- Jean-Louis Quéguiner - Founder of Gladia: A tool using Speech AI for generative audio, with a focus on language models and high-speed transcription.
- Olivier Moindrot - CEO and Co-founder of Fragment: An AI-powered task management system for human-in-the-loop operations, emphasizing the collaboration between AI and human agents.
- Fadeelah Al-horaibi - Head of Product at Slite: A knowledge base on autopilot that has just launched a major update to their AI assistant Ask that now connects with external resources to give answers based on all internal communication.
- The panel was moderated by Louis Steenbrink, Forest Admin's Director of Revenue, and followed by a demo from Forest Admin's CEO, Sandro Munda, who showed how AI can speed up customer operations.
To write this summary we used Gladia, Slite, ChatGPT and a "human in the loop" :)
Watch the recording or read the summary below 👇
Scaling Customer Operations with AI
First off, all the panelists shared their experiences and insights on the real-world applications of AI in their respective domains.
Olivier talked about his experience at MetaMap, a customer identity platform, where they focused on increasing the automation rate of identity verification. He stressed the need to break down complex processes into smaller tasks for effective automation, always involving humans for optimal performance.
Our main initial [use cases] have been fintech and insurance because there were some repetitive manual tasks that need to be done. But another area that is very interesting for us is whenever you have a company that is doing automation, they have their own AI team, but they still need to have humans in the loop because the AI is not perfect. - said Olivier.
Jean-Louis discussed the primary use cases for Gladia, including meeting recordings, multilingual support (which is, by the way, harder for AI than people tend to think) and applications in the medical industry, particularly in tele-consultations, where AI assists in note-taking and summarizing patient-doctor conversations.
The top [use case] I would say is meeting recordings. Nobody wants to take notes anymore, we are getting lazy. It's like the GPS - nobody opens the map, we use GPS. - said Jean-Louis
Fadeelah explained how Slite's AI-powered knowledge base, with its Ask feature, transformed customer operations. This AI system allows users to ask questions in natural language, providing human-like answers and significantly reducing the reliance on human support for routine queries.
[Slite's Ask] has saved people a lot of time with being autonomous and independent in their roles, which ultimately leads them to feeling successful. - said Fadeelah.
The Importance of Human Involvement
Throughout the discussion, the panelists emphasized the crucial role of humans in AI operations. They highlighted scenarios where AI may struggle, such as detecting sentiment or handling creative problem-solving. The concept of "human in the loop" was recurrent, underscoring the idea that while AI can automate many tasks, human oversight and intervention are essential for optimal outcomes.
Since ChatGPT came in, the price of human annotation has exploded. Literally exploded. And one of the most rare resources right now is human annotation. (...) AI won't replace humans. - said Jean-Louis from Gladia.
Challenges of Using AI in Production
In the third part of the panel discussion, the panelists discussed challenges of building AI products. Fadeelah outlined three key challenges:
- Assessing Feature Quality, as traditional user research methods may not be directly applicable, and nuances in understanding the quality of answers require internal testing, refining questions, and observing user behavior.
- Users' preconceived understanding of AI tools and their expectations can impact the effectiveness of the AI system. Educating users about what type of knowledge should be provided, how it should be structured, and addressing security concerns becomes crucial.
- Finding the right prompts for AI systems remains a challenge. Users often struggle to understand how to phrase questions or prompts effectively.
Jean-Louis talked about the security challenges when shipping AI projects to production. He mentioned the concept of "prompt escaping," where inappropriate questions can cause the AI system to deviate from its intended domain. Guardrailing becomes essential to ensure the AI stays within its designated scope. Additionally, he highlighted challenges related to GPU and hardware complexities, and the difficulty of ensuring 100% secure operations.
All panelists stressed the importance of continuous measurement and a mix of quantitative and qualitative approaches to ensure AI system accuracy. They discussed the potential impact of voice as a major mode of data ingestion, emphasizing its significance in preserving and utilizing knowledge exchanged in daily interactions.
Looking ahead, they predict that voice will play a transformative role in data knowledge. They foresee a paradigm change in search engines, with companies possibly having to pay for indexing and ranking, challenging traditional SEO practices. The emergence of companies built around AI as a core component is noted, signifying a shift in how businesses are structured, with tech expertise leading the way.
What about AI in Forest Admin?
The last part of the event was a demo given by Sandro Munda, the CEO of Forest Admin, emphasizing the role of AI in enhancing internal operations. Forest Admin has explored various industry-specific processes and AI use cases through interviews with customers, categorizing them into data extraction, data enrichment, data computation, and data entry.
Sandro showed a specific use case related to onboarding processes. The demo involved using Forest Admin AI to detect identity information from documents, such as driving licenses or passports. The tool automatically extracted details like names, ages, and associated them with the relevant companies. However, it was emphasized that AI, while powerful, requires human verification, especially in sensitive processes like KYC.