Turn an issue into a tested patch
Coding agents inspect repositories, edit files, run tests and return evidence. The useful deliverable is not plausible code; it is code proven to work.
Beyond the blank chat box
What can modern AI actually do?
Every entry links to public evidence and labels vendor reports. Find a relevant starting point by work, output or control.
120source-linked use casesPUBLIC EXAMPLES + CAVEATS
12work + industry categoriesSOFTWARE, BANKING + SCIENCE
7source collectionsINDEPENDENT + VENDOR EVIDENCE
Search and source collection combine with the category filter. Try a task, deliverable, risk or tool.
Reproduce and repair real repository bugs
SWE-bench gives agents real GitHub issues and grades the resulting patches with tests. It is a practical model for bounded maintenance work.
Build from a durable specification
OpenSpec and GitHub Spec Kit make the specification a reviewable artifact before generation, reducing ambiguity and keeping the workflow transferable between models.
OpenSpec → · Spec Kit →Continue a project across agent sessions
A harness can preserve progress notes, git history and test state so a later context window resumes work instead of rebuilding an understanding from scratch.
Vendor implementation report →Plan a migration before touching the code
A specification workflow can inventory affected behavior, propose milestones and preserve non-goals before an agent starts a cross-cutting upgrade.
Spec Kit → · OpenSpec →Turn a reported bug into a regression test
Before accepting a fix, require a test that fails on the original behavior and passes after the patch. This is a practitioner verification rule, not a finding from the benchmark.
Practitioner method →Produce a cited research brief
Deep-research agents browse many sources, synthesize them and return citations. The researcher still checks provenance, coverage and whether each citation supports the claim.
Vendor capability report →Recover primary sources from weak clues
LLMs can act as secondary librarians: turn a remembered quote, chart or claim into candidate sources, while the human verifies the original material.
Practitioner examples →Read and compare a document collection
AnythingLLM provides a public implementation of retrieval over local documents, letting users ask cross-document questions while retaining the original files as evidence.
Extract structure from PDFs and reports
Docling converts complex PDFs, tables and page layouts into structured representations that downstream retrieval or analysis workflows can actually use.
Project documentation →Turn a draft into a claim-checking queue
A proposed workflow can extract consequential claims, locate candidate primary sources and flag partial support. The linked system documents source-finding and citations, not this exact review queue.
Underlying vendor capability →Compare revisions across a policy set
A proposed workflow can load named document versions into a bounded collection, ask for candidate differences and verify each one against the originals. The linked project supplies retrieval, not a purpose-built policy-diff guarantee.
Retrieval implementation →Run long, bounded workflows
Agent harnesses can retain task state, resume after context limits, run checks and stop on explicit completion conditions instead of relying on a single conversation.
Vendor implementation report →Operate software that has no API
Computer-use agents inspect screenshots, click and type in existing interfaces. Use narrow accounts, allowlisted sites and approval gates because visual automation is inherently brittle.
Vendor capability release →Transcribe and search meetings
Whisper turns recordings into searchable transcripts. A follow-on model can extract decisions, owners and unanswered questions while the recording remains the canonical source.
Transcription implementation →Route, classify and draft from an inbox
A proposed workflow can route inbound messages, retrieve policy and prepare a response behind human approval. The linked guide documents routing and tool-use patterns, not this exact inbox deployment.
Vendor engineering pattern →Run repeatable agent loops with state and review
A practitioner pattern combines scheduled runs, a state file, isolated worktrees, tools, a separate verifier and human review. It is a method proposal, not evidence of a measured reliability gain.
Extract invoice lines and table fields for review
Parse pages and tables into a fixed schema, validate totals with deterministic code and send low-confidence fields to a person before anything reaches accounting.
Document parsing implementation →Prepare a weekly operations brief
A proposed workflow can gather allowlisted metrics, summarize exceptions and draft next actions while deterministic queries remain the source of every number. The source documents the component patterns, not this specific use case.
Vendor engineering patterns →Prioritize suspicious payments for human investigation
AI can score transactions, surface the signals behind an alert and assemble an investigation packet. Analysts must retain disposition and reporting decisions, with false positives, missed cases and drift measured against reviewed outcomes. A Bank of England and FCA survey found fraud and money laundering among financial firms' established AI uses; it does not prove that every model improves detection.
Bank of England and FCA survey →Review customer and counterparty risk in cross-border payments
A bounded KYC and AML workflow can retrieve approved customer records, support counterparty due diligence and flag unusual payment patterns. For EU or EEA data, document purpose and legal basis, minimize fields and retention, restrict access and assess whether profiling requires a data protection impact assessment. The BIS identifies these as important AI applications in correspondent banking, but sanctions screening, escalation and account decisions still need governed sources, audit trails and accountable reviewers.
Support credit underwriting without black-box denials
Models can organize application evidence and estimate risk, but the lender must own the decision. Keep reason codes tied to factors actually used, test for unfair outcomes, provide meaningful human intervention and let applicants contest a significant decision. EU or EEA deployments need an Article 22 automated-decision assessment and often an Article 35 data protection impact assessment. In the United States, the CFPB says creditors using AI still must give applicants accurate, specific principal reasons for adverse action.
Assist customer-support agents
A generative assistant can retrieve answers and suggest responses while the support worker remains responsible. A 5,179-worker field study measured higher issues resolved per hour.
Field study →Answer questions and call a quote tool
A support agent can ground answers in product documentation and call a deterministic quote generator instead of inventing prices in free-form text.
Vendor implementation guide →Prepare for a sales meeting
Anthropic reports that ServiceNow combined enterprise context and web research for sales preparation and saw up to a 95% reduction in preparation time in early testing. The page provides no independent audit or study methodology.
Vendor/partner report →Draft routine business writing
First drafts of emails, press releases, reports and short analyses are a measured use case: a randomized experiment found faster completion and higher evaluator-scored quality.
Peer-reviewed experiment →Assist with inventory, without owning the money
Project Vend's agent managed pricing, stock decisions and customer requests but lost money through concrete errors. It supports testing recommendations before granting spending authority.
Vendor research report →Qualify inbound requests before a person replies
A proposed routing workflow can classify requests by explicit criteria, retrieve the relevant offer and prepare follow-up questions without making commitments. The source documents the routing pattern, not a measured sales result.
Vendor engineering pattern →Turn a storyboard into short video shots
Runway documents a shot-by-shot workflow for longer videos: plan the narrative, create character references, generate short clips and assemble them in an editor. Treat each generated shot as footage to review, not a finished film.
Runway production guide →Animate a still image into a controlled clip
An image-to-video workflow keeps the source image responsible for subject, composition, lighting and style, while the prompt specifies motion and camera behavior. Current Runway outputs are short, so budget for iteration and editing.
Runway prompting guide →Remove a distraction from a photo
Lightroom's Generative Remove lets an editor brush over an unwanted object, inspect several generated fills and regenerate when the result is weak. Keep the original file and check edges, reflections and repeated textures at full size.
Adobe Lightroom guide →Restage a product photo for a new campaign
Shopify's media editor can remove or generate backgrounds and change lighting from a prompt while retaining the product image. Verify that the result still represents the item, and preserve AI-source metadata when the image will enter Google Merchant Center.
Adapt one idea for several social channels
Buffer's AI Assistant can draft, rephrase, shorten or expand captions with channel-aware limits. Give it the source material, audience, objective and constraints, then fact-check and approve every post before scheduling.
Buffer workflow guide →Turn post performance into the next content brief
Buffer Insights summarizes patterns across recent posts and can open an AI-assisted draft based on what performed well. Treat the suggestion as a hypothesis: compare like formats and audiences, and do not infer causation from engagement alone.
Buffer Insights guide →Classify pages while crawling a website
Screaming Frog can send page text, HTML or extracted fields to a hosted or local model during a crawl. Use bounded prompts to classify intent, language or missing image descriptions, test on a few URLs first and review the resulting queue before changing the site.
SEO Spider tutorial →Research and structure a people-first page
Google says generative AI can help research a topic and add structure to original content. The acceptance test is still accuracy, relevance and distinct reader value; mass-producing pages without added value can violate its scaled-content-abuse policy.
Turn a concept or lyrics into a complete song sketch
Suno can turn a plain-language description or supplied lyrics into a song, making it useful for demos and rapid exploration. Treat release rights separately: its paid-tier terms assign only the rights Suno owns and warn that copyright may not vest, while free-tier output is non-commercial.
Develop a hummed tune or recorded idea
Record or upload a short audio idea, describe the intended result and generate continuations to audition. Use only material you have permission to upload, and keep the original recording as the reference for deciding what the model actually preserved.
Suno audio workflow →Compose an instrumental, then edit the MIDI
AIVA generates music across more than 250 styles, accepts audio or MIDI influences and exports editable formats. Its free plan is non-commercial and AIVA-owned; the Pro plan advertises user-owned copyright, so select the licence before composing for a client or release.
AIVA features and plans →Replace a weak section without restarting the song
Udio Sessions lets paid users extend their own tracks, replace a selected waveform region, revise aligned lyrics, compare takes and undo edits. It turns generation into an iterative timeline workflow rather than a one-prompt lottery.
Udio Sessions guide →Generate a full song locally with MiniMax Music 3
MiniMax Music 3 accepts lyrics plus a detailed music description and produces up to five minutes of stereo audio. The official reference path fits full precision in under 24 GB of VRAM and documents offloading down to 8 GB, but the complete repository is about 57 GB and its community licence adds attribution, revenue and public-disclosure conditions.
Generate accompaniment, stems or revisions locally
ACE-Step 1.5 supports text and reference-audio generation, covers, selective repainting, track separation and vocal-to-accompaniment work. Its project reports support for CUDA, Apple, AMD, Intel and CPU paths, with the base model able to run below 4 GB of VRAM; verify the exact model variant and input rights before publishing.
ACE-Step 1.5 repository →Prototype short instrumental samples offline
MusicGPT packages Meta MusicGen behind a local desktop UI and command line on macOS, Linux and Windows. It generates 10-30 second samples and keeps prompts and audio on the host, but its downloaded model weights use a non-commercial licence, so this is an experimentation route rather than a commercial production default.
MusicGPT repository →Run a persistent personal assistant
Hermes maintains memories, searches previous conversations and creates reusable skills. A phone can be the messaging surface while the agent and model stay on another machine; the chosen messaging provider remains part of the conversation path.
Search a personal knowledge base
Store notes as local Markdown, then index a copy in a retrieval tool. Obsidian supplies the file-based notes; AnythingLLM supplies the AI retrieval layer. Keep the notes, not generated summaries, canonical.
Reach a self-hosted agent from messaging
OpenClaw can keep sessions and state on an always-on host while an iPhone or Android device acts as the operator or a narrowly scoped node. Keep the gateway private, give it narrow permissions and require approval for sending, deleting, purchasing or changing accounts.
Transcribe and search private voice notes
Run speech recognition locally, attach timestamps and retain the audio as the source so later summaries and task extraction remain checkable.
Whisper implementation →Coach the tutor during a lesson
Tutor CoPilot suggests questions and teaching moves to a human tutor in real time. The preprint reports a 4-percentage-point mastery gain overall and 9 points for students of lower-rated tutors.
Randomized-trial preprint →Give learners an always-available speaking practice partner
Duolingo's Video Call uses an animated AI character for real-time speaking practice, followed by a transcript and feedback. In a company-run 30-day study with 567 completers, the Video Call group showed 43.7% greater speaking-score improvement than the control group; the report tests the complete feature, not the avatar alone.
Duolingo research report →Prepare replies to patient messages
AI can prepare drafts for clinician review. A five-week single-group study found 20% adoption and no measured time savings; lower surveyed burden and exhaustion were observational, so the safe claim is feasibility, not proven efficiency.
Single-group QI study →Predict molecular structures and interactions
The peer-reviewed AlphaFold 3 system predicts joint structures containing proteins, nucleic acids, small molecules, ions and modified residues. Predictions are hypotheses to test, not experimental confirmation.
Peer-reviewed system paper →Build and revise a playable 3D scene through an agent
Unity's beta assistant can inspect scenes and components, drive editor actions and verify changes. The community Godot AI plugin exposes scene, node, script, UI, material and test operations through MCP. Both routes still need human playtesting, performance checks and asset-rights review.
Debug and profile an XR app from the editor
Meta's VR command-line and MCP tooling can query platform documentation, inspect device logs and screenshots, analyze Perfetto traces and GPU frame metrics, find 3D assets and check store readiness. This is first-party capability documentation, not measured evidence of time saved.
Meta Horizon AI tooling overview →Give game characters perception, decisions and navigation
Unreal Engine combines perception signals, Blackboards, Behavior Trees and navigation meshes so non-player characters can sense events, select branches and move through a level. This is game-runtime control logic, not a claim that a language model creates believable behavior automatically.
Create and revise a parametric CAD model from intent
Autodesk's Fusion MCP lets an AI tool act inside the design environment for repetitive modeling steps, multi-step changes and workflow automation. A community beta exposes concrete sketching, constraints and feature tools, but its optional network socket has no authentication and must stay off untrusted networks.
Turn building drawings into an energy model
SimulationHub says Buildings AI can detect walls, windows and doors in drawings, convert them into BIM geometry, guide material and schedule setup, then run EnergyPlus-based load, energy, airflow and thermal-comfort analysis. Treat automated geometry and defaults as inputs to validate, not a compliant result by themselves.
Buildings AI capability documentation →Generate and compare early building layouts
Autodesk's experimental Building Layout Explorer generates floor-plan options from a massing model, building type and structural material so teams can compare alternatives before detailed design. Availability is limited, and generated layouts still require architectural, structural, accessibility and code review.
Autodesk experimental feature note →Predict a new design's physics from past simulations
Ansys SimAI trains surrogate models on previous 3D simulation data and predicts physical fields for new geometries across fluids, structures, electromagnetics and optics. Use its confidence and evaluation reports to identify out-of-distribution designs, then confirm decisions with the original solver and physical tests.
Ansys SimAI product documentation →Test factory layouts and robot fleets before deployment
Physics-based digital twins can connect 3D facility data with sensors and simulation to compare layouts, material flow and robot-fleet behavior before changing a real site. NVIDIA's route is vendor architecture guidance; teams still need calibrated models, real operating data and field acceptance tests.
NVIDIA digital-twin workflow →Extract assets and change from maps and imagery
GeoAI models can detect and classify buildings, roads, damage and other features in aerial imagery, video and point clouds, then feed spatial analysis or a digital twin. Validate against representative regions and dates because sensor conditions, geography and label quality can shift accuracy.
Esri GeoAI overview →Optimize chip design flows against engineering goals
Cadence Cerebrus applies machine learning and reinforcement learning across digital implementation so engineers can explore power, performance and area trade-offs for multiple blocks. Published gains are vendor and customer reports; sign-off tools and silicon results remain the acceptance gates.
59 vendor-reported deployments
From prototypes to work at scale.
These additional examples come from Microsoft, AWS, Google Cloud, OpenAI, GitLab and Anthropic customer stories. They show concrete workflows and reported outcomes, but remain customer/vendor case studies, not independent experiments or proof that a result transfers unchanged.
Microsoft · 28 stories
Azure, Foundry and Copilot deployments
Customer-reported implementations published by Microsoft.
Surface answers during customer calls
Microsoft reports that AT&T's digital coworkers reduced information-search time for customer-care staff by 33%. The wider platform had 71 generative-AI solutions in use by more than 100,000 employees.
Microsoft customer story →Question test-vehicle telemetry in natural language
BMW's multi-agent system retrieves, analyzes and visualizes test-fleet data for engineers. The story labels its early internal result as up to 12 times faster analysis potential, not a completed independent evaluation.
Microsoft customer story →Summarize board materials before meetings
Nasdaq says Boardvantage's AI summarization saved governance teams more than 100 hours annually, with directors reporting up to 25% less preparation time and up to 60% less reading time.
Microsoft customer story →Analyze full audit datasets and draft documentation
KPMG Clara uses agents to analyze datasets, prepare documentation and surface risk signals. The customer story emphasizes deployment scale, 95,000 auditors across member firms in more than 140 countries, rather than a controlled productivity effect.
Microsoft customer story →Turn campaign history into targeting proposals
HicMobile converted unstructured campaign files into a searchable knowledge base and targeting tool. Its story reports a six-month prototype-to-launch cycle, described as a 70% faster timeline, and a prototype delivered 11.25 times faster than working without the co-innovation lab.
Microsoft customer story →Give a governed assistant to a global workforce
Aon built AonGPT to connect internal knowledge and automate routine analysis. Microsoft reports more than 62,000 users, about 31,000 monthly active users and over 6.4 million messages exchanged.
Microsoft customer story →Replan production when factory conditions change
Sight Machine used AI-generated optimization models for production scheduling. Microsoft reports that one beverage manufacturer cut non-value-added production time by 75% and increased production capacity by more than 5%.
Microsoft customer story →Resolve routine banking requests around the clock
Commerzbank's Ava agent handles more than 30,000 customer conversations a month. The bank reports that roughly 75% of requests in the designed scope are resolved autonomously.
Microsoft customer story →Complete financial tasks through conversation
Astra Tech embedded a multilingual voice assistant in botim for tasks such as money transfer. The story reports 3.6 million users and 375% wallet-transaction growth, while attributing that growth to broader product initiatives with the assistant as one driver.
Microsoft customer story →Personalize sports stories for millions of fans
The Premier League Companion combines live and historical data into individualized feeds. The league reports about 20% year-over-year growth in app and website consumption and 60 million active fans early in the season.
Microsoft customer story →Orchestrate routines in a consumer assistant
SK Telecom expanded A.(A-Dot) from single-turn responses into multi-step personal workflows. Microsoft reports growth from about 1.1 million monthly active users to more than 10 million subscribers and monthly active users by 2025, plus three to four months trimmed from feature delivery.
Microsoft customer story →Route customer and employee banking questions
Banco Bradesco built a governed multi-agent platform for internal and external processes. The bank reports an 83% resolution rate in digital customer service and 80% for employee queries.
Microsoft customer story →Answer routine patient calls and return voicemails
healow is piloting Genie as an AI medical receptionist for appointment information, common questions and voicemail handling. The story describes expected workload and satisfaction benefits, but no controlled outcome measurement.
Microsoft customer story →Draft clinical notes from patient conversations
healow's Sunoh.ai transcribes visits and prepares a clinical note for provider approval. Microsoft reports clinicians saving up to two hours per day; that is a customer-reported result, not an independent trial.
Microsoft customer story →Prepare emergency charts and discharge instructions
Sayvant transcribes acute-care conversations and drafts charts plus instructions in more than 30 languages. The company reports cutting charting from 10 minutes to under 90 seconds per patient and saving an estimated 50,000 clinician hours.
Microsoft customer story →Draft appeal determination letters for nurses
Acentra Health's MedScribe prepares letters for nurse review. The company reports about 50% less time per letter, 11,000 nursing hours and nearly $800,000 saved, with a 99% approval rate for generated drafts.
Microsoft customer story →Provide a curriculum-grounded tutor at distance-learning scale
Universitas Terbuka deployed an AI tutor across 500 classes and roughly 100,000 students. Its internal research across four courses and nearly 38,000 students found more discussion participation and higher assignment scores, reported as statistically significant.
Microsoft customer story →Prepare teacher assessment from photographed work
A Discovery Trust teacher used Copilot to assess 33 pieces of work against Year 6 standards in 30 minutes instead of six hours. The trust estimates 4,500 hours reclaimed annually across its first 59 licenses.
Microsoft customer story →Give teachers feedback across millions of essays
São Paulo's education department uses AI to analyze and draft personalized feedback on nearly 10 million essays. Microsoft says the system saves educators thousands of hours; teachers remain responsible for instruction and use of the feedback.
Microsoft customer story →Retrieve regulated crop guidance at the edge
Bayer fine-tuned a small language model on crop-protection labels. Early users report 5-10% productivity gains and answers to complex questions in under 30 seconds instead of days; the result remains an early customer estimate.
Microsoft customer story →Search decades of R&D without duplicating experiments
Almirall built a multilingual assistant over 400,000 documents spanning more than 50 years. Scientists reportedly locate past experiments in seconds rather than hours or days, while subject experts validate the results.
Microsoft customer story →Test clinical hypotheses with governed code execution
Novo Nordisk's reasoning agent generates code and statistical analysis over harmonized clinical data, with human validation built in. Microsoft reports time to insight falling from weeks to minutes.
Microsoft customer story →Provide general AI inside a high-security laboratory
Sandia National Laboratories deployed a custom AI chat service to nearly 17,000 employees in eight months. Its customer story reports about three minutes saved per question while keeping the service inside established security controls.
Microsoft customer story →Turn freight-request emails into quotes
C.H. Robinson's workflow classifies inbound freight emails, extracts details, requests missing information and prepares a response. The company reports cutting average quote time from hours to 32 seconds and being on pace for a further 15% productivity increase.
Microsoft customer story →Check documents for accessibility, tone and clarity
Scope's staff use an internal engine to review documents against organizational accessibility standards. The disability charity reports more than 26,000 prompts in a typical month and over 640 task-specific agents, but no controlled productivity estimate.
Microsoft customer story →Generate personalized character interactions at campaign scale
DEPT built an interactive Sinterklaas experience using language and custom-voice models. Its retail client reports 300,000 users, more than 3 million interactions and 233% higher engagement than the previous year's holiday campaign.
Microsoft customer story →Retrieve the current operating procedure in seconds
Rumo gives train drivers authenticated access to a regulatory-team-approved knowledge base. It reports reducing average lookup time from more than four minutes to three seconds, reclaiming 7,644 hours annually and achieving payback in under two months.
Microsoft customer story →Extract trade confirmations and compile ETF reports
CSOP built tools for inconsistent trade-confirmation documents and daily ETF reporting. The firm reports 99% automation in selected workflows, 10-minute tasks completed in 30 seconds and monthly reporting effort reduced by 75%.
Microsoft customer story →AWS · 9 stories
Bedrock and AWS AI deployments
Customer-reported implementations published by Amazon Web Services.
Investigate security alerts with traceable evidence
eSentire's agent forms hypotheses, selects tools, gathers evidence and produces drill-down investigation reports that analysts can extend. AWS says a validation against 1,000 real investigations reached 95% alignment with senior SOC experts and that hours of analysis now arrive in minutes. This vendor case supports triage and investigation, not unsupervised trust in containment decisions.
AWS customer story →Prioritize CVs against a job description
Ubidy uses an LLM pipeline to rank CVs for enterprise recruiting clients, and tests new models against fixed CV/job-description matches with known results. AWS reports a 95% reduction in CV-evaluation time; its story also identifies candidate-suitability AI as a high-risk use. Treat ranking as recruiter support, with documented fairness, privacy and human-decision controls, not automatic rejection.
AWS customer story →Summarize long email threads and stage updates
Epilot built an Amazon Bedrock workflow that summarizes utility-customer email chains and proposes record updates for human confirmation. AWS reports an 87% reduction in email handling time across the evaluated workflow.
AWS customer story →Field high-volume delivery-support calls
DoorDash built a voice-operated support flow using Amazon Bedrock and Claude. The story reports hundreds of thousands of calls handled per day, response latency of 2.5 seconds or less and 50% less application-development time.
AWS customer story →Automate invoice intake and customer onboarding
Ellby used Amazon Bedrock to raise automated invoice processing from under 60% to more than 94%. The company reports saving over 300 maintenance hours per month and cutting onboarding time by more than 55%.
AWS customer story →Search scientific data and literature with citations
Genentech's gRED Research Agent searches internal data and PubMed, then synthesizes cited answers to multi-step questions. AWS says work that took weeks can take minutes and projects more than 43,000 manual hours automated in biomarker validation.
AWS customer story →Add drafting and slide generation to office software
WPS AI adds rewriting, proofreading and presentation generation to WPS Office for more than 200 million overseas users. WPS reports 30% higher R&D efficiency and 35% lower operational costs after its Amazon Bedrock rollout.
AWS customer story →Monitor tax rules and summarize business impacts
Amazon Finance built World Wide Watch to identify, prioritize and summarize VAT-policy changes. Its AWS case study reports more than 90% accuracy and a 92% reduction in time to insight, from 26 minutes to two minutes per update.
AWS self-case study →Retrieve support knowledge while agents work
Fractal Analytics built Knowledge Assist over enterprise content with Amazon Bedrock and semantic search. Its clients report 10-15% shorter average call handling and a 30% deflection rate for supported self-service questions.
AWS customer story →Google Cloud · 9 stories
Gemini and Vertex AI deployments
Customer-reported implementations published by Google Cloud.
Review litigation documents and explain relevance decisions
Altumatim uses Gemini to classify and analyze large eDiscovery collections, with reports that preserve evidence and explanations. Google Cloud says the platform moved some analyses from weeks or months to hours or days and improved document classification from 67% to over 90%. Those are vendor and customer figures: counsel must validate relevance, privilege, confidentiality and sampled accuracy before relying on a review.
Google Cloud customer story →Question data across a mine-to-port operation
Golden Energy Mines built GEMVIS, a multi-agent system that connects data across more than 50 applications. Google Cloud reports data retrieval falling from two days to under an hour and executive decision speed improving by over 90%.
Prepare engineering quality assessments
Cognizant fine-tuned Gemini on its audit knowledge to help assess thousands of projects consistently. The company reports preparation time falling from as much as six hours to one and a functional agent prototype delivered in one week.
Google Cloud customer story →Review advertising content at platform scale
Taboola moved advertising-content review to Gemini while rolling Google AI tools across the company. Google Cloud reports a 75% reduction in moderation costs and AI tools supporting daily workflows for 90% of employees.
Google Cloud customer story →Unify support knowledge and resolve more tickets
Mosaic AI uses Gemini and Vertex AI to connect customer-facing teams with company knowledge. Google Cloud reports over 50% more tickets resolved, response times more than 35% faster and up to 75% more tickets handled overall.
Google Cloud customer story →Automate telecom infrastructure management
IT-Development uses Gemini to automate maintenance workflows for telecom operators and tower companies. Its Google Cloud story reports onboarding reduced from weeks to days and about 30% lower infrastructure-management costs.
Google Cloud customer story →Evaluate AI applications at production scale
Galileo uses Gemini-based evaluation agents to test model behavior and risk. Google Cloud reports more than 1,000 AI applications assessed and over 20 million requests processed daily at roughly 300-millisecond latency.
Google Cloud customer story →Work across documents in one AI workspace
Macro uses Gemini for multi-document chat, editable maps and actions over connected work. Google Cloud reports more than 125,000 users and says 80% interact with the workspace through its AI chat and agentic features.
Google Cloud customer story →Classify customer feedback before it becomes a product decision
Mattel built a feedback-classification system over social posts, reviews and direct communications, using the results to spot product, brand and supply-chain signals. Google Cloud reports analysis falling from a month to a minute and about $1 million in savings. This vendor account is not an accuracy study: inspect labels, samples and representative raw feedback before acting on a trend.
Google Cloud implementation report →OpenAI · 8 stories
ChatGPT and OpenAI deployments
Customer-reported implementations published by OpenAI.
Replace an obsolete test framework across a large codebase
Asana used Codex agents in separate codebase copies to remove Enzyme, with an engineer checking progress twice daily and reviewing every proposed change. OpenAI reports the migration took two calendar weeks and about $12,000 in model and infrastructure costs, against Asana's prior estimate of at least five years and roughly $6 million in staffing. The comparison is a company estimate, not a controlled benchmark.
OpenAI customer story →Build editable finance decks and workbooks with traceable sources
Model ML agents carry a brief through research, calculations and native PowerPoint or Excel output. At one asset manager, OpenAI reports a custom tearsheet fell from about an hour to five minutes. Finance professionals still check assumptions, sources, formulas and message before sharing; Model ML's own benchmark found only 43.3% of GPT-5.6 Sol decks cleared its professional-readiness gate.
OpenAI customer story →Remediate software vulnerabilities under human review
HP teams used ChatGPT and Codex to analyze and remediate critical software vulnerabilities across tools. OpenAI reports several bugs resolved in a day that the security team estimated could otherwise have taken up to a month, plus a directional estimate of about 82 security-team hours freed weekly. The work remains permissioned, evaluated and reviewable; these are pilot estimates, not an independent security study.
OpenAI partnership report →Audit lead-funnel failures and stage fixes at scale
Zapier's marketing team uses ChatGPT Work to review thousands of dropped leads each month, diagnose recurring issues, patch system problems and update an executive dashboard. OpenAI reports that reviewing one lead manually took 35-45 minutes and attributes seven figures of monthly pipeline to the system. That is customer and vendor attribution, so teams should verify classifications, fixes and revenue accounting.
OpenAI customer story →Prepare structured utilization-review rationales
AdventHealth uses ChatGPT for Healthcare to summarize charts, surface relevant details and draft rationales while physician advisors keep final judgment. OpenAI reports an 80% reduction in time spent on the measured administrative workflow.
OpenAI customer story →Turn operational knowledge into a usable first draft
Recycling-equipment maker STADLER uses ChatGPT across drafting, summarization, translation and analysis. It reports 30-40% time savings on common knowledge tasks, 2.5 times faster first drafts and daily use above 85%.
OpenAI customer story →Compare a contract with a company playbook and stage redlines
Ironclad's GPT-4-based AI Assist flags irregularities and suggests playbook-grounded clauses, while users can accept or reject every suggestion. OpenAI reports Ironclad users cutting an initial redlining pass from about 40 minutes to two; this is vendor evidence, not an independent legal-quality study. A qualified reviewer must own interpretation, negotiation and approval.
OpenAI implementation report →Localize content and widen access to data work
Holiday Extras uses ChatGPT for multilingual content, data analysis, code debugging and support. The company reports more than 500 hours saved weekly, with 92% of employees saving over two hours a week.
OpenAI customer story →GitLab · 1 story
GitLab Duo deployment
Customer-reported implementation published by GitLab.
Investigate a failed pipeline and prepare a proposed fix
Barclays uses GitLab Duo to explain failed job logs, identify a likely root cause and suggest a fix inside its developer workflow. GitLab's customer story describes one manager resolving a pipeline issue in seconds; that anecdote is not a controlled delivery-speed measure. Review the diagnosis and change before merge, and keep log and code access within approved controls.
GitLab customer story →Anthropic · 4 stories
Claude deployments
Customer-reported implementations published by Anthropic.
Review every expense and escalate policy exceptions
Brex uses Claude to interpret spend policies, categorize expenses and flag out-of-policy line items, so finance teams can review exceptions rather than every transaction. Anthropic reports 100% of transactions checked and a 94% compliance rate against a stated 70% industry baseline. These customer and vendor figures do not replace policy ownership or human judgment on ambiguous spend.
Anthropic customer story →Accelerate analytics and query handling
IG Group uses Claude across analytics, content and strategic work. Anthropic reports about 70 analyst hours saved each week, full payback in under three months and a days-long executive-analysis exercise completed in under two hours.
Anthropic customer story →Help developers understand and change complex code
Palo Alto Networks put Claude into developer tools through Google Cloud. Anthropic reports 20-30% higher feature-development velocity, onboarding reduced from months to weeks and a 70% faster pilot task for junior developers.
Anthropic customer story →Research and draft client proposals faster
Norwegian communications group TRY uses Claude across research, proposals, content and project work. Anthropic reports 30% less time on routine tasks, 40% faster proposal development and more than 50 use cases in operation.
Anthropic customer story →No use case matches this category, source collection and search. Try a broader choice.
Hosted music app
Hosted Codex Music siteOpen Now That’s What I Call Slop!
↗
When sharing the album becomes the app.
Codex Music presents three Andrew Ambrosino releases as a searchable library with album pages, liked songs, lyrics and a persistent player. It is a concrete example of building the exact distribution surface a piece of work needs.
3 releases
Library + search
Playback + lyrics
chat-only mental model
Ask, receive text, copy it somewhere
- No tools
- No direct inspection
- No memory of the work environment
- No deterministic completion check
agent mental model
Observe, act, test, recover and report
- Reads relevant files and systems
- Uses terminals, APIs and applications
- Maintains task state and reusable skills
- Stops when tests or approval gates say it is done
Turn an example into your workflow
Copy the pattern, not the demo. Define the artifact, boundary and acceptance test.