Over 2 million customer conversations and few insights

Maria Almenara started as a Peruvian schoolgirl selling slices of cake at break time. 

Today, it’s an artisanal pastry company running 24 shops in Lima (called “casas”) and employing more than 800 people.

It is the first Peruvian pastry chain run on data and artificial intelligence, using demand forecasting to decide how much to bake and when to bake it.

Over six years, Maria Almenara has collected more than two million customer interactions across seven communication channels.

However, these interactions sit unstructured. 

While customer interactions were already categorised at a basic level, the available structure was not detailed enough to unlock their full value for customer experience and drive product innovation.

Here are some examples of what Maria Almenara could better understand:

1. What customers want and when. Which flavours, sizes, and formats are requested, in which district, and at which time of year?
   
2. What customers ask for that does not exist. A request can be repeated a thousand times for a product that is not on the menu.
3. Where operations can fail. If complaints were clustered by shop, delivery window, and day of the week, Maria Almenara could get a specific list of what to fix first.
4. What good service looks like. Answers that resolve a case on the first attempt become the standard answers and the training material for the next similar case.

Maria Almenara has already looked at the standard fixes: traditional knowledge management tools and rule-based chatbots aren’t the best solutions.

As a result:

• Service costs climb as the company opens more shops, 
• Consistency remains hard to maintain,
• Commercial insights buried in six years of customer messages remain buried.

Both require constant manual upkeep and struggle to handle large volumes of unstructured, multilingual, multichannel conversations.

AI agents can understand customer intent, maintain context across interactions, make decisions, and execute actions across multiple systems autonomously. 

This allows Maria Almenara to provide 24/7 customer support, significantly reduce wait times, and deliver faster, more personalised, and scalable experiences, while improving operational efficiency and enabling teams to focus on higher-value activities. 

“

We transformed millions of customer interactions into actionable knowledge through AI, creating a powerful foundation for more accurate and consistent customer experiences. By launching JULIA, our AI-powered assistant, we automate routine inquiries, allowing our teams to focus on the moments that matter most: delivering empathy, human connection, and exceptional service. The result is a more efficient, scalable, and people-centred operation.

One source of truth for all channels 

Linkhub.ai is a Spanish enterprise agentic AI company operating across Spain, Peru and Mexico.

These AI systems can decide what to do next and carry it out (rather than simply answering one question at a time and being blocked after each question).

Its platform does four things for Maria Almenara:

1. It extracts and processes text and audio interactions from every channel, so a recorded phone call and a WhatsApp thread end up in the same searchable pool.
2. The platform builds semantic taxonomies, meaning it groups conversations by what they actually mean rather than by the words they happen to contain. Consequently,  “my cake arrived late”, “the delivery never showed up”, and “pedido tarde” all fall under the same theme.
3. It builds the knowledge base. That structured base is called “La Biblia” (i.e. “the Bible”), and it is what the AI agent “JULIA” reads from when a customer writes in. The same question gets the same answer whether it arrives on WhatsApp at 3:00 am or by phone on a Saturday afternoon.
4. The platform grades itself. Linkhub.ai runs evaluator agents, a second layer of AI whose only job is to score the answers produced by the first layer, measuring relevance and hallucination rates (a hallucination is an answer that sounds confident and is simply wrong, such as inventing a delivery slot that does not exist).

Human supervision sits atop all of it, guiding what the agents are allowed to do on their own.

“

EU-LAC helped us deepen our presence in LATAM with a flagship food-service case, worth more than any amount of cold outreach. It gave us the opportunity to test our AI agents in a real food-service environment, with real customers and real-world complexity. Seeing our technology hold up in those conditions is exactly the outcome we set out to validate.

Mario ARMAS RAMIREZ, CEO & Co-founder, LINKHUB.AI

From a well-defined pilot to rollout

• Service n°1: PoC design & Prototyping

This service gave the partnership a structured framework to define, run, monitor and validate the pilot. 

Before a single customer conversation was automated, both sides agreed on what success would look like: 

• a knowledge base covering at least 80% of recurring queries, 
• an answer relevance score above 80%, 
• a hallucination rate held below 3%. 

It also settled a few unglamorous points that quietly sink pilots:

• who is responsible for what, 
• which of the seven channels the pilot runs on, 
• at which milestone each result gets measured. 

The effect is to de-risk the collaboration, so that neither partner discovers a mismatch in expectations halfway through.

• Service n°2: Access to funding

A validated pilot is not the same thing as a funded rollout. 

This service provides advisory services tailored to corporate-startup partnerships on European, Latin American, and Caribbean grants, loans, and equity funds.

It checks which programmes the partnership is eligible for, advises on how to structure financing, and sets a funding strategy to scale the pilot’s results. 

For Linkhub.ai, it means identifying relevant funding opportunities that could support its future growth.

A knowledge base that can cover 80% of customer queries