Why AI agents fail to understand your customers in one or two exchanges (and how to fix it)

In retail, every second counts. A customer hesitating in front of a product online, comparing prices, or voicing frustration in a chat often has only a few exchanges in which to be convinced. Yet most AI agents deployed today have to guess the intent and the context from one or two messages — a near-impossible mission with their current architectures. Their dataset is too narrow to catch weak signals (a doubtful tone, a browsing history, an implicit preference), and their specialization blinds them to anything outside their pre-trained script.

The result: 73% of online shoppers expect contextual personalization, but only 14% of customer problems are resolved in self-service without human intervention, according to Gartner (links to the sources are at the end of the article).

The answer is not to replace these agents, but to augment them so they map intents and personas upstream — and so turn a basic conversation into an interaction that is genuinely useful.

That is what this article is about.

The challenge: understanding intent in one or two messages AI agents (retail ones in particular) are built today for precise tasks: answering an FAQ, tracking a parcel, or recommending a product from keywords. Their main challenge is to understand an intent and a context that are often implicit.

For example, a customer writes: "This sweater is nice, but I'm not sure…" → A classic agent will reply with size availability, without detecting what may be a purchase hesitation that calls for a styling video or a similar customer review.

ChatGPT, or Mistral's Le Chat, draw on a broader dataset of the user's history, including other conversations on other subjects — foundational ones when it comes to refining the prompter's profile, who might be a risk-taker or focused on safety, a fashionista or a specialist in the ecological transition.

Why is this so hard for AI agents? A limited dataset: agents rely on predefined scenarios (for example, "product return"), whereas large models analyze billions of interactions to extract behavioral patterns. A lack of contextual memory: they do not remember the customer's past preferences, unlike platforms such as Zalando or Nike, which link the agent to the user account. Time pressure: in retail, an answer has to be fast and relevant. And 61% of customers prefer an instant AI answer to waiting for a human — provided it fits.

Consequence: according to Gartner, only 25% of customers use a chatbot again after a first interaction, because they sense a lack of understanding.

What are the keys to getting past keywords?

For an agent to become useful, it has to take in two dimensions that are often ignored:

Hidden intents. Behind a simple question lie complex needs. For example: The context of personas. A customer is not an anonymous user but a persona with a history, preferences, and expectations.

Three steps toward "context/intention-ready" agents: how to move from theory to practice

Step 1: Map intents with GPT (or an LLM) The aim is to identify the hidden intents behind customer questions, beyond the obvious keywords.

Analyze conversation logs: use GPT to sift through chat histories, emails, and customer reviews. The aim is to extract recurring intent patterns. For example: Build an intent taxonomy: sort intents into actionable categories. Sephora, for instance, distinguishes: Validate with user tests: submit the identified intents to a customer panel to check that they hold. Sephora cut its false positives by 30% by crossing GPT analysis with human feedback.

And of course (see our other publications on this subject), do not confuse intent with a keyword. For example, "delivery" can hide an urgency ("I need it by tomorrow") or simple curiosity ("What are the lead times?"). GPT can make that distinction by analyzing the context of the sentence.

Step 2: Enrich personas with contextual data The aim is to turn an anonymous customer into a dynamic persona, with preferences, a history, and a purchase context.

Cross CRM and behavioral data: feed the AI agent information such as: Create "micro-personas": rather than broad segments ("women aged 25–34"), refine personas with dynamic attributes: Use real-time context: enrich every interaction with live data:

Step 3: Inject context into the conversation (without weighing the experience down) The aim is to ask 1 or 2 questions at most in order to sharpen understanding, then deliver an ultra-personalized answer.

Ask "intelligent" questions: the questions have to be: Adapt the tone and the format of the answer: Escalate to a human when necessary. According to Gartner, 64% of customers would rather avoid AI for customer service, but only if it does not solve their problem quickly.

A synthesis of good practice Step 1: Use GPT to analyze the logs and build an intent taxonomy. Avoid confusing intent with a keyword. Step 2: Cross CRM and behavioral data to create dynamic micro-personas. Audit regularly to avoid bias. Step 3: Limit questions to 1 or 2, adapt the tone, and escalate to a human when necessary. Do not weigh the experience down.

In retail, as in a large share of customer relationships, the difference between a lost sale and a loyal customer often comes down to a weak signal that was misread. AI agents have a key role to play — provided they stop limiting themselves to keywords and instead take in intent and context.

The question is no longer whether your agents should understand your customers, but how to help them do so without weighing the experience down. And you: which weak signals are your current tools letting slip past?

Sources:

• McKinsey, "LLM to ROI: How to scale gen AI in retail", 2024: read the study: - Gartner, "Only 8% of customers used a chatbot during their last service interaction", 2023: read the study - Gartner, "Only 14% of customer service issues are fully resolved in self-service", 2024: read the release - Gartner, "64% of customers would prefer that companies didn't use AI for customer service", 2024: read the study - DigitalDefynd, "Sephora AI chatbot case study", 2025: read the analysis - Yellow.ai, "Sephora chatbot case study": read the case - DigitalDefynd, "Nike AI use cases", 2025: read the article: https://digitaldefynd.com/IQ/ways-nike-use-ai/ - AI Expert Network, "Zalando AI case study", 2023: read the case - Zalando Corporate, "Zalando Assistant update", 2025: read the release : - VUX World, "Decathlon AI chatbot interview", 2022: read the interview : - Cleverence, "Nordstrom AI customer service", 2025: read the analysis :