{"name":"Agents in the Wild","url":"https://agentsinthewild.co","count":35,"items":[{"title":"Octopus Energy's AI email assistant Arlo scored higher satisfaction than human replies","summary":"UK energy supplier Octopus trialled Arlo, an AI that answers routine customer emails. In a three-month trial it handled about 8,000 emails a week and scored 76% satisfaction, versus 72% for comparable human replies.","category":"customer-support","organization":"Octopus Energy","industry":"Energy","tools":["Arlo","Kraken"],"outcome":"76% customer satisfaction for AI replies vs 72% for comparable human replies, on ~8,000 emails a week (4% of UK customer emails).","metrics":[{"label":"CSAT, AI replies","value":"76%"},{"label":"CSAT, human replies","value":"72%"},{"label":"Emails per week","value":"~8,000"},{"label":"Share of UK emails","value":"4%"}],"evidence":"company-reported","caveats":["Arlo only handles routine questions. Vulnerable customers, sensitive cases, and complex complaints go straight to people, so the comparison is not like-for-like across all emails.","Earlier press reports gave much larger figures (about a third of emails, '250 people's worth of work'). Those do not match this later trial data."],"sources":[{"title":"Octopus Energy's AI trial wins customer approval","url":"https://octopus.energy/press/more-news-press-releases/octopus-energy-s-ai-trial-wins-customer-approval/","publisher":"Octopus Energy","date":"2026-07"},{"title":"AI satisfies customers better than human customer support, CEO says","url":"https://thewrap.com/ai-satisfies-customers-better-than-humans","publisher":"TheWrap","date":"2024"}],"happenedAt":"2026-07","addedAt":"2026-10-16","tags":["email support","CSAT","utilities","human handoff"],"sponsored":false,"slug":"octopus-energy-arlo-emails","toolSlugs":["arlo","kraken"],"url":"https://agentsinthewild.co/use-cases/octopus-energy-arlo-emails"},{"title":"Stripe's Minions coding agents produce 1,000+ merged pull requests every week","summary":"Stripe built homegrown coding agents called Minions. An engineer starts a task, often from Slack, and gets back a tested pull request. Over a thousand merged PRs a week contain no human-written code, though humans review them.","category":"software-engineering","organization":"Stripe","industry":"Payments","tools":["Minions (internal)","Goose"],"outcome":"More than a thousand pull requests merged each week are fully agent-produced and human-reviewed.","metrics":[{"label":"Merged agent PRs","value":"1,000+ per week"},{"label":"Human-written code in them","value":"None"}],"evidence":"company-reported","caveats":["Figures come from Stripe's own engineering blog.","Every PR is human-reviewed. Merged PR counts do not show the size or difficulty of each change.","Some secondary sources quote 1,300+ per week. We use Stripe's own wording."],"sources":[{"title":"Minions: Stripe's one-shot, end-to-end coding agents","url":"https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents","publisher":"Stripe","date":"2026-02"},{"title":"How Stripe's Minions ship 1,300 PRs a week","url":"https://blog.bytebytego.com/p/how-stripes-minions-ship-1300-prs","publisher":"ByteByteGo","date":"2026"}],"happenedAt":"2026-02","addedAt":"2026-10-16","tags":["background agent","Slack","one-shot","MCP","developer productivity"],"sponsored":false,"slug":"stripe-minions-coding-agents","toolSlugs":["minions-internal","goose"],"url":"https://agentsinthewild.co/use-cases/stripe-minions-coding-agents"},{"title":"Anthropic's growth marketer cut ad creation from 30 minutes to 30 seconds with Claude Code","summary":"A growth marketer at Anthropic with no coding background used Claude Code to build a Figma plugin for ad creative and a Google Ads copy workflow. Ad production went from about 30 minutes to about 30 seconds per ad.","category":"sales-marketing","organization":"Anthropic","industry":"AI","tools":["Claude Code","Figma","Google Ads"],"outcome":"Work that took about 30 minutes per ad now takes about 30 seconds; the plugin can make up to 100 ad variations at once.","metrics":[{"label":"Time per ad","value":"~30 min → ~30 sec"},{"label":"Variations per run","value":"up to 100"}],"evidence":"vendor-case-study","caveats":["Anthropic makes Claude Code, so this is a vendor story about its own team.","Time saved is self-reported by the marketer. Ad performance results are not given."],"sources":[{"title":"How Anthropic's Growth Marketing team cut ad creation time from 30 minutes to 30 seconds","url":"https://claude.com/blog/how-anthropic-uses-claude-marketing","publisher":"Anthropic","date":"2026-01"},{"title":"How Anthropic teams use Claude Code","url":"https://www.anthropic.com/news/how-anthropic-teams-use-claude-code","publisher":"Anthropic","date":"2025"}],"happenedAt":"2026-01","addedAt":"2026-10-16","tags":["paid ads","creative variations","non-developer","Figma plugin","CSV workflow"],"sponsored":false,"slug":"anthropic-growth-marketing-ads","toolSlugs":["claude-code","figma","google-ads"],"url":"https://agentsinthewild.co/use-cases/anthropic-growth-marketing-ads"},{"title":"Spotify's Honk background agent merges fleet-wide code migrations: 1,500+ PRs and counting","summary":"Spotify connected a background coding agent, later built on Claude Code, to its Fleet Management platform. It opens migration pull requests across many repositories at once. Review is now the bottleneck.","category":"software-engineering","organization":"Spotify","industry":"Media & Streaming","tools":["Claude Code","Fleet Management (internal)"],"outcome":"1,500+ merged AI-generated PRs; 1,000 merged PRs took three months at first and about ten days later on.","metrics":[{"label":"Merged agent PRs","value":"1,500+"},{"label":"Time for 1,000 PRs","value":"3 months → ~10 days"}],"evidence":"company-reported","caveats":["Figures come from Spotify's engineering blog and conference talks.","Spotify says code review, not code generation, became the bottleneck."],"sources":[{"title":"1,500+ PRs later: Spotify's journey with our background coding agent (Honk, Part 1)","url":"https://engineering.atspotify.com/2025/11/spotifys-background-coding-agent-part-1","publisher":"Spotify Engineering","date":"2025-11"},{"title":"Spotify Honk rewrite","url":"https://infoq.com/news/2026/03/spotify-honk-rewrite","publisher":"InfoQ","date":"2026-03"}],"happenedAt":"2025-11","addedAt":"2026-10-16","tags":["background agent","code migration","fleet management","code review"],"sponsored":false,"slug":"spotify-honk-background-agent","toolSlugs":["claude-code","fleet-management-internal"],"url":"https://agentsinthewild.co/use-cases/spotify-honk-background-agent"},{"title":"Amazon says its Rufus shopping assistant is on track for $10B+ in extra yearly sales","summary":"Amazon's AI shopping assistant Rufus answers product questions inside the Amazon app. On the Q3 2025 earnings call, Andy Jassy said 250 million shoppers used it that year.","category":"sales-marketing","organization":"Amazon","industry":"E-commerce","tools":["Rufus"],"outcome":"Amazon expects over $10B in annualized incremental sales; shoppers who used Rufus were 60% more likely to buy.","metrics":[{"label":"Users in 2025","value":"250M"},{"label":"Incremental sales (annualized)","value":"$10B+ (estimate)"},{"label":"Purchase likelihood","value":"+60%"},{"label":"Monthly users, YoY","value":"+140%"}],"evidence":"company-reported","caveats":["Amazon's own estimate. It did not publish how 'incremental' or 'user' is defined.","Reports say sales are credited to Rufus over a 7-day window after use, which can overcount.","Shoppers who choose to use Rufus may already be more likely to buy."],"sources":[{"title":"Amazon's Rufus AI shopping assistant is on track for $10 billion in sales","url":"https://www.fortune.com/2025/11/02/amazon-rufus-ai-shopping-assistant-chatbot-10-billion-sales-monetization","publisher":"Fortune","date":"2025-11"}],"happenedAt":"2025-10","addedAt":"2026-10-16","tags":["shopping assistant","conversion","e-commerce","earnings call"],"sponsored":false,"slug":"amazon-rufus-shopping-assistant","toolSlugs":["rufus"],"url":"https://agentsinthewild.co/use-cases/amazon-rufus-shopping-assistant"},{"title":"BNY runs 100+ AI 'digital employees' with their own logins and human managers","summary":"The bank BNY gives its AI agents a login, an email, an employee number, and a human manager who assigns tasks and reviews output. They work on jobs such as payment instruction checks and code security fixes.","category":"finance-operations","organization":"BNY","industry":"Banking","tools":["Eliza (internal)","OpenAI"],"outcome":"Over 100 digital employees in production by October 2025, later reported as 130+, each with a human supervisor.","metrics":[{"label":"Digital employees","value":"100+ (Oct 2025), 130+ later"}],"evidence":"press-report","caveats":["Headcounts differ between sources and dates.","Sources describe the setup, not measured results such as time or cost saved."],"sources":[{"title":"This Wall Street bank has over 100 'digital employees'","url":"https://www.axios.com/2025/10/17/ai-wall-street-digital-workers","publisher":"Axios","date":"2025-10"},{"title":"BNY built its digital workforce backward — and it's working","url":"https://www.forrester.com/blogs/bny-built-its-digital-workforce-backward-and-its-working/","publisher":"Forrester","date":"2026"},{"title":"BNY","url":"https://openai.com/index/bny/","publisher":"OpenAI","date":"2025"}],"happenedAt":"2025-10","addedAt":"2026-10-16","tags":["agent identity","access control","governance","payments operations"],"sponsored":false,"slug":"bny-digital-employees","toolSlugs":["eliza-internal","openai"],"url":"https://agentsinthewild.co/use-cases/bny-digital-employees"},{"title":"Google DeepMind's CodeMender agent upstreamed 72 security fixes to open-source projects","summary":"CodeMender finds and fixes security flaws in code. In its first six months, DeepMind used it to send 72 security fixes to open-source projects, some with millions of lines of code. Humans review every patch.","category":"security","organization":"Google DeepMind","industry":"Security Research","tools":["CodeMender","Gemini Deep Think"],"outcome":"72 security fixes upstreamed to open-source projects in six months, including projects as large as 4.5 million lines of code.","metrics":[{"label":"Fixes upstreamed","value":"72"},{"label":"Largest codebase","value":"4.5M lines"}],"evidence":"company-reported","caveats":["CodeMender is a research project, not a product you can use.","Human researchers review every patch before it is submitted. The count is of accepted human-reviewed patches, not fully autonomous work."],"sources":[{"title":"Introducing CodeMender: an AI agent for code security","url":"https://deepmind.google/discover/blog/introducing-codemender-an-ai-agent-for-code-security/","publisher":"Google DeepMind","date":"2025-10"},{"title":"Google DeepMind unveils CodeMender AI agent that autonomously patches software vulnerabilities","url":"https://siliconangle.com/2025/10/06/google-deepmind-unveils-codemender-ai-agent-autonomously-patches-software-vulnerabilities/","publisher":"SiliconANGLE","date":"2025-10"}],"happenedAt":"2025-10","addedAt":"2026-10-16","tags":["vulnerability repair","open source","memory safety","self-critique"],"sponsored":false,"slug":"google-deepmind-codemender","toolSlugs":["codemender","gemini-deep-think"],"url":"https://agentsinthewild.co/use-cases/google-deepmind-codemender"},{"title":"Aider, the open-source coding agent, writes most of its own code","summary":"The Aider project tracks how much of each release Aider itself wrote. Recent releases report figures as high as 88%. It is a public, line-level record of an agent building its own product.","category":"software-engineering","organization":"Aider (open source)","industry":"Developer Tools","tools":["Aider"],"outcome":"Aider's release notes report that Aider wrote up to 88% of the code in a release (v0.86.0, August 2025).","metrics":[{"label":"Share written by Aider","value":"up to 88% per release"}],"evidence":"company-reported","caveats":["The project measures this itself, by counting lines in git blame. Lines are not the same as effort or difficulty.","The figure changes a lot from release to release."],"sources":[{"title":"Aider release history","url":"https://aider.chat/HISTORY.html","publisher":"Aider","date":"2025-08"},{"title":"Self-assembly: how much of Aider's code was written by Aider","url":"https://aider.chat/2024/05/24/self-assembly.html","publisher":"Aider","date":"2024-05"}],"happenedAt":"2025-08","addedAt":"2026-10-16","tags":["open source","pair programming","dogfooding","terminal agent"],"sponsored":false,"slug":"aider-writes-its-own-code","toolSlugs":["aider"],"url":"https://agentsinthewild.co/use-cases/aider-writes-its-own-code"},{"title":"Salesforce runs its own support on Agentforce and cut support headcount from ~9,000 to ~5,000","summary":"Salesforce calls itself 'customer zero' for Agentforce. Its AI agents now handle a large share of customer support conversations on its help site, and the company reduced support roles.","category":"customer-support","organization":"Salesforce","industry":"Enterprise Software","tools":["Agentforce"],"outcome":"About 1.5M customer conversations handled by agents, with CSAT about the same as humans; support headcount went from ~9,000 to ~5,000.","metrics":[{"label":"Agent conversations","value":"~1.5M"},{"label":"Support headcount","value":"~9,000 → ~5,000"},{"label":"Support cost","value":"-17% in 2025"}],"evidence":"company-reported","caveats":["Figures come from CEO Marc Benioff on a podcast and from company spokespeople, not an audit.","Salesforce sells Agentforce, so it has a strong reason to show success.","Benioff called the change a 'rebalance'; Salesforce says hundreds of staff moved to sales and customer success."],"sources":[{"title":"Salesforce CEO says it cut 4,000 support jobs - and replaced them with AI","url":"https://techradar.com/pro/salesforce-says-it-cuts-4-000-support-jobs-and-replaced-them-with-ai","publisher":"TechRadar","date":"2025-09"},{"title":"AI agents drive 4,000 job cuts in Salesforce support division","url":"https://www.salesforceben.com/ai-agents-drive-4000-job-cuts-in-salesforce-support-division/","publisher":"Salesforce Ben","date":"2025-09"}],"happenedAt":"2025-08","addedAt":"2026-10-16","tags":["chat support","workforce impact","dogfooding"],"sponsored":false,"slug":"salesforce-agentforce-support","toolSlugs":["agentforce"],"url":"https://agentsinthewild.co/use-cases/salesforce-agentforce-support"},{"title":"Goldman Sachs pilots Devin as an AI engineer next to its 12,000 developers","summary":"Goldman Sachs started a pilot with Cognition's Devin. The tech chief said the bank could run hundreds and later thousands of Devin instances on tasks like moving code to newer languages, under human supervision.","category":"software-engineering","organization":"Goldman Sachs","industry":"Banking","tools":["Devin"],"outcome":"Pilot started with plans for hundreds of instances; the CIO expects 3–4x developer output on routine code updates.","metrics":[{"label":"Human developers","value":"~12,000"},{"label":"Expected output gain","value":"3–4x (estimate)"}],"evidence":"press-report","caveats":["This is a pilot announcement, not a measured result.","The 3–4x figure is the CIO's expectation."],"sources":[{"title":"Goldman Sachs is piloting its first autonomous coder","url":"https://www.nbcchicago.com/news/business/money-report/goldman-sachs-is-piloting-its-first-autonomous-coder-in-major-ai-milestone-for-wall-street/3787263/","publisher":"CNBC (via NBC Chicago)","date":"2025-07"}],"happenedAt":"2025-07","addedAt":"2026-10-16","tags":["code modernization","enterprise rollout","pilot"],"sponsored":false,"slug":"goldman-sachs-devin-pilot","toolSlugs":["devin"],"url":"https://agentsinthewild.co/use-cases/goldman-sachs-devin-pilot"},{"title":"Ramp's policy agent reviews expenses against company policy for 1,000+ finance teams","summary":"Ramp launched AI agents for controllers that read a company's expense policy, approve low-risk spend, flag suspicious receipts, and answer employee policy questions, with an audit trail for each decision.","category":"finance-operations","organization":"Ramp","industry":"Fintech","tools":["Ramp Policy Agent","OpenAI reasoning models"],"outcome":"Used by 1,000+ finance teams within six months; Ramp claims 99% approval accuracy and 7x more out-of-policy spend caught.","metrics":[{"label":"Finance teams","value":"1,000+"},{"label":"Approval accuracy","value":"99% (vendor)"},{"label":"Out-of-policy caught","value":"7x (vendor)"}],"evidence":"vendor-case-study","caveats":["All performance figures come from Ramp.","'Accuracy' is measured against Ramp's own labels, method not published in the source."],"sources":[{"title":"Ramp Policy Agent is generally available","url":"https://ramp.com/blog/ramp-policy-agent-ga-launch","publisher":"Ramp","date":"2026-01"},{"title":"Ramp introduces AI agents suited for controllers","url":"https://www.cpapracticeadvisor.com/2025/07/10/ramp-introduces-ai-agents-suited-for-controllers/164684/","publisher":"CPA Practice Advisor","date":"2025-07"}],"happenedAt":"2025-07","addedAt":"2026-10-16","tags":["expense management","policy enforcement","fraud","audit trail"],"sponsored":false,"slug":"ramp-policy-agent-expenses","toolSlugs":["ramp-policy-agent","openai-reasoning-models"],"url":"https://agentsinthewild.co/use-cases/ramp-policy-agent-expenses"},{"title":"Cloudflare built an OAuth 2.1 library with Claude and published the prompts in git","summary":"Cloudflare engineers wrote workers-oauth-provider largely with Claude and kept the prompts in the commit history. A security expert later reviewed it and found real flaws, which makes this a useful case study.","category":"software-engineering","organization":"Cloudflare","industry":"Cloud & Security","tools":["Claude"],"outcome":"A working OAuth 2.1 provider library for Cloudflare Workers, written mostly by Claude and reviewed by Cloudflare engineers.","metrics":[{"label":"Prompts public","value":"Yes, in commit history"}],"evidence":"independent","caveats":["An independent review by Neil Madden found an implicit-grant flow (removed in OAuth 2.1) and a Basic-auth bug. Human review still matters for security code.","The library was labelled beta at the time."],"sources":[{"title":"cloudflare/workers-oauth-provider","url":"https://github.com/cloudflare/workers-oauth-provider","publisher":"GitHub","date":"2025"},{"title":"A look at CloudFlare's AI-coded OAuth library","url":"https://neilmadden.blog/?p=3982","publisher":"Neil Madden","date":"2025-06"}],"happenedAt":"2025-06","addedAt":"2026-10-16","tags":["security","open source","prompts in git","code review"],"sponsored":false,"slug":"cloudflare-oauth-library-claude","toolSlugs":["claude"],"url":"https://agentsinthewild.co/use-cases/cloudflare-oauth-library-claude"},{"title":"XBOW's autonomous pentesting agent reached #1 on HackerOne's US leaderboard","summary":"XBOW, an AI agent for offensive security testing, submitted over 1,000 vulnerability reports to bug bounty programs and reached the top of HackerOne's US leaderboard in 2025.","category":"security","organization":"XBOW","industry":"Cybersecurity","tools":["XBOW"],"outcome":"#1 on the HackerOne US leaderboard, with 1,000+ reports in a few months and 130+ confirmed and fixed by program owners.","metrics":[{"label":"Reports submitted","value":"1,000+"},{"label":"Confirmed and resolved","value":"132"},{"label":"Triaged, not yet resolved","value":"303"}],"evidence":"press-report","caveats":["Humans reviewed findings before submission, to follow HackerOne's AI policy.","Most figures come from XBOW's own blog and interviews.","Leaderboard rank reflects volume as well as quality."],"sources":[{"title":"Top 1: how XBOW did it","url":"https://xbow.com/blog/top-1-how-xbow-did-it/","publisher":"XBOW","date":"2025-06"},{"title":"AI bug hunter claims top spot on HackerOne's leaderboard","url":"https://www.techrepublic.com/article/news-ai-xbow-tops-hackerone-us-leaderboad/","publisher":"TechRepublic","date":"2025-06"}],"happenedAt":"2025-06","addedAt":"2026-10-16","tags":["pentesting","bug bounty","web security","validation"],"sponsored":false,"slug":"xbow-hackerone-leaderboard","toolSlugs":["xbow"],"url":"https://agentsinthewild.co/use-cases/xbow-hackerone-leaderboard"},{"title":"Airbnb migrated 3,500 test files with an LLM pipeline in 6 weeks instead of 1.5 years","summary":"Airbnb moved nearly 3,500 React test files from Enzyme to React Testing Library using an automated, step-by-step LLM pipeline with retry loops. Engineers had estimated 1.5 years of manual work.","category":"software-engineering","organization":"Airbnb","industry":"Travel","tools":["LLM pipeline (custom)"],"outcome":"~3,500 test files migrated in 6 weeks; about 97% fully automated and the rest finished by hand.","metrics":[{"label":"Files migrated","value":"~3,500"},{"label":"Time","value":"6 weeks vs ~1.5 years est."},{"label":"Automated","value":"~97%"}],"evidence":"company-reported","caveats":["The 1.5-year manual estimate is Airbnb's own.","Going from about 75% to 97% automated took much more effort than the first 75%."],"sources":[{"title":"Accelerating Large-Scale Test Migration with LLMs","url":"https://airbnb.tech/?p=366","publisher":"Airbnb Tech Blog","date":"2025-03"},{"title":"Airbnb uses LLMs to migrate tests","url":"https://www.infoq.com/news/2025/03/airbnb-llm-test-migration/","publisher":"InfoQ","date":"2025-03"}],"happenedAt":"2025-03","addedAt":"2026-10-16","tags":["code migration","testing","react","retry loop","pipeline"],"sponsored":false,"slug":"airbnb-enzyme-rtl-migration","toolSlugs":["llm-pipeline-custom"],"url":"https://agentsinthewild.co/use-cases/airbnb-enzyme-rtl-migration"},{"title":"Sakana's AI Scientist-v2 wrote a paper that passed peer review at an ICLR workshop","summary":"Sakana AI's agent generated a full research paper, from hypothesis to experiments to write-up. It was accepted at an ICLR 2025 workshop with scores of 6, 7, and 6, then withdrawn as planned.","category":"research","organization":"Sakana AI","industry":"AI Research","tools":["The AI Scientist-v2"],"outcome":"One of three fully AI-generated submissions cleared the acceptance bar at the ICBINB workshop (scores 6, 7, 6).","metrics":[{"label":"Reviewer scores","value":"6 / 7 / 6"},{"label":"AI papers submitted","value":"3 (1 accepted)"}],"evidence":"research-paper","caveats":["Humans chose the topic and picked which papers to submit.","Sakana said none of the papers would meet its bar for the main conference.","Outside reviewers reported hallucinations and overstated novelty in the accepted paper.","The experiment ran with the organizers' cooperation, and the paper was withdrawn after acceptance."],"sources":[{"title":"The AI Scientist generates its first peer-reviewed scientific publication","url":"https://sakana.ai/ai-scientist-first-publication/","publisher":"Sakana AI","date":"2025-03"},{"title":"Sakana claims its AI paper passed peer review, but it's a bit more nuanced","url":"https://techcrunch.com/2025/03/12/sakana-claims-its-ai-paper-passed-peer-review-but-its-a-bit-more-nuanced-than-that","publisher":"TechCrunch","date":"2025-03"}],"happenedAt":"2025-03","addedAt":"2026-10-16","tags":["automated research","machine learning","peer review","experiments"],"sponsored":false,"slug":"sakana-ai-scientist-peer-review","toolSlugs":["the-ai-scientist-v2"],"url":"https://agentsinthewild.co/use-cases/sakana-ai-scientist-peer-review"},{"title":"Google's AI co-scientist proposed the same superbug hypothesis an Imperial College team spent years on","summary":"Google's multi-agent AI co-scientist was asked how some antibiotic-resistant bacteria spread genes. In about two days it proposed the mechanism that Imperial College researchers had found but not yet published.","category":"research","organization":"Google Research & Imperial College London","industry":"Life Sciences","tools":["AI co-scientist","Gemini 2.0"],"outcome":"The system's top hypothesis matched the Imperial team's unpublished finding about how bacteria spread resistance genes.","metrics":[{"label":"Time to hypothesis","value":"~2 days (reported)"},{"label":"Agents in the system","value":"6 specialized roles"}],"evidence":"research-paper","caveats":["This is one test where the answer was already known. It does not show the system finds new truths reliably.","Google's authors list limits: literature review, fact-checking, and larger expert evaluations need work.","The 'decade vs. two days' framing comes from press coverage, not from a controlled comparison."],"sources":[{"title":"Accelerating scientific breakthroughs with an AI co-scientist","url":"https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist","publisher":"Google Research","date":"2025-02"},{"title":"Google's AI co-scientist could enhance research, say Imperial researchers","url":"https://www.imperial.ac.uk/news/261293/googles-ai-co-scientist-could-enhance-research/","publisher":"Imperial College London","date":"2025-02"}],"happenedAt":"2025-02","addedAt":"2026-10-16","tags":["multi-agent","hypothesis generation","biology","scientific discovery"],"sponsored":false,"slug":"google-ai-co-scientist","toolSlugs":["ai-co-scientist","gemini-2-0"],"url":"https://agentsinthewild.co/use-cases/google-ai-co-scientist"},{"title":"Lyft cut average support resolution time by 87% with a Claude-based assistant","summary":"Lyft built a customer care assistant on Claude through Amazon Bedrock. It tries to solve rider and driver issues first, and hands hard cases to people.","category":"customer-support","organization":"Lyft","industry":"Ride-hailing","tools":["Claude","Amazon Bedrock"],"outcome":"Average resolution time fell by more than 87%, with thousands of requests resolved per day (company claim).","metrics":[{"label":"Resolution time","value":"-87%"},{"label":"Decision accuracy","value":"+30% (vendor)"}],"evidence":"vendor-case-study","caveats":["Numbers come from Lyft and Anthropic, not an independent audit.","An issue counts as 'resolved' when the customer answers yes to 'Did we resolve your issue?'. This is self-reported.","Safety, deactivation, and fraud issues still go to human agents."],"sources":[{"title":"Lyft reduces customer support time by 87% with Claude","url":"https://claude.com/customers/lyft","publisher":"Anthropic","date":"2025"},{"title":"Lyft's new AI customer assistant is powered by Anthropic's Claude","url":"https://techcrunch.com/2025/02/06/lyfts-new-ai-customer-assistant-is-powered-by-anthropics-claude","publisher":"TechCrunch","date":"2025-02"}],"happenedAt":"2025-02","addedAt":"2026-10-16","tags":["chat support","escalation","two-sided marketplace"],"sponsored":false,"slug":"lyft-claude-customer-care","toolSlugs":["claude","amazon-bedrock"],"url":"https://agentsinthewild.co/use-cases/lyft-claude-customer-care"},{"title":"Google uses LLMs for internal code migrations and reports about 50% less migration time","summary":"Google engineers published papers on using LLMs to run large internal code migrations, such as JUnit3 to JUnit4 and int32 to int64 IDs. Most of the landed code changes were generated by the model.","category":"software-engineering","organization":"Google","industry":"Technology","tools":["Gemini (internal)"],"outcome":"Across 39 migrations, about 74% of code changes were LLM-generated; developers estimated a 50% cut in total migration time.","metrics":[{"label":"Migrations studied","value":"39"},{"label":"LLM-generated changes","value":"~74%"},{"label":"Time saved","value":"~50% (dev estimate)"}],"evidence":"research-paper","caveats":["Time savings are developer estimates, not a controlled experiment.","Results differ a lot between migrations."],"sources":[{"title":"Migrating Code At Scale With LLMs At Google (FSE 2025)","url":"https://conf.researchr.org/details/fse-2025/fse-2025-industry-papers/14/Migrating-Code-At-Scale-With-LLMs-At-Google","publisher":"FSE 2025","date":"2025"},{"title":"Google reports halving code migration time with AI help","url":"https://www.theregister.com/2025/01/16/google_ai_code_migration/","publisher":"The Register","date":"2025-01"}],"happenedAt":"2025-01","addedAt":"2026-10-16","tags":["code migration","monorepo","research"],"sponsored":false,"slug":"google-llm-code-migrations","toolSlugs":["gemini-internal"],"url":"https://agentsinthewild.co/use-cases/google-llm-code-migrations"},{"title":"IBM's AskHR agent resolves 94% of common HR questions without a person","summary":"IBM has refined its internal HR agent, AskHR, for six years. It now automates more than 80 HR tasks and handled over 11.5 million employee interactions in 2024.","category":"hr-recruiting","organization":"IBM","industry":"Technology","tools":["AskHR (watsonx Orchestrate)"],"outcome":"94% containment of common questions, 75% fewer HR support tickets since 2016, and a 40% cut in HR operating costs over four years.","metrics":[{"label":"Containment (common questions)","value":"94%"},{"label":"Interactions in 2024","value":"11.5M+"},{"label":"Support tickets since 2016","value":"-75%"},{"label":"HR operating cost (4 yrs)","value":"-40%"}],"evidence":"company-reported","caveats":["IBM sells the platform behind AskHR, so it has a reason to show success.","94% applies to common questions, not all HR issues. Sensitive matters such as misconduct still go to people.","The cost cut is described as a contribution from the agent, not caused by it alone."],"sources":[{"title":"Transforming HR support with agentic AI","url":"https://www.ibm.com/case-studies/ibm-askhr","publisher":"IBM","date":"2025"},{"title":"IBM CHRO lessons from initial AI rollout","url":"https://www.fortune.com/2024/07/12/ibm-chro-lessons-from-initial-ai-rollout","publisher":"Fortune","date":"2024-07"}],"happenedAt":"2025","addedAt":"2026-10-16","tags":["employee self-service","HR operations","internal agent","containment"],"sponsored":false,"slug":"ibm-askhr-agent","toolSlugs":["askhr-watsonx-orchestrate"],"url":"https://agentsinthewild.co/use-cases/ibm-askhr-agent"},{"title":"Intercom's Fin AI agent resolves about two-thirds of support conversations on average","summary":"Intercom's Fin answers customer support conversations for thousands of companies. Intercom reports an average resolution rate of about 66–67% across its customer base, rising over time.","category":"customer-support","organization":"Intercom (and its customers)","industry":"Customer Service Software","tools":["Fin","Claude"],"outcome":"Average resolution rate of 66–67% across about 6,000 customers, up from about 51% at launch.","metrics":[{"label":"Avg. resolution rate","value":"66–67%"},{"label":"Customers","value":"~6,000"},{"label":"Out-of-box (earlier)","value":"~51%"}],"evidence":"vendor-case-study","caveats":["Intercom sells Fin and defines 'resolution' itself.","Results vary a lot by industry, question type, and how good the help content is."],"sources":[{"title":"Intercom announces latest evolution of its AI customer service agent","url":"https://diginomica.com/intercom-announces-latest-evolution-its-ai-customer-service-agent-just-getting-started","publisher":"diginomica","date":"2025"},{"title":"Intercom case study","url":"https://aws.amazon.com/solutions/case-studies/intercom-anthropic/","publisher":"AWS","date":"2025"}],"happenedAt":"2025","addedAt":"2026-10-16","tags":["chat support","help center","resolution rate","benchmark"],"sponsored":false,"slug":"intercom-fin-resolution-rate","toolSlugs":["fin","claude"],"url":"https://agentsinthewild.co/use-cases/intercom-fin-resolution-rate"},{"title":"McKinsey's Lilli platform is used regularly by over 70% of the firm's staff","summary":"McKinsey built Lilli, an internal generative AI platform that searches the firm's knowledge and helps with research, summaries, and drafting. It reports that a large majority of its people use it regularly.","category":"internal-tools","organization":"McKinsey & Company","industry":"Consulting","tools":["Lilli (internal)"],"outcome":"More than 70% of the firm uses Lilli regularly, with millions of prompts answered (company figures).","metrics":[{"label":"Regular users","value":"70%+ of staff"}],"evidence":"company-reported","caveats":["McKinsey also sells AI consulting, so its own story is also marketing.","'Regularly' is not defined, and figures differ between McKinsey pages and dates.","Usage is not the same as time saved or better work."],"sources":[{"title":"Rewiring the way McKinsey works with Lilli, our generative AI platform","url":"https://www.mckinsey.com/capabilities/tech-and-ai/how-we-help-clients/rewiring-the-way-mckinsey-works-with-lilli","publisher":"McKinsey","date":"2025"},{"title":"Meet Lilli, our generative AI tool","url":"https://www.mckinsey.com/about-us/new-at-mckinsey-blog/meet-lilli-our-generative-ai-tool","publisher":"McKinsey","date":"2023-08"}],"happenedAt":"2025","addedAt":"2026-10-16","tags":["knowledge management","enterprise search","adoption","training"],"sponsored":false,"slug":"mckinsey-lilli-platform","toolSlugs":["lilli-internal"],"url":"https://agentsinthewild.co/use-cases/mckinsey-lilli-platform"},{"title":"Rakuten engineer let Claude Code work alone for 7 hours on a vLLM feature","summary":"A Rakuten engineer asked Claude Code to add an activation vector extraction method to vLLM, a large open-source library. The agent worked for seven hours in one run with only occasional guidance.","category":"software-engineering","organization":"Rakuten","industry":"E-commerce & Tech","tools":["Claude Code"],"outcome":"Seven hours of mostly autonomous coding in one run. Rakuten reports 99.9% numerical accuracy versus the reference method.","metrics":[{"label":"Autonomous run","value":"7 hours"},{"label":"Numerical accuracy","value":"99.9%"}],"evidence":"vendor-case-study","caveats":["Published as an Anthropic customer story. No independent verification.","The engineer gave occasional guidance, so this was not fully hands-off."],"sources":[{"title":"Rakuten accelerates development with Claude Code","url":"https://www.claude.com/customers/rakuten","publisher":"Anthropic","date":"2025"}],"happenedAt":"2025","addedAt":"2026-10-16","tags":["long-running agent","open source","ML infrastructure"],"sponsored":false,"slug":"rakuten-claude-code-vllm","toolSlugs":["claude-code"],"url":"https://agentsinthewild.co/use-cases/rakuten-claude-code-vllm"},{"title":"Vercel's lead agent took inbound qualification from a 10-person team to about one person","summary":"Vercel built an agent that researches and qualifies inbound sales leads. A GTM engineer shadowed the best SDR, copied her workflow, and ran the agent in shadow mode before going live.","category":"sales-marketing","organization":"Vercel","industry":"Developer Platform","tools":["Custom agent (Vercel AI SDK)"],"outcome":"Inbound SDR work went from about 10 people to 1 plus a part-time person; the agent costs about $5,000 a year to run (as reported).","metrics":[{"label":"SDR team","value":"~10 → ~1"},{"label":"Run cost","value":"~$5,000/yr"},{"label":"Maintenance","value":"~20% of one engineer"}],"evidence":"press-report","caveats":["Numbers come from a conference talk by Vercel's COO, reported by SaaStr.","Team-size figures differ between sources."],"sources":[{"title":"Vercel took a 10-person SDR team down to 1","url":"https://www.saastr.com/vercel-took-a-10-person-sdr-team-down-to-1-the-whole-thing-costs-5000-a-year-with-vercels-coo-jeanne-dewitt-grosser/","publisher":"SaaStr","date":"2025"},{"title":"What we learned building agents at Vercel","url":"https://vercel.com/blog/the-sweet-spot-for-current-generation-agentic-ai","publisher":"Vercel","date":"2025-11"}],"happenedAt":"2025","addedAt":"2026-10-16","tags":["lead qualification","inbound sales","shadow mode","GTM engineering"],"sponsored":false,"slug":"vercel-lead-qualification-agent","toolSlugs":["custom-agent-vercel-ai-sdk"],"url":"https://agentsinthewild.co/use-cases/vercel-lead-qualification-agent"},{"title":"Google's Big Sleep agent found an exploitable SQLite bug that fuzzing missed","summary":"Google Project Zero and DeepMind built Big Sleep, an LLM agent for vulnerability research. It found a stack buffer underflow in a development version of SQLite before the code was released.","category":"security","organization":"Google (Project Zero & DeepMind)","industry":"Security Research","tools":["Big Sleep","Gemini 1.5 Pro"],"outcome":"Found a previously unknown, exploitable memory-safety bug in SQLite; developers patched it the same day, before any release.","metrics":[{"label":"Users exposed","value":"0 (unreleased code)"},{"label":"Time to patch","value":"Same day"}],"evidence":"research-paper","caveats":["Big Sleep was a research project at the time, not a product.","Google says traditional fuzzing did not find this bug, but one case does not show the agent beats fuzzing in general."],"sources":[{"title":"From Naptime to Big Sleep","url":"https://googleprojectzero.blogspot.com/2024/10/from-naptime-to-big-sleep.html","publisher":"Google Project Zero","date":"2024-11"},{"title":"Google says its AI found SQLite vulnerability that fuzzing missed","url":"https://www.securityweek.com/google-says-its-ai-found-sqlite-vulnerability-that-fuzzing-missed/","publisher":"SecurityWeek","date":"2024-11"}],"happenedAt":"2024-11","addedAt":"2026-10-16","tags":["vulnerability research","variant analysis","memory safety","open source"],"sponsored":false,"slug":"google-big-sleep-sqlite","toolSlugs":["big-sleep","gemini-1-5-pro"],"url":"https://agentsinthewild.co/use-cases/google-big-sleep-sqlite"},{"title":"LinkedIn's Hiring Assistant agent finds candidates and drafts outreach for recruiters","summary":"LinkedIn launched Hiring Assistant, its first AI agent, to take over repetitive recruiting work such as sourcing candidates, writing outreach, and scheduling, while recruiters stay in control.","category":"hr-recruiting","organization":"LinkedIn","industry":"HR Tech","tools":["LinkedIn Hiring Assistant"],"outcome":"Launched to charter customers including AMD, Canva, Siemens, and Zurich Insurance, aiming at the 20+ hours a week recruiters spend on admin.","metrics":[{"label":"Recruiter admin time (baseline)","value":"20+ hrs/week"}],"evidence":"company-reported","caveats":["This is a product launch. LinkedIn has not published outcome data in the source.","Matching on skills reduces some bias risks, but recruiters still need to audit shortlists."],"sources":[{"title":"Introducing Hiring Assistant","url":"https://news.linkedin.com/2024/October/introducing-hiring-assistant-to-help-recruiters-spend-more-time-","publisher":"LinkedIn","date":"2024-10"},{"title":"LinkedIn launches new AI assistant for hiring","url":"https://www.hr-brew.com/stories/2024/10/29/linkedin-launches-new-ai-assistant-for-hiring-to-help-ta-pros-get-back-to-strategic-work","publisher":"HR Brew","date":"2024-10"}],"happenedAt":"2024-10","addedAt":"2026-10-16","tags":["sourcing","outreach","scheduling"],"sponsored":false,"slug":"linkedin-hiring-assistant","toolSlugs":["linkedin-hiring-assistant"],"url":"https://agentsinthewild.co/use-cases/linkedin-hiring-assistant"},{"title":"Uber's Genie answers on-call questions in Slack and saved ~13,000 engineering hours","summary":"Uber built Genie, a retrieval-based copilot that answers engineers' questions in internal Slack support channels using wiki pages, internal Stack Overflow, and docs.","category":"internal-tools","organization":"Uber","industry":"Ride-hailing & Delivery","tools":["OpenAI embeddings","RAG","Slack"],"outcome":"70,000+ questions answered across 154 Slack channels, with an estimated 13,000 engineering hours saved.","metrics":[{"label":"Questions answered","value":"70,000+"},{"label":"Slack channels","value":"154"},{"label":"Hours saved","value":"~13,000 (estimate)"},{"label":"Rated helpful","value":"48.9%"}],"evidence":"company-reported","caveats":["Hours saved is Uber's own estimate.","Helpfulness (48.9%) is rated by users, so about half of answers were not marked helpful."],"sources":[{"title":"Genie: Uber's Gen AI On-Call Copilot","url":"https://www.uber.com/blog/genie-ubers-gen-ai-on-call-copilot/","publisher":"Uber Engineering","date":"2024-10"},{"title":"Uber's Genie RAG copilot","url":"https://www.infoq.com/news/2024/10/uber-genie-rag-copilot","publisher":"InfoQ","date":"2024-10"}],"happenedAt":"2024-10","addedAt":"2026-10-16","tags":["RAG","on-call","Slack bot","knowledge base"],"sponsored":false,"slug":"uber-genie-on-call-copilot","toolSlugs":["openai-embeddings","rag","slack"],"url":"https://agentsinthewild.co/use-cases/uber-genie-on-call-copilot"},{"title":"Uber's QueryGPT turns plain-English questions into SQL, cutting query time from ~10 to ~3 minutes","summary":"Uber built QueryGPT to write SQL from natural-language questions over its huge data warehouse. It started as a hackathon idea and grew into a tool with hundreds of daily users.","category":"data-analytics","organization":"Uber","industry":"Ride-hailing & Delivery","tools":["QueryGPT (internal)","RAG","LLM"],"outcome":"Uber estimates a query that took about 10 minutes to write by hand now takes about 3 minutes; the limited release reached ~300 daily active users.","metrics":[{"label":"Time per query","value":"~10 → ~3 min (estimate)"},{"label":"Daily active users (limited release)","value":"~300"}],"evidence":"company-reported","caveats":["Uber calls the 10-to-3-minute figure a conservative estimate, not a measured before/after result.","Some third-party write-ups extrapolate huge monthly hour savings. Those are not Uber's numbers."],"sources":[{"title":"QueryGPT – Natural Language to SQL Using Generative AI","url":"https://www.uber.com/blog/query-gpt/","publisher":"Uber Engineering","date":"2024-09"},{"title":"Natural language to SQL query generation at scale","url":"https://www.zenml.io/llmops-database/natural-language-to-sql-query-generation-at-scale","publisher":"ZenML LLMOps Database","date":"2024"}],"happenedAt":"2024-09","addedAt":"2026-10-16","tags":["text-to-SQL","RAG","data warehouse","analytics","hackathon"],"sponsored":false,"slug":"uber-querygpt-text-to-sql","toolSlugs":["querygpt-internal","rag","llm"],"url":"https://agentsinthewild.co/use-cases/uber-querygpt-text-to-sql"},{"title":"Amazon upgraded 30,000 internal apps to Java 17 with Amazon Q Developer","summary":"Amazon used its own coding agent to move tens of thousands of internal applications from Java 8 and 11 to Java 17. Andy Jassy said it saved about 4,500 developer-years of work.","category":"software-engineering","organization":"Amazon","industry":"E-commerce & Cloud","tools":["Amazon Q Developer"],"outcome":"30,000 apps upgraded. Amazon estimates 4,500 developer-years saved and $260M a year from performance gains.","metrics":[{"label":"Apps upgraded","value":"30,000"},{"label":"Developer-years saved","value":"~4,500 (estimate)"},{"label":"Annual savings","value":"$260M (estimate)"},{"label":"Avg. upgrade time","value":"~50 dev-days → hours"}],"evidence":"company-reported","caveats":["Amazon's own internal estimates. Amazon also sells the tool.","Time saved was estimated by counting migrated dependencies, at about one developer-day each.","The $260M figure comes from hosts removed thanks to Java 17 performance, not from the agent itself."],"sources":[{"title":"Amazon Q Developer just reached a $260 million dollar milestone","url":"https://aws.amazon.com/blogs/devops/amazon-q-developer-just-reached-a-260-million-dollar-milestone/","publisher":"AWS DevOps Blog","date":"2024-08"}],"happenedAt":"2024-08","addedAt":"2026-10-16","tags":["code migration","java","dependency upgrades","platform engineering"],"sponsored":false,"slug":"amazon-q-java-upgrades","toolSlugs":["amazon-q-developer"],"url":"https://agentsinthewild.co/use-cases/amazon-q-java-upgrades"},{"title":"Kaiser Permanente rolled out Abridge ambient AI notes to 24,000+ doctors","summary":"Kaiser Permanente deployed Abridge's ambient documentation tool across about 40 hospitals and 600+ medical offices. It listens to visits, with consent, and drafts clinical notes for the doctor to review.","category":"healthcare","organization":"Kaiser Permanente","industry":"Healthcare","tools":["Abridge"],"outcome":"One of the largest generative AI rollouts in healthcare: 24,000+ physicians across ~40 hospitals and 600+ offices.","metrics":[{"label":"Physicians","value":"24,000+"},{"label":"Hospitals","value":"~40"},{"label":"Medical offices","value":"600+"}],"evidence":"press-report","caveats":["The sources give rollout scale, not measured outcomes like time saved or burnout.","Kaiser's venture arm is an Abridge investor.","Patient consent is required and doctors review every note."],"sources":[{"title":"Kaiser Permanente rolls out Abridge's gen AI clinical tech across 40-hospital system","url":"https://www.fiercehealthcare.com/health-tech/kaiser-permanente-rolls-out-abridges-gen-ai-clinical-tech-across-40-hospitals-60","publisher":"Fierce Healthcare","date":"2024-08"}],"happenedAt":"2024-08","addedAt":"2026-10-16","tags":["ambient scribe","clinical documentation","EHR"],"sponsored":false,"slug":"kaiser-abridge-ambient-notes","toolSlugs":["abridge"],"url":"https://agentsinthewild.co/use-cases/kaiser-abridge-ambient-notes"},{"title":"Morgan Stanley's Debrief agent takes notes in advisor meetings and drafts the follow-up email","summary":"Morgan Stanley Wealth Management launched AI @ Morgan Stanley Debrief, an OpenAI-powered tool that joins client meetings with consent, writes notes, drafts a follow-up email, and saves a note to Salesforce.","category":"finance-operations","organization":"Morgan Stanley","industry":"Wealth Management","tools":["OpenAI GPT-4","Salesforce","Zoom"],"outcome":"Rolled out to the firm's financial advisors in 2024; the firm cites about 30 minutes saved per client meeting (testimonial).","metrics":[{"label":"Advisors targeted","value":"~15,000"},{"label":"Time saved per meeting","value":"~30 min (testimonial)"}],"evidence":"company-reported","caveats":["The time saving comes from an advisor testimonial, not a measured study.","No public data on how many advisors use it regularly.","Client consent is needed every time."],"sources":[{"title":"Morgan Stanley Wealth Management announces AI @ Morgan Stanley Debrief","url":"https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch","publisher":"Morgan Stanley","date":"2024-06"},{"title":"Morgan Stanley Wealth rolls out AI note-taking tool","url":"https://thinkadvisor.com/2024/06/26/morgan-stanley-wealth-rolls-out-ai-note-taking-tool","publisher":"ThinkAdvisor","date":"2024-06"}],"happenedAt":"2024-06","addedAt":"2026-10-16","tags":["meeting notes","CRM","financial advisors","consent","follow-up email"],"sponsored":false,"slug":"morgan-stanley-debrief","toolSlugs":["openai-gpt-4","salesforce","zoom"],"url":"https://agentsinthewild.co/use-cases/morgan-stanley-debrief"},{"title":"Klarna's AI assistant took two-thirds of support chats in its first month","summary":"Klarna put an OpenAI-powered assistant in front of customer service chats. It handled 2.3 million conversations in month one. In 2025 the CEO said quality had suffered and Klarna started hiring human agents again.","category":"customer-support","organization":"Klarna","industry":"Fintech","tools":["OpenAI GPT-4"],"outcome":"2.3M conversations in the first month, about two-thirds of all service chats, the work of roughly 700 full-time agents (company claim).","metrics":[{"label":"Conversations, month one","value":"2.3M"},{"label":"Share of chats","value":"~2/3"},{"label":"Resolution time","value":"11 min → <2 min"},{"label":"Markets / languages","value":"23 / 35+"}],"evidence":"company-reported","caveats":["All launch numbers come from Klarna and OpenAI. No independent audit.","In May 2025 the CEO said a cost-first AI push lowered service quality, and Klarna began recruiting human agents again so customers can always reach a person."],"sources":[{"title":"Klarna's AI assistant does the work of 700 full-time agents","url":"https://openai.com/index/klarna","publisher":"OpenAI","date":"2024-02"},{"title":"Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It","url":"https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396","publisher":"Entrepreneur","date":"2025-05"}],"happenedAt":"2024-02","addedAt":"2026-10-16","tags":["chat support","deflection","multilingual","lessons learned"],"sponsored":false,"slug":"klarna-ai-assistant-customer-service","toolSlugs":["openai-gpt-4"],"url":"https://agentsinthewild.co/use-cases/klarna-ai-assistant-customer-service"},{"title":"DoorDash built a voice AI agent for Dasher support calls with answers in under 2.5 seconds","summary":"DoorDash gets hundreds of thousands of support calls a day. With AWS, it built a voice self-service agent on Amazon Connect, Bedrock, and Claude 3 Haiku that answers routine Dasher questions.","category":"customer-support","organization":"DoorDash","industry":"Delivery","tools":["Claude 3 Haiku","Amazon Bedrock","Amazon Connect","Amazon Lex"],"outcome":"A voice agent ready for live testing in about two months, answering in 2.5 seconds or less, with 50x more testing capacity.","metrics":[{"label":"Response latency","value":"≤ 2.5 sec"},{"label":"Testing capacity","value":"50x"},{"label":"Build time to live test","value":"~2 months"},{"label":"Call volume","value":"100Ks per day (all calls)"}],"evidence":"vendor-case-study","caveats":["Published by AWS, which sold the platform.","The case study gives speed and testing numbers, not resolution rate or customer satisfaction."],"sources":[{"title":"DoorDash builds a generative AI contact center with Amazon Bedrock","url":"https://aws.amazon.com/solutions/case-studies/doordash-bedrock-case-study","publisher":"AWS","date":"2024"}],"happenedAt":"2024","addedAt":"2026-10-16","tags":["voice agent","contact center","latency","automated testing"],"sponsored":false,"slug":"doordash-dasher-voice-support","toolSlugs":["claude-3-haiku","amazon-bedrock","amazon-connect","amazon-lex"],"url":"https://agentsinthewild.co/use-cases/doordash-dasher-voice-support"},{"title":"Nubank used Devin to split an 8-year-old ETL monolith into sub-modules","summary":"Nubank had a core ETL codebase with millions of lines of code. Engineers gave the repetitive refactoring work to Devin and reviewed the output. Cognition reports large time and cost savings.","category":"software-engineering","organization":"Nubank","industry":"Banking","tools":["Devin"],"outcome":"Cognition reports 8–12x engineering-time efficiency and over 20x cost savings on the data-class migrations.","metrics":[{"label":"Engineering efficiency","value":"8–12x (vendor)"},{"label":"Cost savings","value":">20x (vendor)"}],"evidence":"vendor-case-study","caveats":["Published by Cognition, the company that sells Devin.","The case study page shows 8x in its header and 12x in its body text. Its own method note gives a range of 8–12x.","Efficiency is measured against hours spent prompting and reviewing Devin, not against total project time."],"sources":[{"title":"How Nubank refactors millions of lines of code with Devin","url":"https://devin.ai/customers","publisher":"Cognition","date":"2024"}],"happenedAt":"2024","addedAt":"2026-10-16","tags":["code migration","refactoring","monolith","ETL"],"sponsored":false,"slug":"nubank-devin-etl-migration","toolSlugs":["devin"],"url":"https://agentsinthewild.co/use-cases/nubank-devin-etl-migration"},{"title":"Pinterest's text-to-SQL agent helps analysts find tables and write queries 35% faster","summary":"Pinterest added an LLM SQL generator to Querybook, its open-source SQL tool, then added retrieval to pick the right tables from a very large warehouse.","category":"data-analytics","organization":"Pinterest","industry":"Social Media","tools":["Querybook","RAG","LLM"],"outcome":"Pinterest reports 35% faster task completion for SQL writing, and first-shot acceptance rising from 20% to over 40%.","metrics":[{"label":"Task speed","value":"+35%"},{"label":"First-shot acceptance","value":"20% → 40%+"}],"evidence":"company-reported","caveats":["The speed figure is from Pinterest's internal data and does not control for task differences.","We could only confirm figures from secondary write-ups so far."],"sources":[{"title":"Text-to-SQL system with RAG-enhanced table selection","url":"https://www.zenml.io/llmops-database/text-to-sql-system-with-rag-enhanced-table-selection","publisher":"ZenML LLMOps Database","date":"2024"}],"happenedAt":"2024","addedAt":"2026-10-16","tags":["text-to-SQL","RAG","data warehouse","analytics"],"sponsored":false,"slug":"pinterest-text-to-sql","toolSlugs":["querybook","rag","llm"],"url":"https://agentsinthewild.co/use-cases/pinterest-text-to-sql"},{"title":"Allen & Overy rolled out Harvey to 3,500 lawyers after a 40,000-query trial","summary":"Law firm Allen & Overy (now A&O Shearman) trialled the legal AI assistant Harvey for three months, then rolled it out firm-wide across 43 offices.","category":"legal","organization":"A&O Shearman (Allen & Overy)","industry":"Legal","tools":["Harvey","OpenAI GPT-4"],"outcome":"About 3,500 lawyers asked around 40,000 questions during the trial, followed by a firm-wide rollout across 43 offices.","metrics":[{"label":"Lawyers","value":"~3,500"},{"label":"Trial queries","value":"~40,000"},{"label":"Offices","value":"43"}],"evidence":"company-reported","caveats":["Query counts come from the firm's own announcement.","The firm stated that every output needs careful review by a lawyer.","Usage counts do not tell you about accuracy or time saved."],"sources":[{"title":"A&O announces exclusive launch partnership with Harvey","url":"https://www.aoshearman.com/en/news/ao-announces-exclusive-launch-partnership-with-harvey","publisher":"A&O Shearman","date":"2023-02"},{"title":"As Allen & Overy deploys GPT-based legal app Harvey firmwide","url":"https://www.lawnext.com/2023/02/as-allen-overy-deploys-gpt-based-legal-app-harvey-firmwide-founders-say-other-firms-will-soon-follow.html","publisher":"LawSites","date":"2023-02"}],"happenedAt":"2023-02","addedAt":"2026-10-16","tags":["legal research","drafting","professional services"],"sponsored":false,"slug":"allen-overy-harvey-legal","toolSlugs":["harvey","openai-gpt-4"],"url":"https://agentsinthewild.co/use-cases/allen-overy-harvey-legal"}]}