How to: deploy an AI agent in production in 4 months for lacoste

Benoît Dugelay helps large organizations design and deploy agentic platforms at scale. At converteo, he leads conversational and agentic AI projects for retail, e-commerce and services.

Key takeaways

• From deterministic chatbot to conversational agent: lacoste moved from a chatbot limited to after-sales service to an agentic platform capable of handling the entire customer journey — from FAQ to escalation — while embodying the brand’s voice.
• Four months end to end: two months of design, two months of development and testing. An industrial product, integrated with salesforce commerce cloud, service cloud and manhattan associates, live in production on lacoste.com US.
• The product builder as keystone: this hybrid profile — bridging business vision and technical execution — enabled continuous risk mitigation, fast prototyping and validation before building.
• Results from day one: consumer satisfaction above 4/5 from the first days, escalation rate on track for -30%, complex cases handled in under eight minutes on the advisor side.

For two years, organizations have been stacking AI experiments without crossing into industrialization. lacoste has just provided a concrete answer to the question: how do you move from prototype to product in production, fast and well? Here is a project delivered in four months, from vision to deployment.

The vision: a digital advisor in the image of the boutique

It all starts from a simple but demanding observation. When a customer walks into a lacoste boutique in paris or monaco, they are welcomed with a tone, a posture, a desire to introduce them to the brand’s codes — and ideally, the ability to personalize the relationship based on their profile. A streetwear teenager is not approached the same way as a customer looking for a gift for someone.

On lacoste.com, none of that existed. The previous chatbot, purely deterministic, was limited to after-sales service, unable to cover product discovery, recommendation or purchase assistance. The gap between the in-store experience and the digital experience was stark.

The vision formulated by lacoste’s teams was therefore clear: deploy a digital advisor capable of reproducing, at scale and in natural language, what the best in-store salesperson does. Personalized recommendation, catalog discovery, handling of post-purchase requests — all while respecting the brand’s codes, right down to the wording used when escalating to a human advisor.

Four months to industrialize AI: mission impossible?

The project kicked off in november, with a hard constraint: the existing chatbot on the US market had to be replaced by february-march. Two months of design, two months of development and testing. That is the timeframe in which converteo, in partnership with google, delivered a complete agentic platform to production.

What made this timeline achievable was a combination of three factors.

A solution with pre-built components. google’s gemini enterprise for CX comes with pre-configured agents and customer experience building blocks already available. This allowed the converteo team to focus on customization, business integration and tone of voice, rather than starting from scratch on infrastructure.

Native integration into the existing stack. The agent was connected to salesforce commerce cloud for the catalog and e-commerce layer, to salesforce service cloud for ticket management and escalation to advisors, and to manhattan associates for order tracking and logistics. All critical integrations delivered on time.

An integrated team, not siloed. The converteo teams worked directly on lacoste’s premises, within the brand’s digital factory. No handoffs, no cross-team tickets: a shared agile squad with a shared commitment to KPIs.

The expert agents: an orchestrated architecture

The deployed architecture is built around a central orchestrator agent that routes each request to the most appropriate expert agent based on detected intent and the criticality level of the request.

• The knowledge agent answers frequently asked questions — general terms and conditions, return policy, pre-sales questions — from an enriched and trained FAQ base. It handles a large share of incoming traffic without human intervention.
• The order tracking agent handles all delivery-related requests: parcel tracking, cancellation requests, delay management. It is directly interfaced with manhattan associates, lacoste’s OMS tool, to retrieve real-time information.
• The product quality agent manages returns related to defects or post-wash issues — a significant flow in fashion retail. Its role is to precisely qualify the issue and prepare the case so that a human advisor can handle it quickly, with all the necessary context.

It is on this last point that the notion of symmetry of attention takes on its full meaning: the agent does not only serve the customer, it also serves the advisor. When an escalation to a human is required, the advisor retrieves the entire conversation (including any photos shared by the customer) directly in their salesforce service cloud interface, with no additional data entry required.

The tone of voice: when the brand speaks through the agent

One of the most structuring workstreams of the project was work on the brand’s voice. lacoste has precise guidelines: words to avoid, a register of elegance to uphold, a welcoming posture drawn from the codes of the physical boutique.

Training the language model on these constraints (empathy, short sentences, avoiding overly verbose phrasing, rapid escalation in case of tension) required careful prompting and fine-tuning work. This work is ongoing: the agent learns, the teams adjust. It is precisely for this reason that the project did not stop at go-live.

The agent is called ask rené, a nod to rené lacoste, founder of the house. A way of embodying the relationship, even within an automated system.

The product builder: the profile that de-risks

Agentic projects do not look like classic digital projects. They carry a degree of uncertainty — possible hallucinations, variable quality of source data, behaviors that are difficult to anticipate — that makes the classic “spec → dev → test → prod” cycle ill-suited.

This is where the product builder comes in: a hybrid profile, capable of understanding business challenges, rapidly prototyping with AI tools and vibe coding, and validating feasibility before development teams even begin to build.

On this project, the product builder took ownership of tone of voice work from the design phase, producing functional prototypes submitted directly to lacoste’s business teams for validation. Only once those prototypes were validated did industrial development begin, which avoided costly back-and-forth at the end of the project.

This role becomes structuring in any agentic project: it is not about managing a backlog, but about continuously de-risking at the boundary between business and technology.

Early results and next steps

Four days after launch on the US market, early indicators are encouraging:

• Consumer satisfaction: above 4/5, exceeding the target set
• Escalation rate: on track for -30% compared to the previous chatbot
• Handling time for complex cases: under 8 minutes on the human advisor side

Deployment continues. Canada is the next step, followed by europe and latin america. Teams are currently working to adapt the agent to each market’s specificities — different carriers, local offers, languages —, work that is more about business testing and validation than pure development.

The roadmap goes beyond care. Personalized product recommendation, conversational personal shopping, multimodal interaction (voice, image) and omnichannel integration with in-store salespeople are the next identified workstreams. The question of navigation, and of the place of search in an increasingly conversational journey, is already on the table.