Architects of Transformation: Foundational Models in Generative AI | RAISE Summit 2024 | Paris

Panel Context & Participants

• Moderated by Daphné Leprince-Ringuet (Sifted), the panel "Architects of Transformation" focused on the gap between enterprise expectations and the reality of foundational AI models.
• Pablo (Rave, US): Builds multimedia foundational models prioritizing IP rights and legal compliance.
• Milos (DeepSets, Berlin): Provides open-source platforms for enterprises to build customized LLMs.
• Nan (Gena, Berlin): Offers open-source tools for neural search and retrieval-augmented generation (RAG).
• Marco (Defector, Paris): Fintech providing instant financing to SMBs, utilizing AI agents for operations.

Enterprise Expectations vs. Reality

• Legal Compliance & IP: Businesses prioritize legal safety over raw capability; many avoid open-source models (e.g., Stable Diffusion, MidJourney) due to training data uncertainty, shifting instead to Adobe Firefly's licensed data, despite mixed satisfaction with brand consistency.
• Performance Metrics: Organizations expect deterministic quality assurance and clear metrics (e.g., "99.9% success rate") similar to classic software, but generative AI remains stochastic, requiring acceptance of inherent uncertainty.
• Product Integration: There is a misconception that deploying a generative AI feature guarantees business success (leads/clicks); experts note AI must be embedded within a broader product strategy rather than acting as a standalone solution.
• Amara's Law: Customers and executives tend to overestimate the immediate impact of GenAI (over the next 16 months) while underestimating its long-term structural integration.

Current Use Cases & Business Implementation

• AI as "New Employees": Defector utilized AI agents to double origination (financing volume) while hiring only four new staff members in one year, specifically for underwriting document review and automated repayment collection.
• Sales Automation: AI agents were deployed to close inbound sales leads automatically, eliminating the need for dedicated sales representatives for initial qualification.
• Cost & ROI Hurdles: Decision-makers hesitate to adopt production-grade RAG solutions (e.g., Llama-70B) due to high infrastructure costs (thousands of dollars/month) without guaranteed immediate ROI.
• Data Silos: A primary technical challenge remains the difficulty of integrating segregated enterprise data into AI agents for personalized performance.

Barriers to Adoption

• Talent Mismatch: Companies often hire research-focused AI engineers instead of product managers; success requires teams capable of "applying" models to specific workflows and designing engaging user experiences rather than just training foundational models.
• Risk Aversion: Decision-makers struggle with the non-deterministic nature of AI, fearing liability for the "bad 10%" of outputs in high-stakes industries like finance.
• Testing Infrastructure: Lack of robust sandbox environments and validation tools makes it difficult for companies to iterate safely without risking system stability or regression.
• IP & Transparency: The financial services sector requires 99.999% reliability for tasks like payment reconciliation; currently, companies rely on human-in-the-loop systems where AI proposes resolutions subject to operator verification to meet regulatory SLAs.

Future Trends & Capabilities

• Controllability: The next frontier for image models is moving from "scariness/creativity" to "brand consistency" and precise output control to match IP requirements.
• Video Generation: Video is identified as the upcoming frontier, though personalization and cost barriers mean widespread enterprise adoption is expected later this year or next.
• Open Source Sustainability: Experts warn that true open source requires access to original training data, not just model weights; without this, the ability to maintain model capabilities after fine-tuning is compromised.
• Ecosystem Shift: A potential move toward "walled gardens" is occurring where businesses pay for licensed, IP-safe models (e.g., Adobe) rather than free open-source alternatives due to trust and legal risk concerns.

Talent & Workforce Strategy

• Engineering Needs: Requires personnel with familiarity in adopting models and building agentic systems, rather than pure research talent.
• Go-to-Market Challenge: Finding sales and product staff comfortable navigating an early market with high awareness but low adoption and no established playbooks is a significant hiring bottleneck.
• Geographic Dynamics: While the US talent pool is larger, European startups compete by offering meaningful missions, such as building "legal-first" creative tools for the media industry.