Case Studies

17LIVE · 2025

Offering more interactive, personalized live-streaming content with Gemini

30% increase in user satisfaction; 30% fewer SRE resources

Challenge

Because all its content is live and unrecorded, 17LIVE lacked visibility into what happened in streaming rooms, making troubleshooting slow and personalization difficult.

Solution

17LIVE adopted Gemini and Vertex AI (plus Veo, Vertex AI Vision, and BigQuery) to analyze feedback and logs in real time, run a Gemini-powered incident agent, build AI cohosts, and generate viewer and streamer personas.

Results

Faster, more accurate troubleshooting drove a 30% increase in user satisfaction and required 30% fewer site-reliability engineering resources, while persona-based personalization raised comment rates 10% and follow rates 20%.

Leading to a 30% increase in user satisfaction.

• Gemini
• Vertex AI
• Veo
• Vertex AI Vision
• BigQuery

Source: Google Cloud Customer Stories · Google Cloud ↗

Advantage Solutions · 2025

Advantage Solutions gives frontline managers 70,000 hours back with Claude

70,000+ labor hours redirected annually

Challenge

Manual compliance checks, finance forecast validation, and workforce communications across a large frontline organization consumed enormous labor hours.

Solution

Advantage Solutions deployed Claude Enterprise and Claude Code across operations, finance, and communications — automating photo-based compliance verification, forecast validation, and a store-routing model.

Results

Redirected more than 70,000 labor hours annually from manual compliance checks, compressed finance forecast validation from 10 hours weekly to under 30 minutes, and scaled from a 150-person pilot to thousands of users.

That's real people, real shifts, and real time that was going back to our teammates.

• Claude Enterprise
• Claude Code

Source: Anthropic Customer Stories · Anthropic ↗

Affinda · 2025

Affinda cuts document-processing setup time by 90% with Amazon Bedrock

90% reduction in setup time for new document-extraction use cases

Challenge

Affinda's traditional machine-learning document processing required extensive training data and weeks-to-months of configuration to onboard each new document extraction use case.

Solution

Affinda re-architected its intelligent document processing on Amazon Bedrock large language models (Anthropic Claude Sonnet 3.5 v2, Claude 3.7 Sonnet, and Claude Sonnet 4), alongside Amazon SageMaker and Amazon EKS.

Results

The generative AI approach delivered about a 90% reduction in configuration time for new use cases and roughly 90% cost savings, cutting customer configuration from weeks to minutes.

We've seen a 90 percent reduction in configuration time for new document extraction use cases.

• Amazon Bedrock
• Anthropic Claude Sonnet 3.5 v2
• Anthropic Claude 3.7 Sonnet
• Amazon SageMaker
• Amazon EKS

Source: AWS Customer Stories · Amazon Web Services ↗

CareSource · 2025

CareSource scales member care with Azure OpenAI automation

Documentation process cut from two months to two weeks; over $125,000 saved

Challenge

Rapid growth and an expanding member base created operational complexity, and CareSource needed to scale while keeping care personalized and human-centered.

Solution

The managed-care organization adopted Azure OpenAI Service, Dynamics 365, Azure Functions and Cosmos DB, and GitHub Copilot to automate documentation and streamline workflows.

Results

A key documentation process dropped from two months to two weeks, the organization saved over $125,000 through automation, and developers reported 20-30% productivity gains with GitHub Copilot.

Reduced a key documentation process from two months to two weeks.

• Azure OpenAI Service
• Dynamics 365
• GitHub Copilot
• Azure Functions
• Azure Cosmos DB

Source: Microsoft Customer Stories · Microsoft ↗

Cox Communications · 2025

Cox Communications scales Claude to a 7x first-year ROI

7x ROI on first-year AI investment

Challenge

Cox needed to scale AI across a 15,000-person workforce to strengthen B2B marketing and sales and empower non-engineers to build their own solutions.

Solution

Cox deployed Claude Code and Claude Cowork company-wide and built multi-agent B2B marketing and sales systems (lead validation, campaign creation, seller enablement) running on Claude through Amazon Bedrock.

Results

Reported a 7x ROI on its first year of AI investment, with about 2,500 Claude Code users — more than half of them outside engineering.

We now know we got a 7x ROI on our first year of investment.

• Claude Code
• Claude Cowork
• Amazon Bedrock

Source: Anthropic Customer Stories · Anthropic ↗

CRESCO LTD. · 2025

CRESCO saves ~10 hours per user monthly with Microsoft 365 Copilot

Saved the average user approximately ten hours of work per month

Challenge

Executives and staff faced time-consuming manual work: interview data entry took 30 minutes per candidate and contract reviews required extensive back-and-forth with legal teams.

Solution

CRESCO created a Generative AI Business Transformation Lab and rolled out Microsoft 365 Copilot company-wide (scaling from 30 to 1,000 licenses), integrating it with Teams, Word, and Excel.

Results

Copilot saved the average user about ten hours per month, cut interview data processing from 30 minutes to 5 minutes per candidate, and reached over 90% adoption within six months.

It saved the average user approximately ten hours of work per month.

• Microsoft 365 Copilot
• Azure OpenAI
• GitHub Copilot

Source: Microsoft Customer Stories · Microsoft ↗

Etsy · 2025

Etsy connects ~90 million buyers with special items using gen AI and 'algotorial curation'

SEO-driven visits +5% and conversions +3%; ~80x more listings per theme

Challenge

Etsy needed to match its nearly 90 million shoppers with the right items from a daily-changing inventory of more than 130 million listings offered by over 5 million sellers.

Solution

Etsy used Gemini models, the Gemini Enterprise Agent Platform, BigQuery, and Dataflow to enrich inventory data, infer shopper intent, generate listing alt text, and personalize search and discovery at scale.

Results

Gemini-driven alt text lifted SEO-driven visits 5% and conversions 3% for sellers, while 'algotorial curation' produced roughly an 80x increase in listings per theme.

Etsy improved alt text generation for listings, increasing SEO-driven visits by 5% and boosting conversions by 3% for our sellers.

• Gemini
• Gemini Enterprise Agent Platform
• BigQuery
• Dataflow

Source: Google Cloud Customer Stories · Google Cloud ↗

Golden Energy Mines (GEMS) · 2025

Accelerating pit-to-port mining decisions by over 90% with Gemini

Executive decision-making accelerated by over 90%

Challenge

GEMS's roughly 50 fragmented application dashboards created a cognitive bottleneck, leaving executives with a siloed view that made fast, cross-functional pit-to-port decisions difficult.

Solution

GEMS built GEMVIS, a hierarchical multi-agent system using a Gemini Dispatcher Agent and the Gemini Enterprise Agent Platform (on GKE, Compute Engine, and Cloud Armor) with a hybrid on-prem RAG architecture to unify data across all applications.

Results

GEMVIS accelerated executive decision-making by over 90%, cut multi-operational data retrieval from two days to under one hour, unified insights across 50+ portfolios, and empowered more than 4,000 users.

Accelerated executive decision-making speed by over 90% using Google Gemini.

• Gemini
• Gemini Enterprise Agent Platform
• Google Kubernetes Engine
• Cloud Armor

Source: Google Cloud Customer Stories · Google Cloud ↗

Immigration, Refugees and Citizenship Canada · 2025

IRCC Artificial Intelligence Strategy for immigration processing

Challenge

IRCC processes very high volumes of immigration and citizenship applications and needs to improve efficiency and program integrity while protecting client privacy and fairness.

Solution

Since 2018 IRCC has used advanced analytics and machine learning on applicant data; its AI Strategy applies AI to triage applications, create summaries, respond to enquiries, and flag low-risk files for expedited officer decisions — the tools never refuse applications.

Results

AI helps employees complete routine administrative tasks faster so they can focus on complex assessments of risk, fraud, and admissibility; the inaugural strategy aligns with the AI Strategy for the Federal Public Service 2025-2027.

Since 2018, the department has used advanced analytics and machine learning to help us gain deeper insights from program and applicant data.

• Machine Learning
• Advanced Analytics
• Automation

Source: Immigration, Refugees and Citizenship Canada ↗

LG CNS · 2025

LG CNS modernizes 20-year-old enterprise systems with Claude

99.1% of APIs converted (2,888 of 2,913) at ~50% lower cost

Challenge

LG CNS needed to modernize decades-old legacy enterprise systems (including 20-year-old MiPlatform applications) — a high-risk, hard-to-start migration.

Solution

LG CNS built its 'Build Factory' service with Claude Code (accessed through Amazon Bedrock, relying on Claude Opus 4.6) as the engine, wrapping it in an orchestration layer for legacy analysis, code generation, testing, and QA.

Results

Converted 2,888 of 2,913 APIs (99.1%), migrated 1,340 screens to React, and delivered the seven-month migration at roughly 50% of the cost of a conventional rebuild.

The ROI is that we made a previously difficult-to-start project possible.

• Claude Code
• Claude Opus 4.6
• Amazon Bedrock
• React

Source: Anthropic Customer Stories · Anthropic ↗

Matillion · 2025

Matillion builds an AI-powered data engineer on Snowflake

50% reduction in analyst workload; ~40% lower infrastructure costs

Challenge

A small data team needed to scale complex data-engineering and self-service analytics without proportionally growing headcount.

Solution

Matillion built 'Maia,' an AI agent for data engineering, on the Snowflake AI Data Cloud, using Cortex AI (Cortex Search, Cortex Sentiment, Semantic Views) and Snowflake Intelligence for natural-language analytics via Slack.

Results

Cut infrastructure costs by roughly 40% and reduced analyst workload by 50% — the equivalent of saving one full-time analyst.

It's allowed us to grow exponentially as a business, even with a small data team.

• Snowflake AI Data Cloud
• Snowflake Cortex AI
• Cortex Search
• Semantic Views
• Snowflake Intelligence

Source: Snowflake Customer Stories · Snowflake ↗

National Oceanic and Atmospheric Administration · 2025

Project EAGLE: AI-based global weather-forecast models

Challenge

NOAA needs a way to rapidly test and adopt fast, skillful AI weather models to complement its computationally expensive physics-based global forecast systems.

Solution

Through Project EAGLE — NOAA Research, the Earth Prediction Innovation Center, and the National Weather Service — NOAA builds AI global and ensemble forecast systems by fine-tuning DeepMind's GraphCast on NOAA's own data, running 31-member ensembles to produce twice-daily 16-day forecasts.

Results

The initiative lets NOAA rapidly test and demonstrate AI forecast models (Global-EAGLE-Solo, Global-EAGLE-Ensemble, and the HRRRCast regional emulator) to identify the best-performing AI innovations for operational forecasting.

Project EAGLE will allow NOAA to identify the best performing AI innovations rapidly.

• Machine Learning
• GraphCast
• Ensemble Weather Modeling

Source: NOAA Earth Prediction Innovation Center · Google DeepMind ↗

Novo Nordisk · 2025

Novo Nordisk scales to 2,500+ generative AI use cases on Amazon Bedrock

2,500+ generative AI use cases built by 25,000+ employees

Challenge

Novo Nordisk wanted to democratize AI-driven innovation across the enterprise while keeping data secure across nonregulated business processes.

Solution

Novo Nordisk built a secure self-service generative AI platform on Amazon Bedrock (with AWS Lambda and Amazon DynamoDB) that lets employees create their own chatbots and agents.

Results

More than 25,000 employees built over 2,500 use-case chatbots at about $10 per chatbot per month, with some day-long tasks completed in 30 minutes.

The democratization of innovation has been the biggest outcome.

• Amazon Bedrock
• AWS Lambda
• Amazon DynamoDB

Source: AWS Customer Stories · Amazon Web Services ↗

Reveleer · 2025

Bringing value-based healthcare to clinicians with Vertex AI

Up to 50% less noise than legacy NLP

Challenge

Clinicians often have only minutes to review a patient's history, so Reveleer needed to extract and reason over structured and unstructured records to surface high-value care gaps at the point of care while reducing false positives.

Solution

Reveleer built a prospective risk-adjustment engine on Vertex AI, using an agentic pipeline of Gemini Flash and Flash-Lite models (with Cloud Run, Firestore, and BigQuery) to interpret clinical charts and standardize evidence.

Results

The Gemini-based engine delivers fewer, more relevant suspects with up to 50% less noise than legacy NLP, improving accuracy, cutting false positives, and letting customers update logic in minutes instead of months.

Delivering fewer, more relevant suspects — with up to 50% less noise than legacy NLP.

• Vertex AI
• Gemini Flash
• Gemini Flash-Lite
• BigQuery
• Cloud Run

Source: Google Cloud Customer Stories · Google Cloud ↗

Robinhood Markets · 2025

Robinhood transforms financial-crime investigations with Amazon Bedrock

Up to 20% cumulative efficiency gain in investigative workflows

Challenge

Robinhood needed to scale its financial-crimes (FinCrimes) investigations to keep pace with rapid platform growth and evolving risks while remaining compliant with regulations.

Solution

Robinhood built its FinCrimes Agent on Amazon Bedrock, using Anthropic Claude (Haiku and Sonnet) and DeepSeek with orchestration via Amazon RDS, to synthesize customer and transactional data into investigative summaries.

Results

The FinCrimes Agent put relevant information at investigators' fingertips and delivered up to 20% cumulative efficiency gains in investigative workflows.

The FinCrimes Agent has put all relevant information at our investigators' fingertips, helping them to quickly identify illicit activity with increased focus and accuracy.

• Amazon Bedrock
• Anthropic Claude Haiku
• Anthropic Claude Sonnet
• DeepSeek
• Amazon RDS

Source: AWS Customer Stories · Amazon Web Services ↗

Scribd · 2025

Scribd turns content into sign-ups and scale with generative AI on Databricks

90% reduction in generative-AI operational costs; 7% more sign-ups

Challenge

Scribd needed to improve content discovery across a 250M+ document library while controlling the cost of generative AI operations.

Solution

Scribd unified its data on the Databricks Data Intelligence Platform and deployed open LLMs (Llama, Mistral, Claude) via Databricks Model Serving for auto-generated metadata, semantic search, content moderation, and an 'Ask AI' chat discovery feature.

Results

Achieved a 7% increase in new-user sign-ups, a 7% reduction in churn, and a 90% reduction in generative AI operational costs.

Our data scientists are able to test, tune and deploy in one continuous flow.

• Databricks Data Intelligence Platform
• Databricks Model Serving
• Llama
• Mistral
• Claude

Source: Databricks Customer Stories · Databricks ↗

U.S. Department of Veterans Affairs · 2025

AI-assisted colonoscopy for earlier cancer detection

21% increase in the odds of adenoma detection

Challenge

Colorectal cancer screening depends on detecting adenomas (precancerous polyps) during colonoscopy; adenomas missed by clinicians are associated with later-stage cancers and higher mortality.

Solution

The Veterans Health Administration deployed FDA-approved devices that use computer vision to assist clinicians in real time during colonoscopy, enhancing detection of potential adenomas.

Results

A VA study found the AI-assisted devices produced a statistically significant 21% increase in the odds of adenoma detection; higher detection rates are associated with lower late-stage cancer incidence and reduced mortality.

A VA study demonstrated that the provision of colonoscopy AI devices resulted in a statistically significant 21% increase in the odds of adenoma detection.

• Computer Vision
• Deep Learning
• Medical Imaging AI

Source: Federal AI Use Case Inventory — U.S. Dept. of Veterans Affairs ↗

U.S. Patent and Trademark Office · 2025

DesignVision: AI image search for design-patent examination

Federated image-similarity search across 80+ global design registers

Challenge

Design-patent examiners must assess novelty by searching enormous collections of prior designs spread across dozens of separate U.S. and foreign registers — a slow, fragmented manual task.

Solution

USPTO launched DesignVision, an AI-based image-search tool in the Patents End-to-End (PE2E) search suite that lets examiners query design collections by image and returns results ranked and sortable by visual similarity.

Results

DesignVision gives examiners centralized, federated image search across design patents, registrations, trademarks, and industrial designs from over 80 global registers, augmenting their existing search tools.

DesignVision will augment — not replace — design examiners' other search tools.

• Computer Vision
• Image Similarity Search

Source: U.S. Patent and Trademark Office ↗

Worten · 2025

Worten saves 11,000 hours a year with an Azure OpenAI store assistant

11,000 hours saved annually on in-store information search

Challenge

Store employees manually searched an operations manual of over 1,000 documents, averaging about 3.5 minutes per query across roughly 75,000 searches a year, around 11,000 wasted hours.

Solution

Worten, with EY, built a smartphone chatbot on Azure OpenAI and Azure AI Search that returns instant answers and direct links to the operations manual.

Results

Documentation search dropped from minutes to seconds, saving an estimated 11,000 hours annually across its Portuguese store network.

Searching for documentation became a much simpler process, from minutes to seconds.

• Azure OpenAI
• Azure AI Search

Source: Microsoft Customer Stories · Microsoft ↗

Aberdeen City Council · 2024

Aberdeen City Council frees staff capacity with Microsoft 365 Copilot

Projected 241% ROI and an estimated $3 million (USD) saved annually

Challenge

Rising demand for healthcare, mental-health, and housing services combined with budget cuts overloaded staff with repetitive administrative work, limiting their ability to serve 230,000 residents.

Solution

The council deployed Microsoft 365 Copilot to 700 staff across city services to automate meeting minutes, reports, and policy documents, and piloted Dynamics 365 contact-center capabilities.

Results

The council projects a 241% ROI in time savings and productivity, an estimated $3 million (USD) saved annually, plus higher job satisfaction and faster resident services.

We project a 241% ROI in time savings and improved productivity, saving an estimated $3 million in US dollars annually.

• Microsoft 365 Copilot
• Dynamics 365

Source: Microsoft Customer Stories · Microsoft ↗

Australian Government — The Treasury · 2024

Whole-of-government trial of Microsoft 365 Copilot in the Treasury

Almost two-thirds of users found Copilot beneficial for administrative tasks

Challenge

The Treasury wanted to understand whether generative AI could be safely deployed to improve staff productivity, efficiency, and employee outcomes.

Solution

As part of the Digital Transformation Agency's whole-of-government trial, Treasury staff trialled Microsoft 365 Copilot, and its impact was independently evaluated across administrative and analytical tasks.

Results

Almost two-thirds of users found Copilot appropriate and beneficial for basic administrative tasks, and the evaluation concluded generative AI can be deployed in the Treasury, with some limitations.

Generative AI can be deployed within the Treasury environment, but there are some limitations.

• Generative AI
• Microsoft 365 Copilot
• Large Language Models

Source: Australian Treasury — Evaluation Unit · Microsoft ↗

Bynder · 2024

Bynder reduces digital asset search time by 75% using Amazon Bedrock

75% reduction in asset search time

Challenge

Marketing teams using Bynder's digital asset management platform spent excessive time manually searching large libraries to find the right assets for campaigns.

Solution

Bynder built visual-similarity and natural-language search using Amazon Bedrock with Amazon Titan Multimodal Embeddings, converting images into vectors for similarity search.

Results

A Bynder customer reported asset-search time for a typical campaign task fell by 75%, with roughly 50% more relevant options returned per search.

Amazon Bedrock with Titan Multimodal Embeddings solved the puzzle for us. We could very easily turn this into a scalable solution.

• Amazon Bedrock
• Amazon Titan Multimodal Embeddings

Source: AWS Customer Stories · Amazon Web Services ↗

EXL · 2024

EXL cuts insurance underwriting costs by up to 80% with Amazon Bedrock

Underwriting costs cut by up to 80%

Challenge

Insurance underwriting was slow and costly, requiring analysts to manually review lengthy documents over several days while staying compliant with PII regulations.

Solution

EXL built LDS Underwriting Assist, a retrieval-augmented-generation assistant on Amazon Bedrock using Anthropic Claude 3 Sonnet, with Amazon Textract, Amazon Comprehend (PII redaction), and Amazon Kendra.

Results

The solution was built in roughly 60 days and cuts underwriting costs by up to 80%, reducing processing time from several days to a few hours.

We've created a generative AI solution with AWS that significantly increases the efficiency of insurance underwriting while lowering the associated costs.

• Amazon Bedrock
• Anthropic Claude 3 Sonnet
• Amazon Kendra
• Amazon Textract
• Amazon Comprehend

Source: AWS Customer Stories · Amazon Web Services ↗

Federal Bank · 2024

Humanizing the banking experience with Vertex AI and Gemini

First bank in India to launch a complete consumer generative AI solution

Challenge

Federal Bank wanted to humanize and modernize customer interactions, upgrading its virtual assistant 'Feddy' to deliver seamless, personalized, multilingual, 24/7 support rather than serving customers static links.

Solution

Starting in 2022 with Dialogflow CX and Vertex AI, the bank added generative AI after Gemini's release to power human-like chatbot and website-search responses, plus a support portal and a Firebase-based employee app.

Results

Federal Bank became the first bank in India to launch a complete consumer-facing generative AI solution, delivering 24/7 multilingual support and fully automated processes.

We are the first bank in India to launch a complete generative AI solution that is available to consumers, and it has been transformational.

• Vertex AI
• Gemini
• Dialogflow CX
• BigQuery
• Firebase

Source: Google Cloud Customer Stories · Google Cloud ↗

Finastra · 2024

Copilot accelerates financial-services marketing at Finastra

One of its largest marketing campaigns took 75% less time with Copilot

Challenge

Building a major marketing campaign took up to three-to-six months, with content production and manual data analysis creating bottlenecks for the financial-software vendor's marketing and communications teams.

Solution

Finastra deployed Microsoft Copilot for Microsoft 365 (with Viva Engage and Power BI) across marketing, communications, and campaign functions to automate content creation, summarize meetings, and analyze campaign performance.

Results

Campaign timelines dropped from three months to under one, one of its largest campaigns took 75% less time, and staff report saving 20-50% of their time.

Copilot has saved me between 20% and 50% of my time.

• Microsoft 365 Copilot
• Microsoft Viva Engage
• Power BI

Source: Microsoft Customer Stories · Microsoft ↗

Pfizer · 2024

Pfizer accelerates life-sciences R&D with generative AI on Amazon Bedrock

Up to 16,000 scientist-hours saved annually; 55% lower infrastructure costs

Challenge

Pfizer scientists spent large amounts of time searching scientific and manufacturing documents, and needed to accelerate innovation while controlling infrastructure costs.

Solution

Through the Pfizer-Amazon Collaboration Team (PACT), Pfizer used generative AI on Amazon Bedrock (Anthropic Claude 2.1) via an internal platform called Vox, plus Amazon SageMaker and Amazon Kendra.

Results

PACT projects saved scientists up to 16,000 hours of search time annually and cut infrastructure costs by 55%, with prototype development falling from 3+ months to 6 weeks.

With access to the talents and technologies of AWS, we've changed our innovation culture and done a lot in a very short time.

• Amazon Bedrock
• Anthropic Claude 2.1
• Amazon SageMaker
• Amazon Kendra

Source: AWS Customer Stories · Amazon Web Services ↗

Sandvik · 2024

Sandvik's Manufacturing Copilot unlocks decades of product knowledge

Manufacturing Copilot saves employees on average 20% to 30% of their time

Challenge

A growing manufacturing skills gap, rising component complexity, and a younger workforce made it hard for employees to access decades of technical product documentation.

Solution

Sandvik built its proprietary Manufacturing Copilot on Azure OpenAI Service and Azure AI Search, giving employees and customers natural-language access to product documentation across 19 languages.

Results

Manufacturing Copilot saves employees on average 20-30% of their time (up to 50% for novel problems) and cut new-salesperson onboarding time by 50%.

Manufacturing Copilot saves time for our employees, on average, close to 20% to 30%.

• Azure OpenAI Service
• Azure AI Search
• GitHub Copilot

Source: Microsoft Customer Stories · Microsoft ↗

Telkomsel · 2024

Powering a super app's digital experience with AI-powered search

88% more search clicks and 20% more purchases within three months

Challenge

Indonesia's largest mobile operator needed to unify search across the content and services of its MyTelkomsel super app, replacing rigid keyword matching that hid relevant results and required constant manual keyword-database updates.

Solution

Telkomsel integrated Gemini and Vertex AI Search (with BigQuery and Datastore) into MyTelkomsel to power personalized, multimodal search, a virtual assistant, and the Digital Smart Care support chatbot.

Results

Within three months, MyTelkomsel saw an 88% rise in search clicks, a 20% increase in purchases, a 355% jump in search engagement, more than 3x revenue from search interaction, and a 27% reduction in customer complaints.

It has led to a 355% increase in search engagement and a more than three times revenue boost from search interaction.

• Gemini
• Vertex AI Search
• BigQuery
• Datastore

Source: Google Cloud Customer Stories · Google Cloud ↗

U.S. Census Bureau · 2024

BEACON: machine learning for industry (NAICS) classification

Challenge

Economic Census respondents must select the correct North American Industry Classification System (NAICS) code for their establishment — a difficult task that affects data quality across the Economic Census.

Solution

The Census Bureau developed BEACON, a machine-learning tool that analyzes respondent-provided text in real time and uses a hierarchical ensemble method to predict the code first at the 2-digit then the 6-digit level, trained on past Economic Census responses, NAICS descriptions, and IRS data.

Results

BEACON lets respondents self-classify their industry in real time during the Economic Census, improving the accuracy and consistency of assigned NAICS codes.

BEACON uses the respondent-provided text, in real time, to predict the respondent's most likely NAICS code.

• Machine Learning
• Ensemble Modeling
• Text Classification

Source: U.S. Census Bureau ↗

U.S. Social Security Administration · 2024

AI to support disability-benefits claim processing

Challenge

SSA faces a growing backlog of disability-benefits claims and long processing times, straining timely and accurate decisions for Americans with disabilities.

Solution

With Technology Modernization Fund support, SSA is applying AI and machine learning — enhancing the National Case Processing System and its Intelligent Medical-Language Analysis Generation (IMAGEN) tool — to flag decision-rationale issues, surface occupational information, and provide real-time decision support to adjudicators.

Results

The effort is intended to reduce processing times, improve determination accuracy, and deliver real-time feedback and decision support that could save adjudicators thousands of work hours annually.

Through the responsible and ethical use of artificial intelligence, SSA plans to tackle a growing backlog of cases.

• Machine Learning
• Natural Language Processing
• Clinical Decision Support

Source: Technology Modernization Fund (GSA) ↗

UK Government Digital Service · 2024

GOV.UK Chat: a generative AI assistant for government guidance

Nearly 70% of users found the responses useful

Challenge

GOV.UK holds information across hundreds of thousands of pages, and GDS wanted to test whether people could find what they need by asking questions in natural language rather than searching and browsing.

Solution

GDS built GOV.UK Chat, an experimental chatbot using retrieval-augmented generation (RAG) over GOV.UK content with OpenAI's GPT-3.5-turbo model, and tested it across five phases including a 1,000-user private pilot.

Results

In the first experiment, nearly 70% of users found the chatbot's responses useful and just under 65% were satisfied, giving GDS confidence to scale testing.

We wanted to see if we could use this approach to enable users to find the information they need by asking questions.

• Generative AI
• Large Language Models
• Retrieval-Augmented Generation
• OpenAI GPT-3.5-turbo

Source: GOV.UK — Government Digital Service · OpenAI ↗

GovTech Singapore · 2023

Pair: an LLM-based AI assistant for public officers

Over 11,000 users across 100+ agencies within the first two months

Challenge

Public officers spend significant time on routine tasks such as writing emails, conducting research, and generating ideas.

Solution

GovTech built Pair, a government AI chatbot for public officers powered by large language models contextualised for the Singapore government, available on government-issued devices with Custom Assistants for tasks like summarising.

Results

Within the first two months of its trial launch, Pair reached over 11,000 users across more than 100 government agencies, and today has 4,500-plus weekly active users.

Pair is a government AI chatbot assistant for public officers.

• Generative AI
• Large Language Models
• Conversational AI

Source: Government Technology Agency of Singapore ↗

NASA · 2023

Prithvi: open-source geospatial AI foundation model for Earth observation

State-of-the-art performance on flood mapping (per NASA)

Challenge

NASA generates vast volumes of Earth-observation imagery, and analyzing it for challenges like floods, wildfires, and land-use change is slow and labor-intensive without scalable AI tools.

Solution

NASA's IMPACT team and IBM Research openly released Prithvi, a geospatial AI foundation model built on NASA's Harmonized Landsat and Sentinel-2 data and trained on IBM's supercomputer using the watsonx foundation-model stack.

Results

The open-source model can be fine-tuned for flood mapping (state-of-the-art performance in NASA's testing), burn-scar identification, and land-cover and crop-type mapping, and is released openly to the science community.

AI foundation models for Earth observations present enormous potential to address intricate scientific problems and expedite deployment.

• Geospatial Foundation Model
• Transformer
• Deep Learning
• watsonx

Source: NASA Earthdata · IBM ↗

TS Imagine · 2023

TS Imagine automates email and ticket triage at scale with Snowflake Cortex AI

4,000 hours saved annually and ~30% lower AI costs

Challenge

The trading and risk-management SaaS firm manually monitored over 100,000 emails and tens of thousands of annual support and data tickets — an error-prone, time-consuming process.

Solution

TS Imagine built a RAG-based email-intake and ticket-classification system on Snowflake Cortex AI, using Snowflake's Arctic LLM alongside Mistral and Llama, with a Streamlit-in-Snowflake interface.

Results

Saved roughly 4,000 hours of manual effort per year and cut AI costs about 30% versus external LLM APIs, with new RAG use cases shipped in about four business days.

Doing everything exclusively in Snowflake was game-changing. Now we design something on a Thursday, and by Tuesday it's in production.

• Snowflake Cortex AI
• Snowflake Arctic LLM
• Mistral
• Llama
• Streamlit in Snowflake

Source: Snowflake Customer Stories · Snowflake ↗

Estonian Information System Authority · 2022

Bürokratt: national AI virtual assistant for public services

Challenge

Citizens need a simple, always-available way to find information and access public services across many Estonian government institutions.

Solution

RIA operates Bürokratt, a state-developed virtual assistant that understands everyday Estonian, answers citizens' queries, and hands them to a customer service representative when it cannot help.

Results

Bürokratt helps citizens find information about public services around the clock and is used by numerous Estonian public-sector institutions listed on the RIA page.

Bürokratt is a virtual assistant designed for citizens, which understands everyday Estonian and is available around the clock if needed.

• Conversational AI
• Virtual Assistant
• Natural Language Processing

Source: Estonian Information System Authority (RIA) ↗

European Commission — DG Translation · 2020

eTranslation: machine translation for EU public administrations

Challenge

Public administrations and businesses across the EU need to exchange information and documents despite operating in many different languages.

Solution

The European Commission provides eTranslation, a secure neural machine-translation tool built on decades of EU translators' work, covering all official EU languages plus Icelandic and Norwegian.

Results

eTranslation is available to EU institutions, public administrations, SMEs, academia, and NGOs to translate documents and text across all official EU languages, Icelandic, and Norwegian.

A machine translation tool that helps public administrations and businesses exchange information and documents across all official EU languages, Icelandic and Norwegian.

• Neural Machine Translation
• Natural Language Processing

Source: European Commission — Interoperable Europe Portal ↗