AI Atlas - Daily updates on real-world AI deployments worldwide

SMBC Americas migrated a core risk-data workload of 170 attributes to Azure Databricks, where re-architected Delta Lake pipelines reduced shuffle I/O by 67% and improved runtime by 40-60% over the legacy Hadoop-based risk data platform. The migration gave SMBC Americas a single governed lakehouse for cross-region risk analytics and a foundation for AI-driven credit and treasury workflows.

Superhuman and Databricks jointly engineered a GPU inference platform that handles more than 200,000 queries per second for Superhuman's email AI features (Write with AI, Ask AI, Auto Summarize, Instant Reply). The platform combines elastic data-streaming (EDS) power-of-two load balancing with image accelerator and FP8 quantization to push per-pod throughput up by about 60% while keeping p99 latency low.

PepsiCo is replacing a 27-year-old, fragmented BI landscape by moving to a single, governed analytics and AI platform on Azure Databricks SQL and Unity Catalog. A key sales-analytics workload dropped from about $500K to $175K per year (about 80% lower) while running faster, and processing of core financial and commercial tables improved by around 50% across Power BI and Tableau.

Hospital for Special Surgery, the New York orthopedic academic medical center, unified 40-plus source systems and 14,500+ tables spanning 130+ schemas into a single Databricks lakehouse over 10 months. The consolidation gives finance, operations, clinical research, and quality teams a trusted view of the business and lays the groundwork for AI-driven clinical and operational analytics.

Lumen Technologies deployed a generative AI network diagnostics platform, built on Databricks and Meta Llama 3.3, that helps contact-center engineers resolve 3 million-plus annual service incidents faster. The system has lifted Tier-1 deflection rate from 30% to a 50% target and targets savings of 3,675 engineering hours per year on diagnostics and field-dispatch calls.

CAVA deployed Ask Astro, a Databricks multi-agent supervisor, to let employees query business data in plain English. The system routes questions across four Genie spaces, Atlassian APIs and web search, serving thousands of daily users with production guardrails covering PII and content safety.

SAP deployed OpenText Content Management for SAP SuccessFactors as the document layer for its global SAP SuccessFactors Employee Central rollout, migrating 8.5 million HR records, reducing document types by 96% (2,000 to 70), and automating GDPR compliance across 105,000+ employees and more than 100 legal entities.

IBM consolidated its global HR content layer on OpenText Content Management for SAP SuccessFactors, with OpenText Content Aviator as the AI orchestration layer and the watsonx-powered AskHR chatbot as the employee experience. The deployment supports 280,000 employees across dozens of countries and 31 languages, delivering 99% faster document execution, 94% inquiry containment, and 40% lower HR operating costs.

Novartis deployed Market Logic Software's Sherlock + DeepSights generative-AI platform to centralize market insights across 7,500 internal users, cutting primary research spend by 56% ($30M annually) and saving two hours per major research query.

Japan's avatarin built a 24/7 multilingual voice-AI retail agent on OpenAI GPT-Realtime and deployed it at Yamada Denki (one of Japan's largest electronics retailers). 30,000 shoppers engaged in the first two weeks, with 92% satisfaction.

Ring built a semantic video search system on Amazon RDS for PostgreSQL with the pgvector extension that stores 100–200 billion vector embeddings across 3 PostgreSQL clusters holding 140–150+ TB, ingests roughly 2 billion new embeddings per day, and serves billions of read requests daily across 9 AWS Regions on 4 continents while holding query latency under 200ms p50 and under 2 seconds worst case. The CLIP-based AI Video Search converts video frames into vector embeddings and matches natural-language queries like 'package delivery' or 'someone wearing a blue shirt' against them.

Blue Origin used BlueGPT, its internal multi-agent AI platform built on Amazon Bedrock, the open-source Strands Agents SDK, and Amazon EKS, to design the TEAREx thermal battery for lunar hardware with a team of specialized AI agents working alongside 2–3 human engineers — compressing the hardware development cycle by roughly 90% and accelerating analysis workflows by 6x. BlueGPT now hosts more than 2,700 agents that drive roughly 3.5 million interactions per month, with 70% company-wide adoption.

Rakuten France, the French e-commerce marketplace subsidiary of Rakuten Group, runs 80+ Custom Agents inside Notion that automate more than 58,000 workflows company-wide and cut backlog health issues by 50% on the product team. AI Lead Sidney Golstein built a meta-agent called the Rakuten Agent Builder so any employee can spin up a new agent in 10–20 minutes from a plain-language description, with shared SOPs, an Agent Registry, an Agent Logs database, and a separate Delicious Improver agent closing the loop by reviewing logs and proposing updates.

Česká spořitelna's Future Lab branch in Prague runs the Czech Republic's first 24/7 AI-powered bank lobby, where a Promethist Engage digital human greets every visitor in 20+ languages, qualifies intent, and summons bankers with full context. Innovations from the lab are now feeding into Česká spořitelna's 320 branches nationwide and across Erste Group's seven-country network.

Atlassian is shipping four governed agent loop features across Jira and DX that automate the path from backlog to pull request, backed by its 2026 AI SDLC study showing 94% of engineering leaders use AI but only 6% have systems to scale it across the lifecycle. DX analysis of teams using the most Atlassian Teamwork Graph context found they shipped roughly 64% more per developer.

Atlassian deployed a distributed agent harness for Rovo Chat with split-plane architecture, programmatic tool calling ('code mode'), and interactive sub-agents to power an always-on AI assistant for Jira and Confluence workflows; internal evaluations showed code mode cut latency of complex Jira queries by over 50%, reduced token consumption by 55%, and improved accuracy by 30% against a baseline meta-tool harness.

Salesforce's Data 360 team converted roughly 100,000 unstructured Informatica documents into an AI-ready knowledge base in 24 days, beating a 30-day post-acquisition deadline by reusing established ingestion, chunking, and retrieval patterns instead of rebuilding. The resulting Informatica Help Agent resolves 80% of customer questions with only 5% human escalation.

Figma's security engineering team built a dual-model agent system (Claude Code Opus 4.8 + Codex GPT-5.6 Sol) on a precedent-based threat-model policy that catches 75.8% of bugs slipping past human reviewers plus static analysis, at a median cost of $0.50 per PR review. The same policy drives code generation and monorepo-wide auditing, and self-improvement loops update it from HackerOne bug reports and PR-author feedback.

Figma's security engineering team built a multi-layer agentic system on Claude Opus (triage) plus Claude Sonnet (intent routing) orchestrated through Tines, with AWS Bedrock + Kendra for RAG and a Snowflake-backed security data lake, that handles all initial SIEM alert response and auto-files PRs. Outcomes: 71% reduction in time-to-resolution on complex alerts, 20% drop in on-call pages, 25% fewer endpoint software approval requests.

Nokia's Core Networks division is rebuilding its software development life cycle around Cursor as a multi-agent platform. Two engineers analyzed 50M+ lines of code in two weeks (work that would have taken months and a dozen+ experts); one engineer built a PRD-to-tool automation in under a week that automated ~80% of a 6–10 PM workflow; root-cause analysis on customer-impacting defects dropped from weeks to days; SDLC being redesigned around persona-specific agents.