Delivering AI features with editorial restraint and platform foresight
Replacing Google Search across The Economist's platforms whilst shipping AI features required strategic restraint: recognising which capabilities were becoming commoditised, designing brand-appropriate AI integration, and building extensible infrastructure rather than isolated features.
Launched on web (November 2024) and mobile (January 2025) using DeepSet AI and Contextual AI for prototyping, with OpenAI GPT-4o plus custom layers for production. The platform supports semantic search, article summaries and future conversational features including citations and multi-media source integration.
Early metrics show 1% search usage and 1% article summary adoption – honest numbers revealing user behaviour patterns. The work balanced competing priorities: delivering AI capabilities users increasingly expect whilst ensuring editorial journalism remained primary, maintaining brand positioning and ensuring consistency across web, apps, newsletters and service pages.
Strategic restraint: protecting editorial primacy
User research revealed mixed sentiment about AI at The Economist, requiring confident delivery with careful positioning. The challenge wasn't implementing AI features—it was ensuring they augmented journalism rather than competing with it.
Key strategic decisions shaped the integration:
Initial proposals suggested default-open summaries at the top of articles, pushing editorial content down. This risked devaluing journalism from the newsroom, with visitors potentially confusing articles as AI-generated at a glance. Instead, I aimed to make summaries accessible but not primary, sitting within the article meta bar alongside share functions.
The now ubiquitous AI sparkle icons and "AI-powered" labels didn't fit The Economist's analytical tone. It's tools, not magic. Using verb-based microcopy ("summarise", "translate") better suited the current and future features across our products.
Editorial found the subtle integration protected content prominence whilst making AI capabilities discoverable. The semantic search implementation addresses reader and editorial needs whilst maintaining clear boundaries: AI augments expertise, never replaces editorial judgment.
Platform architecture for future AI capabilities
Article summaries and semantic search were recognised as table stakes – baseline expectations rather than differentiation. The strategic opportunity was building infrastructure enabling more sophisticated AI integration: conversational interfaces, multi-turn interactions, citation systems, and multi-media source connections.
Decisions couldn't be made in isolation. The labeling system, interaction patterns and architecture needed to work across web, iOS, Android, newsletters, and service pages. This required collaborative design with platform teams to ensure consistency without rebuilding from scratch for each context.
The platform avoids generic chatbot patterns. Instead, it supports:
• Follow-up questions and multi-turn conversations
• Citation systems connecting to source material
• Multi-media source integration (images, video, quotes)
Early adoption sits around 1% for both search and summaries. Core engagement metrics (article opens, reads, referrals) showed no negative impact, confirming the priority: AI augments journalism consumption rather than replacing it.
External partnerships and transparent implementation
Technical implementation combined external AI partnerships (DeepSet AI, Contextual AI for prototyping) with internal custom layers built around OpenAI GPT-4o for production. External partnerships enabled early collaboration and stakeholder buy-in ahead of internal AI development capabilities being ready.
Production development insights:
• Search infrastructure criticality—when it breaks, user complaints highlight dependence despite modest 1% adoption
• Frequently covered topics create ongoing sorting and relevance challenges requiring continuous refinement
• The gap between prototype enthusiasm and production reality: what works in demos often needs significant adjustment at scale
• Editorial content already being analysis means AI summaries risk losing nuance rather than adding value
Shipping to millions of users established frameworks for ongoing AI integration: what to build in-house versus partner for, where AI genuinely augments journalism versus where it commoditises it, and how to measure success beyond vanity adoption metrics.