Diabolocom

Proud to have contributed to the latest customer experience study by Association Française de la Relation Client - AFRC and frog. Our participation reflects our commitment to shaping the future of CX by turning innovation into clear, actionable outcomes for our customers. 🔎 Discover the key insights in the public version of the study, with the full report available to AFRC members. Thank you to all the teams involved for this meaningful collaboration. 👉 Read more: https://lnkd.in/g3rhDR7p 👉 Full report at AFRC: https://lnkd.in/eSpRjjMQ #CustomerExperience #CX #CCaaS #ThoughtLeadership #Innovation #AIinCX

If your contact center AI isn’t delivering ROI, the problem may not be the technology—but how it was built. ⚠️ Generic AI models often perform well in demos but struggle in real-world CX environments where noise, crosstalk, and industry-specific language are the norm. In this CX Today article, Théo Deschamps-Berger explains why metrics like Word Error Rate (WER) are only a starting point—and how a five-pillar framework focused on real operational performance changes the ROI equation. 👉 Discover what truly drives AI success in contact centers: https://lnkd.in/efTs9u5e #CustomerExperience #CX #AIinCX #CCaaS #ThoughtLeadership

🚀 Big news from the Diabolocom Research Lab: we're releasing BidirLM, a family of 5 frontiers open-source AI models — including an all-in-one model that understands text, voice, and images at once. Why does this matter for our clients? Behind every great customer interaction sits a layer of AI quietly doing the unglamorous work: finding the right knowledge base article, routing a request to the right team, detecting customer intent, measuring conversation quality. These tasks rely on a specific family of AI models called encoders — and while the industry has spent billions training chatbots and generative AI, encoders have been largely left behind. That's the gap our research lab set out to close. With BidirLM, we show how to turn today's most advanced AI models into best-in-class encoders — faster, more accurate, and able to work across languages, voice, and images. Concretely, this means better understanding of your customers, in more languages and channels, at a lower cost. What we're releasing to the open-source community: 🧠 BidirLM-Omni-2.5B — One compact model that handles text, voice, and images together. Outperforming other models at half the cost. 📖 4 text models (from 270M to 1.7B parameters) — Setting a new open-source standard across all practical sizes. 🌐 Truly multilingual — Built to perform well far beyond English. 🧪 The full recipe — So the entire research community can build on our work. 📚 Everything open — Datasets, model checkpoints, and experiments. 📄 Paper: https://lnkd.in/e7MYEn2c   🤗 BidirLM Collection: https://lnkd.in/eaeYYFtM 📝 Blog post: https://lnkd.in/eh3AhBiq This work was led by Nicolas BOIZARD (CentraleSupélec × Diabolocom) and Théo Deschamps-Berger (Diabolocom), with Hippolyte Gisserot-Boukhlef, Pierre Colombo, and Celine Hudelot. Huge thanks to the French supercomputers Adastra (Centre Informatique National de l'Enseignement supérieur), Jean Zay (GENCI), and ROMEO (Université de Reims Champagne-Ardenne) for making this possible. 🇫🇷 A small look at the research powering Diabolocom.

Big news! 📣 Tealium x Diabolocom We are excited to announce our new integration with Diabolocom to deliver unified, AI-powered contact center experiences! 🤖✨ By combining real-time identity resolution with award-winning CCaaS, we’re helping brands turn every conversation into a personalized journey. Highlights: 🚀 Real-time intelligent routing 📉 Increased self-service deflection 📈 58% higher sales conversions (case study) Better data = Better conversations. See the full PR: https://lnkd.in/dvvX52_U #Partnership #Innovation #AI #Tealium #Diabolocom #CustomerData

True or False? 🎥 AI that tries to do everything ends up doing nothing well. Sounds obvious? In CX, this mistake still happens all the time. In this short clip, Rémi Guinier, Head of AI Product at Diabolocom, explains why specialized AI is far better suited than generic, one-size-fits-all models for real contact center environments. 🕹️ Want to keep playing and break down a few more AI myths? No tricks. No shortcuts. Just the difference between trying to do everything and doing the right things well. 🔗 Watch the full video on YouTube : https://lnkd.in/eKiEg8n7 Make Your Move.

When AI fails in the contact center, it’s rarely a technology issue. It’s usually a people issue. 👀 Not because humans slow AI down — but because AI only works when humans are fully part of the equation. In CX Today’s latest article, Diabolocom shares a clear point of view: AI doesn’t deliver value on its own. It delivers value when it’s designed, trained, governed, and challenged by people who understand real contact center operations. The idea of fully autonomous AI may sound futuristic 🚀 In reality, it leads to more risk, less clarity, and blurred accountability. Sustainable AI does the opposite. It empowers agents, supports supervisors, and gives “glue employees” the insight and control they need to continuously improve AI performance, day after day. Because in CX, responsibility never sits with the model. It sits with the humans behind it. 👉 Read the full article to explore what sustainable, human-centric AI really means for contact centers : https://lnkd.in/ehYBKux8

Curate–Train–Refine: a closed-loop agentic framework for zero-shot classification We’re excited to share Curate–Train–Refine, a new research work from the Diabolocom Research team on zero-shot text classification 🧠 One of the core challenges in modern NLP is finding the right balance between the generalisation power of large language models and the efficiency of compact encoders. Our approach takes a different path by clearly separating “intelligence” from “inference”. Instead of using LLMs for test-time prediction, we employ them as active data curators in a closed-loop system. Key research highlights: – An iterative agentic loop where the LLM generates synthetic data, trains a lightweight classifier, analyses error modes, and produces targeted examples to continuously improve performance – Strong data efficiency, with synthetic data often matching or exceeding few-shot baselines trained on human-labelled data – Privacy and sovereignty by design, with no test-time dependency on external LLM APIs 🔒 This work shows how LLMs can move beyond prediction and act as supervisors, enabling accurate, efficient, and production-ready models — without the operational cost of large-model inference ⚙️ 👉 To learn more, explore the full paper and the accompanying blog post below. 📄 Paper: https://lnkd.in/eUj33Ymm 📝 Blog post: https://lnkd.in/ezNXmbWG

When AI is too big, CX slows down. ⚠️ In customer experience, speed and reliability matter more than raw power. Yet many enterprises still bet on massive, general-purpose AI models and pay the price with long deployments, rising costs, and unpredictable outcomes. This is where specialized AI changes the game. ⚡ Purpose-built models, designed for concrete CX workflows, are faster to deploy, easier to control, and far more dependable in live contact center environments. In this CX Today article, we share insights on why the future of CX automation isn’t about scale; but about fit. Smaller models. Clear use cases. Measurable impact. 👉 Learn more: https://lnkd.in/eYNc9YVC

Everyone talks about AI in customer experience. Few talk about what actually breaks when it’s deployed at scale. ⚠️ In this interview with CX Today, James Scott, Senior Solutions Engineer at Diabolocom, explains why so many AI projects struggle to deliver value and why the issue isn’t ambition or innovation, but lack of control. When AI can’t be explained, audited, or measured, trust disappears. And so does ROI. James shares how a different approach to AI, built around clarity, governance, and domain expertise, changes the equation for CX leaders who need predictable outcomes, not experiments. 🎥 Watch the interview to understand what CX leaders should really expect from AI — clarity, control, and results. https://lnkd.in/ev2CWQvG