-commerce. Product research, offer comparison, customer relations, order tracking: conversational interfaces are gradually taking an important place in the purchasing experience. But on post-purchase, a limit quickly appears. The answers generated by AI still often remain imprecise or inconsistent. Not because the models are bad. The problem mainly comes from the data to which they have access.
In many companies, information useful for post-purchase already exists: carrier statuses, logistical incidents, delays, delivery events, marketplace data, etc. But this data generally remains dispersed between several tools and several players. And this is often where AI agents quickly reach their limits.
According to the Metapack Ecommerce Delivery Benchmark Report 2026 study: 28% of consumers already use AI for shopping-related tasks, this figure rises to 40% among those under 45.
Post-purchase becomes a strategic data source for AI agents.
Post-purchase concentrates a significant amount of operational data: carrier events, delivery statuses, delays, returns, after-sales service interactions and even marketplace information.
With the emergence of conversational AI agents, a new need appears: making this data directly understandable and usable by AI. And this is precisely what becomes complex for e-retailers. Because even when the data exists, it often remains fragmented between several systems: carriers, OMS, WMS, CRM, marketplaces, after-sales tools, etc.
The real subject is therefore no longer just AI. The subject becomes the consolidation, structuring and reliability of post-purchase data. An AI is only as relevant as the data it is fed.
Lack of data, primary cause of error and therefore disappointment — but not without remedy.
We have seen it: consumers are getting used to AI in their personal use, which is becoming more widespread. This gives everyone a personal experience of the capabilities of AI, and its limitations. However, even informed of these limits, when the AI hallucinates, the effect remains disappointing.
For the AI agent to be accepted — or better, for it to be appreciated — it must be reliable. It can then be very useful to both the customer and the e-retailer. However, the models suffer from memory loss. All.
A recent Microsoft study confirms this: after several exchanges, models lose memory of the texts encountered at the start of the conversation. The cause? AI is certainly probabilistic, but above all it has an attention window which gives more importance to certain information than to others — especially the most recent. For the answer they formulate to be better, they need to be able to exercise their “attention” on the data that allows them to respond fairly; therefore to have this data as close as possible to their response, just before it is made.
Add to this that, by construction, the AI always responds even when it ignores the response. Faced with a question, AI agents are therefore designed to give an answer, even if it means inventing it. For the AI model, this is not invention, but only the most probable text after the request.
So, so that they do not invent, so that this probability coincides with reality, we must constrain this probability: ensure that the most probable answer is also the most accurate. It depends on the quality of the information present in the prompt before the response.
In short, to respond better, the AI must have real information, provided just before its response — and this information can either be injected by the conversation prompt, or searched by the agentic AI in reliable sources.
Why MCP is becoming a key building block for AI agents.
MCP (Model Context Protocol) allows AI agents to access business data and tools much more easily. Concretely, an MCP server acts as a standardized layer between an AI and a company’s operational systems. Instead of developing complex specific integrations for each AI assistant, MCP allows structured data to be exposed in a much simpler and actionable way. The protocol is gradually being widely adopted in the world of conversational AI agents.
For e-retailers, the benefit is concrete: low integration effort, simplified access to business data, standardization of exchanges, compatibility with new AI environments.
Why post-purchase will be a key use of AI agents.
Post-purchase has all the characteristics of an ideal use case for AI agents: high volume of requests, structured data, need for immediacy and repetitive interactions.
The subject is no longer just about having a tracking page or notifications. The subject becomes: how to make e-commerce data understandable, usable and actionable by AI agents.
