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

Indian digital lender Kissht runs an Arrowhead voice AI agent, in Hindi and English, on two inbound support queues: loan-rejection queries and mandate/NOC cancellations. It is integrated with Kissht's dialer and CRM, hands any mention of the RBI to a human within the same turn, and runs on Arrowhead's own small language model at about 500 ms turn latency. Within two months it handled 40% of all inbound calls, resolving about 88% without a human at roughly 30% lower handle time.

Flock, which insures commercial vehicle fleets, built a pipeline that scores telematics crash files with an XGBoost model on 42 engineered features, then has two Claude Haiku agents on Amazon Bedrock check data quality and the vehicle's claim history before a verified crash is broadcast as an automatic first notification of loss. Every prediction is archived for audit and retraining.

Travelers built AI Claim Assistant, a fully autonomous voice agent on OpenAI's Realtime API and frontier models that handles first notice of loss for auto property damage: it answers policy questions, gathers details and submits the claim. After launching in eight states it went countrywide within two months, and 85-90% of customers who use it finish filing through AI.

Sony Bank and Fujitsu applied Anthropic's Claude on Amazon Bedrock, with Claude Code and MCP-connected agents, to the development of Sony Bank's production core banking system from September 2025. By July 2026 the development period from basic design to integration testing was 30% shorter and man-hours 40% lower, with humans keeping final judgement and quality assurance.

Starling Bank moved its data warehouse to BigQuery and its machine learning to Vertex AI, then built Gemini-based tools: Scam Intelligence checks marketplace ads for scam signs and has raised payment cancellations by 300%, call-transcript summaries for staff save 8,000 hours a month, and a Gemini help centre cut customer service referrals by 50%.

Belgium's AZORG Health System built a single PMcardio AI ECG pathway across six campuses, cutting door-to-balloon time by 48 minutes and STEMI-equivalent delays by six hours, with 32,807 ECGs analysed, 10,854 patients, 377 monthly active clinicians and 50% better guideline adherence.

Swiss digital insurer Simpego and AWS partner copebit replaced manual claims data entry with an event-driven, serverless pipeline that sends each uploaded claim document to Anthropic's Claude Opus on Amazon Bedrock with a strict JSON schema, validates the output with schema and VAT arithmetic checks, and pre-fills the claims platform for specialist review. Built in two months, it handles thousands of documents a month at about 30 seconds each, with 95% needing no manual intervention.

State Farm monitors the generative AI agents that help its chat center representatives by sampling real production conversations, scoring them with Claude Sonnet 4.5 on Amazon Bedrock via DeepEval against reference-free and custom metrics, and synthesising failure patterns with a meta-analysis step. A stricter custom faithfulness metric exposed unsupported inferences in a billing agent, and a prompt fix raised its score 8%; in validation, 78% of judge scores were within 3 points of human reviewers.

Zurich Commercial Insurance uses Cytora as a headless intelligence layer that converts unstructured, multilingual broker submissions into structured risks and feeds them into its Z-Horizon+ underwriting workbench. A shared Intake Master Schema let Zurich go live in five countries in 90 days and 11 by May 2026, with manual triage time per submission falling from 75 to 15 minutes and intake straight-through processing reported at 95%.

Researchers at the SIA-NUS Digital Aviation Corporate Lab built LEAN-LLM-OPT, a three-agent LLM workflow that writes optimisation models from a plain-language query and data. On a booking simulator trained on Singapore Airlines sales data, it reached 93.33% and 92.06% execution accuracy on two revenue management tasks, ahead of Gemini 3 Pro and GPT-5.2.

Envisan-GEM, the Faculty of Civil Engineering at Czech Technical University in Prague and BigHub are developing an AI sorting line for construction and demolition waste: an RGB camera and ML models identify concrete, brick, ceramics and wood on a conveyor, and air jets divert each piece. The data-collection prototype is done; industrial deployment is targeted for 2027.

Czech fintech Patron GO (Antivirus s.r.o.), with 600,000+ users, is building Patron GO 2.0 with BigHub and Czech Technical University: ML models that categorise bank transactions automatically, community self-labelling that retrains them, peer benchmarking, and generative-AI personalised advice. The EU-co-funded project (June 2025 to May 2027) is in a Czech pilot, with Slovakia and Poland as the first foreign markets.

Czech insurers Kooperativa and ČPP (Vienna Insurance Group) worked with BigHub on Data Digger, an AI tool that takes a company registration number and compiles registers, websites, insolvency and land-registry records, property valuations and media signals into one standardised underwriting report. Built in three months on Azure OpenAI, LangChain and Tavily, it is moving from validation toward production, starting with property insurance, and won the VIG Xelerate Audience Award 2026.

Czech electronics retailer PLANEO Elektro deployed a BigHub-built voice AI assistant on its customer line, using ElevenLabs Conversational AI, Azure OpenAI and LangGraph agents with write access to SAP. Of 4,378 calls routed to the voicebot, 52% were resolved without an agent; average call time was 2 minutes 58 seconds.

Since early 2026 the Austrian Parliament's Stenographic Reports Department has used an in-house speech-to-text system for all plenary sittings. Faster Whisper produces the raw transcript, Pyannote detects speaker changes and regex post-editing applies house style. IT developers and a department stenographer built it in about one and a half years; typists now process about twice the text per hour.

Per Anthropic's 15 October 2026 engineering blog, the Claude Code team runs a six-layer code review process with risk-tiered human approval; between June and September 2026, merged PRs to the Claude Code codebase more than doubled while the share that had to be reverted fell by more than 40 percent, and the share of internal PRs getting substantive review comments went from 16% to 54% after the launch of the Code Review for Claude Code research preview. Layer 5's classifier-driven risk tiers (Low / Medium / High) and the rule that an agent's approval never counts as a human's are the structural pieces that let most PRs merge with at most one human in the loop.

Suncorp, an ASX-listed Australian insurer, put Fujifilm Content AI into production to classify and name workers compensation claim documents, part of a workload of over 1.1 million documents a year. Average daily naming corrections fell about 45% (4,297 to 2,345), and the solution is projected to save 68,000 hours of manual effort a year.

Per OpenAI's 16 October 2026 customer story, Stack Overflow's ChatGPT plugin for Stack Overflow for Agents drove a 3× overnight jump in signups and up to 10× growth in agent creation; two engineers built the Agents product in a few weeks with OpenAI Codex, and connector build time dropped from 3-6 weeks to 4 days (an 87% maximum reduction), with OpenAI models like GPT-6 Astra and GPT-6 Sol routing tasks across the platform.

Current, an AI-native tax platform for accounting firms, embedded Nutrient's document infrastructure to consolidate PDF viewing, editing, annotations, and signing in-app. Two engineers integrated; 30-minute initial render; working product for extension tax season in six weeks; full production launch before new year. Roughly $50,000–$100,000 in annual third-party document software savings across participating firms (per Nutrient case study).

Chime Financial deployed Anthropic Claude via Amazon Bedrock to automate customer call summarization, cutting average agent handling time by 18 seconds per call and saving 250,000+ hours annually (~$700K in efficiency gains), with NPS rising 5 points after the February 2024 production launch.