Publicis Sapient CEO Nigel Vaz: 93% of enterprise AI pilots fail to scale — here's why and what the fix looks like

Summary

An MIT study finding that 93% of enterprise AI pilots fail to scale is not a surprise to Nigel Vaz, CEO of Publicis Sapient, a 30-year-old enterprise technology firm whose clients include Goldman Sachs, Walmart, Tesco, Carrefour, and major German, US, and Asian automotive manufacturers. Vaz argues the failure rate is structural, not conceptual: large enterprises are running disconnected pilots on separate ERP systems, meaning AI agents built for sales never communicate with agents built for marketing, effectively digitising the same siloed, baton-passing organisational model they already have.

The Scale Problem

The core barrier is tool-chain fragmentation. When individual engineers adopt different AI coding environments permissionlessly — as happened widely in 2024 — early productivity gains don't compound into enterprise-wide output. In large organisations, shared application context, authentication, and permissions governance are prerequisites that consumer-grade coding tools were never built to handle. Vaz points to a healthcare client on a 10-year COBOL modernisation programme as a concrete example. Publicis Sapient's platform Slingshot is compressing that timeline to 2–3 years by ingesting legacy code, generating human-reviewable specs, and migrating to modern code within a governed, unified tool chain. The COBOL in question dates to the 1960s and 70s, and the healthcare data is HIPAA-regulated, making ad-hoc AI tooling legally and operationally unviable.

Vaz also cites a real-time insurance use case where a client deployed agentic systems to monitor post-wildfire air quality, proactively alerting high-risk asthma patients and saving the insurer hundreds of millions of dollars in healthcare costs. A separate automotive client used agentic orchestration to compress a car build cycle from 18 months to 18 weeks.

Data Strategy Precedes AI Strategy

Vaz is direct that most enterprises claiming an AI strategy lack the prerequisite: a unified data strategy. Hundreds of millions of dollars in accumulated tech debt sit beneath most large organisations, and disconnected data sets mean AI initiatives compound existing fragmentation rather than resolve it. Enterprise leaders at forums like Davos — coming up in a matter of weeks — have moved past debating AI's relevance and are now pressing for practical value extraction within their existing operational contexts.

AI in Advertising and Personalisation

On the advertising side, Vaz argues the bottleneck is not consumer willingness to share data but enterprise inability to act on it. He points to work with Marriott on intent-capture for its villas platform, shifting from structured search fields to open-context inputs that surface genuinely personalised results.

A more operationally significant example involves a large pharmaceutical client launching a vaccine across 115 countries. The conventional compliance-to-creative cycle runs approximately 12 months as marketers and lawyers negotiate country-specific advertising regulations. In New Zealand, for instance, smiling in a vaccine ad is prohibited because it implies a guaranteed outcome. Publicis Sapient's agentic platform ingests all 115 regulatory frameworks, tests creative concepts against them in near real time, regenerates compliant versions using the most efficient available models, deploys at scale, and runs continuous monitoring agents that feed into live A/B testing. The same orchestration architecture connects supply chain inventory data to advertising, preventing a scenario where a sold-out product continues to receive active ad spend during peak retail events like Black Friday.