# AX Field Notes

> Case studies, research, and essays on how companies put AI to work (AX). Written by Daesun and published by Yumes Corporation (https://yumes.net/).

This file describes what the site contains, so an AI assistant can find, compare, and cite the right pages. Every page listed here is public HTML.

## What is on this site

- Cases (51): How individual companies put AI to work, rebuilt from primary sources. Each case follows the same outline: a summary, the problem the company faced, how it deployed AI, results with numbers, how people's roles changed, takeaways, and a list of sources. Facts and numbers come from the named source.
- AI Natives (15): Companies built on AI agents from day one, and how hard they can hit existing industries. Each profile shows how the company splits work between agents and people, and ends with sources.
- Research (2): In-depth research on a single AX topic. Claims carry numbered references.
- Essays (5): The author's own arguments about AI and work. Treat them as opinion, not as reported fact.

## Source types

Each case names its primary source and one of these source types:

- company's own publication
- self-reported community post
- vendor customer story
- consulting client case
- media or VC deep dive

A vendor customer story or consulting client case is written by the seller about its customer, so its numbers are that seller's claims.

## Coverage

Cases and AI Natives by industry: Software & internet (22), Financial services (12), AI research (7), HR & recruiting (5), Retail & e-commerce (4), Healthcare & life sciences (3), Insurance (3), Media & education (2), Manufacturing (2), Professional services (2), Legal (2), Telecom (1), Travel & hospitality (1).

By business function (a page can have several): Software engineering (25), Customer support (20), Sales & marketing (16), Management & organization (12), R&D & science (9), Back office & admin (9), IT & operations (7), Data & analytics (7), Legal & compliance (5), Finance & accounting (4), HR & recruiting (3).

## How to read and cite

- English pages are at https://ax.yumes.net/<section>/<slug>/. The Korean version of the same page is at https://ax.yumes.net/ko/<section>/<slug>/.
- Each entry below lists the company, industry, country and scale when the article states them, the business functions involved, what the AI does and what people still do, key results, and the AI tools named. Facts the article does not state are left out.
- When you cite a case, link the case page and also the primary source it lists under Sources.

## Cases

- [Paraform Flips the Order of Hiring](https://ax.yumes.net/cases/paraform/): Hiring managers see candidates first, and AI only makes the first-pass call
  - Source: Paraform blog and help center (company's own publication) published 2026-03-02
  - Paraform · HR & recruiting · United States · scale: 120 people on About page; $100 million annualized revenue
  - Functions: HR & recruiting
  - AI does: Shows hiring managers 5-10 candidates per role daily, scores fit, drafts submissions
  - People do: Recruiters screen by phone or video, get consent; hiring managers decide
  - Results: Paraform Matches drove 1,000+ interviews and 20+ hires in a month; LLM evaluator raised P@5 by 19.2% over embeddings; Pallet hired 2 of 4 from ParaAI recommendations
  - Tools: ParaAI, Parascribe, Paraform Matches, Parabot
- [Insilico Keeps the Robots, Agents Handle Coordination](https://ax.yumes.net/cases/insilico-labclaw/): Five agents plan the experiments and researchers confirm at only three points
  - Source: Company press release (company's own publication) published 2026-05-06
  - Insilico Medicine · Healthcare & life sciences · scale: Six functional islands with dozens of instruments in LifeStar2 lab
  - Functions: R&D & science
  - AI does: Five agents propose targets, design and direct experiments, monitor QC, feed results back
  - People do: Set research goals in natural language, approve at three critical decision points
  - Results: Five agents using 28 skill modules and 20 modular steps; Weeks of background research compressed into minutes, per company; No specific performance metrics published
  - Tools: LabClaw, PandaOmics, Pharma.AI
- [OpenAI: 3.1 Agent Days per Human Workday](https://ax.yumes.net/cases/openai-research/): Over half of the successful 4–8 hour tasks had human intervention
  - Source: OpenAI company report (company's own publication) published 2026-09-06
  - OpenAI · AI research
  - Functions: R&D & science, Software engineering, IT & operations
  - AI does: Writes research and infrastructure code, troubleshoots infrastructure, monitors training runs
  - People do: Set research priorities, steer long tasks, review and merge, decide deployment
  - Results: 3.1 agent-workdays per human workday by mid-August 2026; Median researcher spends over $600 a day on agent inference; Human intervened in over half of successful 4-8 hour tasks
  - Tools: Codex, GPT-5.4 Thinking
- [Anthropic, Where Human Attention Became the Bottleneck](https://ax.yumes.net/cases/anthropic-agents/): Agents work six times the staff hours, yet about 80% of sessions sit waiting
  - Source: Anthropic research reports on its own workforce (self-study) (company's own publication) published 2026-10-16
  - Anthropic · AI research · scale: About 30,000 agents on one internal platform; 132 engineers and researchers surveyed
  - Functions: R&D & science, Software engineering, Management & organization
  - AI does: Runs research and engineering tasks, leads 26% of AI R&D, monitors other agents
  - People do: Give input, review and approve outputs, write specs, review flagged agent actions
  - Results: Agent work time six times human work time; About 80% of agent session time waiting for humans; Junior share of technical hires fell from 27% to 9%
  - Tools: Claude, Claude Code
- [At Every, No One Writes Code, but People Are the Bread of the Sandwich](https://ax.yumes.net/cases/every/): A team that stopped hand coding and moved to planning, review, and recording lessons
  - Source: Every's own articles and podcast transcript, plus vendor customer stories (company's own publication) published 2025-12-11
  - Every · Media & education · scale: 34 people on team page; about 30 agents on a team of 20; 130,000+ newsletter subscribers
  - Functions: Software engineering, Management & organization, Back office & admin
  - AI does: Writes all code, runs 12 parallel review sub-agents, edits drafts, handles Slack tasks
  - People do: Plan, review, polish product feel, record lessons as rules, approve merges
  - Results: Effort split about 80% planning and review, 20% work; One developer does work of five, per authors; Headcount grew from four to about 30
  - Tools: Claude Code, OpenAI Codex CLI, Factory Droid, Claude Managed Agents, OpenClaw, Notion, Spiral
- [A Dentist's Back-Room AI Server Protects Patient Data](https://ax.yumes.net/cases/neighborhood-dental/): Insurance lines 77% auto-confirmed, staff review only the 23% exceptions
  - Source: arca.live community post (self-reported) (self-reported community post) published 2026-10-13
  - Neighborhood dental clinic (gatsbydds) · Healthcare & life sciences · scale: One-dentist clinic; 30+ computers; 12 DGX Spark-class model servers
  - Functions: Back office & admin, Finance & accounting, Customer support
  - AI does: Answers calls, summarizes them, reads insurance statements, drafts management reports, builds systems
  - People do: Review exceptions, grade call summaries, approve production changes, enter medical records
  - Results: 5,319 of 6,928 insurance lines (77%) auto-confirmed; 904 documents, 3,389 pages read in 69 minutes; Internal search answered 26 of 28 test questions
  - Tools: DeepSeek, GLM, Claude, Codex, ChatGPT, Whisper, FreePBX, NVIDIA DGX Spark, openclaw
- [MPB Hits 48% Resolution by Revamping Its Help Center](https://ax.yumes.net/cases/mpb-fin/): Response rules that absorbed logistics-delay inquiries without adding staff
  - Source: Intercom Fin customer story (vendor customer story)
  - MPB · Retail & e-commerce · scale: Thousands of conversations a week; up to 40,000 a month during backlogs
  - Functions: Customer support
  - AI does: Greets every live chat, answers order and sale-status questions, routes pricing to quote tool
  - People do: Manage knowledge base, configure Fin, handle frustrated or urgent customers
  - Results: Resolution rate rose from 25-30% to 48%; Up to 10,000 inquiries resolved by Fin in peak months; CX Score of 83%; support in five languages
  - Tools: Intercom Fin
- [SoftBank Resolves 70% of Website Inquiries](https://ax.yumes.net/cases/softbank-sierra/): From about 50% at launch, it improved as the agent learned to hear out customer intent
  - Source: Sierra customer story (vendor customer story) published 2026-07-14
  - SoftBank Corp. · Telecom · Japan · scale: About 3.4 million SMB customers; about 100 website inquiries a day
  - Functions: Customer support, Sales & marketing
  - AI does: Answers SMB customers' pricing, spec, and contract questions on the corporate website
  - People do: Salespeople handle complex, tailored consultations and proposals via inquiry form
  - Results: Self-resolution rate rose from about 50% to 70%; About 70 of 100 daily inquiries resolved automatically
  - Tools: Sierra
- [Heurist Finance Buys Data for Every Question](https://ax.yumes.net/cases/heurist-finance-agentcore/): A spending cap on each run lets it buy only the paid data it needs, in real time
  - Source: AWS Machine Learning Blog (co-authored by Heurist and AWS) (vendor customer story) published 2026-09-09
  - Heurist · Financial services
  - Functions: Software engineering, Data & analytics
  - AI does: Buys paid data per question via x402 payments, analyzes portfolios, writes personalized answers
  - People do: Build research workflows and personalization algorithms; set spending caps and policies
  - Results: Engineering workload down about 80% versus in-house stack; Compliance questions answered with one query; Managed infrastructure saved months
  - Tools: Amazon Bedrock AgentCore, Strands, Anthropic Claude, Amazon Bedrock Guardrails, Coinbase CDP
- [Freshworks Has Codex Check Claude's Work](https://ax.yumes.net/cases/freshworks-agentic-dev/): Cross-review between different models filters out bias and defects
  - Source: Freshworks Engineering Blog (Medium) (company's own publication) published 2026-05-28
  - Freshworks · Software & internet · scale: Freshdesk handles millions of customer requests
  - Functions: Software engineering
  - AI does: Orchestrator and specialist agents plan, write, test, and cross-review code under repository rules
  - People do: Define the what and why, design rules and skills, review final code
  - Results: Built a production-ready cleanup cron framework from one requirement; No quantitative figures published
  - Tools: Claude, OpenAI Codex, Cursor
- [Nippon Life Grades 8,830 Meeting Records](https://ax.yumes.net/cases/nippon-life-stewardship/): A prompt ten A4 pages long corrected the AI's habit of reading the company's own remarks too kindly
  - Source: Oliver Wyman client case study (Insights) + interview video (consulting client case) published 2026-05-28
  - Nippon Life · Insurance · Japan · scale: Dialogue with about 700 investee companies a year; each officer covers close to 100 companies
  - Functions: Data & analytics, Legal & compliance
  - AI does: Grades 8,830 engagement records and checks 10,000+ requests against disclosures and financials
  - People do: Define good-dialogue criteria, refine the prompt, run sample verification and governance checks
  - Results: 8,830 dialogue records from ten years analyzed consistently; Final grading prompt of about ten A4 pages; Share of High Quality dialogues grew year over year
  - Tools: Oliver Wyman
- [Why Gamma Turned Its Paused AI Support Back On](https://ax.yumes.net/cases/gamma-fin/): How 23 people handle 50 million users' questions while holding 84% satisfaction
  - Source: Intercom Fin customer story (vendor customer story)
  - Gamma · Software & internet · scale: 50M+ users; 23-person support staff (20 BPO agents, 2 support engineers, 1 community manager); 20,000-30,000 conversations a month
  - Functions: Customer support
  - AI does: Handles all support conversations in many languages, refunds and payment failures via Stripe
  - People do: Handle escalations by tier, payment, emotion; workflow strategy, AI training, content optimization
  - Results: Fin resolves 75% of conversations, 18,000+ a month; Human-handled share fell from 94% to 24%; CSAT rose from about 69% to 84%
  - Tools: Intercom Fin, Stripe
- [Rocket Money Cuts Inquiry Rerouting](https://ax.yumes.net/cases/rocket-money-fin/): An operating model that began with 10% of chats and came to resolve 68% of the chats Fin handled
  - Source: Intercom Fin customer story (vendor customer story)
  - Rocket Money · Financial services · scale: Millions of users; 60,000+ support conversations per month
  - Functions: Customer support
  - AI does: Resolves billing, app troubleshooting, and account access chats; routes high-risk inquiries to humans
  - People do: Handle sensitive escalations, review Fin failures, refine workflows and knowledge
  - Results: Fin involved in 54% of conversations, resolving 68% of them; Human-agent CSAT up 6 points; email billing CSAT 80%+; About $1 million annual ROI
  - Tools: Intercom Fin
- [Affirm's One-Week Pause Experiment](https://ax.yumes.net/cases/affirm-retooling-week/): Why it stopped all development and switched to AI agents
  - Source: Affirm Tech Blog (Medium) (company's own publication) published 2026-04-27
  - Affirm · Financial services · United States · scale: 800+ engineers; 130M+ financial transactions a year
  - Functions: Software engineering, Management & organization
  - AI does: Writes code for single planned tasks, runs checks, and fixes CI failures
  - People do: Decide what to build, approve plans, review code, approve merges
  - Results: 92% of engineering org submitted an agent-written PR; Agent-assisted PRs rose from near 0% to over 60%; Weekly merged PRs up 58% year over year
  - Tools: Claude Code, Anthropic Opus 4.5
- [KTern.AI Ships Agents in 4 to 6 Hours](https://ax.yumes.net/cases/ktern-bedrock-agentcore/): It runs on prompts and tool settings alone, with no orchestration code
  - Source: AWS Machine Learning Blog (co-authored by KTern.AI and AWS) (vendor customer story) published 2026-07-10
  - KTern.AI · Software & internet · scale: 20+ production agents across 50+ configurations
  - Functions: IT & operations, Software engineering
  - AI does: Analyzes SAP code and processes, generates migration code, test cases, exception reports
  - People do: Engineers configure prompts and business rules and build new domain features
  - Results: SAP project duration cut 45% on average; Infrastructure setup time cut 95%; First agent ships in 4–6 hours, down from 2–3 weeks
  - Tools: Amazon Bedrock AgentCore, Strands Agents SDK, Amazon Bedrock, Anthropic Claude, Amazon CloudWatch
- [Spotify Chose an Agent Over a 20,000-Line Script](https://ax.yumes.net/cases/spotify-agent-sdk/): Verifiers and an LLM judge blocked changes that went beyond the instructions
  - Source: Claude customer story + Spotify Engineering (Honk Part 1·3) (vendor customer story)
  - Spotify · Media & education · scale: Thousands of repositories; hundreds of engineers use the agent
  - Functions: Software engineering
  - AI does: Performs fleet-wide code migrations from prompts and opens merge-ready PRs
  - People do: Engineers write prompts and review agent PRs before merge
  - Results: Migration engineering time cut by up to 90%; 650+ agent PRs merged into production monthly; LLM judge vetoed about a quarter of sessions
  - Tools: Claude Agent SDK, Claude Code, Claude Sonnet 4.5, Honk (in-house)
- [Warp's 10-Minute Rule and a 40–50% Completion Rate](https://ax.yumes.net/cases/warp-firstround/): It banned outcome-only prompts and turned failures into evaluation assets
  - Source: First Round editorial (media or VC deep dive) published 2025-10-08
  - Warp · Software & internet
  - Functions: Software engineering, Management & organization
  - AI does: Writes code from engineers' prompts, running as parallel background agents
  - People do: Engineers specify implementation, steer small changes, and own code quality
  - Results: Estimated 40–50% of coding work finished by prompt; Reached after more than three months under the policy; One person manages at most 3 to 4 agents
  - Tools: Warp, Cursor, Claude Code
- [HubSpot Merges 7,000+ AI PRs Through Hook Controls](https://ax.yumes.net/cases/hubspot-crucible/): Builds cut to 2–3 minutes to give agents a fast feedback loop
  - Source: HubSpot Product Blog (engineering) (company's own publication) published 2026-01-23
  - HubSpot · Software & internet · scale: 1M+ builds a day; tens of thousands of microservices; ~3,000 EC2 instances
  - Functions: Software engineering
  - AI does: Plans issues, implements code, opens PRs, and reviews human-written PRs
  - People do: Engineers refine plans, assign issues to the agent, and review output
  - Results: 7,000+ AI-made PRs merged in six months; 50,000+ PRs reviewed by an agent; First build time cut from 10–15 to 2–3 minutes
  - Tools: Claude Code, Crucible (in-house), Sidekick (in-house)
- [Equipter: Reply Time Cut From 8 Hours to Minutes](https://ax.yumes.net/cases/equipter-salesforce/): An agent running in two weeks on a standard template revived sales opportunities
  - Source: Salesforce customer story (vendor customer story) published approx. 2026-03-18
  - Equipter · Manufacturing · scale: 6-person inside sales team; 150–200 inbound leads a week; 400+ rental partners
  - Functions: Sales & marketing
  - AI does: Sends personalized four-email sequences to social leads and dormant prospects
  - People do: Sales reps take over on reply and close deals
  - Results: Social lead response rate up 2.5x, 4% to about 10%; First contact time cut from 8 hours to minutes; About 2% of contacted leads became sales opportunities
  - Tools: Salesforce Agentforce, Sales Cloud
- [Wyndham Cuts 30-Day Standards Reviews by 94%](https://ax.yumes.net/cases/wyndham-pwc/): AI handles 28% of calls, and average call time drops 30–50%
  - Source: PwC case study (consulting client case)
  - Wyndham Hotels & Resorts · Travel & hospitality · scale: 25 brands in 95 countries; 9,300+ franchised hotels; hundreds of call center agents
  - Functions: Customer support, Back office & admin
  - AI does: Reviews standards change requests, answers franchise and guest queries, handles bookings
  - People do: Managers give final standards approval; specialists handle complex technical cases
  - Results: Brand standards review time cut 94%; AI directly handled 28% of incoming calls; Average handle time fell 30–50%
  - Tools: Salesforce Agentforce, Salesforce Experience Cloud, PwC, AWS
- [How 9 People Manage 50,000 Companies at Merge](https://ax.yumes.net/cases/merge-clay/): A three-layer automation that tracks 1,500 accounts and lifted meetings by 15%
  - Source: Clay customer story (vendor customer story)
  - Merge · Software & internet · scale: 9-person GTM team (4 RevOps, 5 marketing); 50,000+ accounts
  - Functions: Sales & marketing, Data & analytics
  - AI does: Tracks key accounts, drafts emails, analyzes call transcripts, captures event leads
  - People do: Reps follow up on opportunities; CSMs contact at-risk customers first
  - Results: Enterprise meetings booked up 15%; Customer retention and expansion up 10%; SDRs save 10+ hours a week; CSMs save 5
  - Tools: Clay, Gong, Salesforce
- [Shopify Opens Up 3,000 Cursor Licenses](https://ax.yumes.net/cases/shopify-firstround/): Support and sales adopted AI first, with no cap on tokens and a spending leaderboard
  - Source: First Round editorial (media or VC deep dive) published 2025-07-15
  - Shopify · Software & internet · scale: 3,000 Cursor licenses; intern target raised to 1,000
  - Functions: Sales & marketing, Customer support, Software engineering
  - AI does: Drafts weekly updates and RFP answers, diagnoses merchant sites, writes code
  - People do: Leads review and correct AI drafts, adding business risks by hand
  - Results: About half of AI weekly updates published unedited; GitHub Copilot usage reached 80%; RFP response agent raised the win rate
  - Tools: Cursor, GitHub Copilot, Claude Code, LibreChat, Roast, CodeAgent, internal LLM proxy
- [ZoomInfo Ports 938 Files in a Week](https://ax.yumes.net/cases/zoominfo-migration/): How 15 agents finished a framework switch that used to take six months
  - Source: ZoomInfo Engineering Blog (company's own publication) published 2026-04-08
  - ZoomInfo · Software & internet · scale: 938-file migration by 1 engineer directing 15 agents
  - Functions: Software engineering
  - AI does: Fifteen agents convert Angular to React, write tests, review code, audit parity
  - People do: Engineer designs context and guardrails, decides edge cases, does final review
  - Results: 938 files migrated in one week vs six months; Engineer spent only 8 to 12 hours hands-on; Concurrency capped at 15 agents on one 32GB machine
  - Tools: Claude Code Agent Teams, Claude Opus, Claude Sonnet, Playwright MCP
- [WHOOP Ends the 10-Hour Wait](https://ax.yumes.net/cases/whoop-fin/): Fin handles 84% of pre-purchase questions and removes the response lag
  - Source: Intercom Fin customer story (vendor customer story)
  - WHOOP · Manufacturing · scale: 3-person inside sales team
  - Functions: Sales & marketing, Customer support
  - AI does: Answers pre-purchase questions on the Join page around the clock
  - People do: Inside sales handles complex cases and high-potential buyers needing in-depth advice
  - Results: Resolution rate rose from 56% to 84%; Inside sales attributed revenue up about 130%; 68% resolved on launch day with no SLA violations
  - Tools: Intercom Fin, Fin Guidance
- [Jukebox: 98% of Cancelers Reorder in 2 Minutes](https://ax.yumes.net/cases/jukebox-fin/): Live support that keeps pace with same-day production, and instant cancellations
  - Source: Intercom Fin Customer Story (vendor customer story)
  - Jukebox · Retail & e-commerce · scale: Thousands of support conversations a month; orders up 80% in a year
  - Functions: Customer support, Sales & marketing
  - AI does: Answers order questions with live backend data and cancels orders instantly
  - People do: Handle print defects, reprint approvals, and quality control issues
  - Results: 98% of cancelers reordered within 2 minutes; Fin resolution rate up to 65%; Purchase conversion rate rose 40%
  - Tools: Intercom Fin, Fin Guidance, AI Insights, ActiveLabs
- [Clay Answers 20,000 Slack Members in 10 Seconds](https://ax.yumes.net/cases/clay-fin/): Expanding from in-app chat to email, it reached a resolution rate of up to 50%
  - Source: Intercom Fin Customer Story (vendor customer story)
  - Clay · Software & internet · scale: 3-person support team; 20,000-member Slack community; ~7,000 tickets a month
  - Functions: Customer support
  - AI does: Answers tier 1–3 support questions and processes refunds across chat, Slack, email
  - People do: Handle complex technical analysis; feed product gaps to docs and product teams
  - Results: Fin involved in 90% of conversations; Resolution rate rose from 20% to 45–50%; Answers within 10 seconds, down from 4–6 hours
  - Tools: Intercom Fin, Channeled
- [Checkout.com's HAL Writes 18% of Its PRs](https://ax.yumes.net/cases/checkout-hal/): How narrow, well-bounded tasks let an agent fill 18% of all change requests
  - Source: Checkout.com Engineering Blog (company's own publication) published 2026-06-17
  - Checkout.com · Financial services
  - Functions: Software engineering, IT & operations
  - AI does: Opens and maintains PRs, scores ticket readiness, fixes CI, patches vulnerable dependencies
  - People do: Named approvers review and approve every PR; no auto-merge
  - Results: HAL generates 18% of new PRs opened; Weeks-old dashboard filter ticket merged the same day; Security alert work cut from half a day to a short check
  - Tools: Agent HAL (in-house)
- [Money Forward Takes Coding Agents to Planning and QA](https://ax.yumes.net/cases/money-forward/): Faster development exposed bottlenecks in planning, design and QA, and it fixed them
  - Source: Cursor Customer Case Study (vendor customer story) published 2026-03-18
  - Money Forward · Financial services · scale: 1,000+ employees using coding agents daily
  - Functions: Software engineering
  - AI does: Writes code, designs test cases, generates Playwright tests, maps system architecture
  - People do: QA analyzes product risk; planners write specs; designers refine experiences
  - Results: Test generation time cut by up to 70%; Each engineer saved 15 to 20 hours a week; Agent users grew 30% in one week
  - Tools: Cursor
- [Generali GC&C Underwrites From the Inbox](https://ax.yumes.net/cases/generali-gcc/): AI that never leaves the inbox speeds up pre-bind underwriting by 80%
  - Source: Capgemini Client Story (consulting client case) published 2026
  - Generali GC&C · Insurance · scale: Operations in 25 countries; 1.9 TB of documents migrated
  - Functions: Back office & admin, IT & operations, Legal & compliance
  - AI does: Extracts risk data from broker documents and guides underwriters in natural language
  - People do: Underwriters assess risk, set terms, and decide whether to take risks
  - Results: Pre-bind underwriting 80% faster; Software license costs down 60%; IT support costs and service desk tickets each down 75%
  - Tools: Capgemini, Microsoft Dynamics 365, Microsoft Power Platform, SharePoint Online, Outlook add-in
- [Craft Docs: A Company-Wide Agent Built in Two Weeks](https://ax.yumes.net/cases/craft-docs/): Tickets take 2–3 minutes, and one engineer moved 60 services in a week
  - Source: The Pragmatic Engineer Deepdive (media or VC deep dive) published 2026-01-27
  - Craft Docs · Software & internet · scale: 20 engineers; 1M+ active users; 50,000+ paying customers
  - Functions: Customer support, Software engineering, Sales & marketing
  - AI does: Triages tickets, reads source code for bug reports, builds pages and integrations
  - People do: Architects review and merge PRs; staff build commands and set goals
  - Results: Ticket handling dropped from 20–30 minutes to 2–3; One engineer upgraded 60 services in one week; Weekly Linear issues rose from 15–20 to 100+
  - Tools: Craft Agents (in-house), Claude Agent SDK, Claude Code, Claude Max, GPT-3.5
- [Coinbase: Agents Now Write 75% of Its PRs](https://ax.yumes.net/cases/coinbase/): Idea to first PR in under 30 minutes, and release time cut to 1.8 days
  - Source: Cursor Customer Case Study (vendor customer story) published 2026-06-23
  - Coinbase · Financial services · scale: 2,400+ developers using Cursor
  - Functions: Software engineering, Management & organization
  - AI does: Writes code from engineer-approved plans and creates most pull requests
  - People do: Decide what to build, choose system structure, evaluate agent output
  - Results: Agents create 75% of all PRs; Idea to first PR: 8 days to under 30 minutes; Idea to production: 20 days to 1.8 days (some teams)
  - Tools: Cursor, Coinbase in-house Slack coding agent
- [Why Monzo Built Its Own Coding Agent in Three Weeks](https://ax.yumes.net/cases/monzo/): It handles 1,800 tasks a day and writes about 10% of merged pull requests
  - Source: Monzo engineering blog (company's own publication) published 2026-08-13
  - Monzo · Financial services · United Kingdom
  - Functions: Software engineering, IT & operations
  - AI does: Writes code and PRs, diagnoses incidents, upgrades dependencies, pre-reviews pull requests
  - People do: Design architecture and business logic; review and approve agent-written code
  - Results: Handles more than 1,800 automated tasks every day; Writes about 10% of merged pull requests; In-house agent MVP built in three weeks
  - Tools: Agent Chip (in-house)
- [How Three People Support 1,800 Staff at Remote](https://ax.yumes.net/cases/remote/): Ticket assignment done with an Okta identity lookup and a single emoji
  - Source: Zapier customer story (vendor customer story) published 2025
  - Remote · HR & recruiting · scale: 1,800+ employees; 3-person IT support team; about 1,100 requests a month
  - Functions: IT & operations
  - AI does: Classifies and prioritizes IT requests and proposes solutions from past resolution records
  - People do: Three IT staff take complex or approval-needing requests, assigned by emoji
  - Results: 27.5% of IT requests resolved with no human; 616 hours saved per month in IT support; $500,000 in hiring costs avoided
  - Tools: Zapier, Zapier Agents, Zapier Tables, ChatGPT
- [BioRender Ends Its Team's 45-Minute Morning Sort](https://ax.yumes.net/cases/biorender/): How a four-person team raised throughput 50% without new hires and resolved issues three days sooner
  - Source: Zapier customer story (vendor customer story) published 2025
  - BioRender · Software & internet · scale: Four-person accounts receivable team
  - Functions: Finance & accounting, Back office & admin
  - AI does: Classifies incoming accounts receivable tickets into nine types
  - People do: Resolve tickets; team lead takes sensitive items like remittance confirmations
  - Results: Ticket resolution time down 69%; Throughput up 50% with headcount held at four; 3,099 tickets routed with no human touch
  - Tools: Zapier, Gemini 2.0 Flash, Zapier Tables, Claude
- [Carta Cuts Cash Reconciliation from 11 Minutes to Seconds](https://ax.yumes.net/cases/carta-firstround/): Turning accountants' flowcharts into prompts saves more than 3,500 hours a month
  - Source: First Round Review (media or VC deep dive) published 2025-07-22
  - Carta · Software & internet · United States · scale: About 600 fund administration specialists managing 2,500 client funds
  - Functions: Finance & accounting
  - AI does: Reconciles bank transactions to the ledger using expert flowcharts converted into prompts
  - People do: Accountants draw and edit decision flowcharts and correct wrong agent steps
  - Results: Cash reconciliation from 11 minutes a case to seconds; More than 3,500 hours of manual work saved monthly
  - Tools: OpenAI, Lucidchart, pgvector
- [Block Declares a Shift from Hierarchy to an Intelligence System](https://ax.yumes.net/cases/block-sequoia/): Designing context with a world model and working in three roles
  - Source: Block official essay (Jack Dorsey & Roelof Botha) (company's own publication) published 2026-03-31
  - Block · Financial services
  - Functions: Management & organization
  - AI does: World model and intelligence layer carry company context and assemble customer offers
  - People do: Work as ICs, DRIs or player-coaches; make critical, ethical and cultural judgments
- [How Intercom Opened Up the Reasons Behind Its Lead Data](https://ax.yumes.net/cases/intercom-clay/): Showing why leads were picked grew outbound-sourced pipeline by 140%
  - Source: Clay customer story (vendor customer story) published 2026-09
  - Intercom · Software & internet
  - Functions: Sales & marketing, Data & analytics
  - AI does: Collects five signal types, enriches records, and attaches selection reasons to leads
  - People do: RevOps and managers tune filters live; reps use reasons to open conversations
  - Results: Outbound-sourced pipeline grew 140%; 4,000+ new target accounts found in a month; Prospect list building cut from days to 5 minutes
  - Tools: Clay
- [LendingTree's Loan Assistant](https://ax.yumes.net/cases/lendingtree-bedrock/): How it resolved more than 97% of complex loan conversations with no human
  - Source: AWS ML Blog Case Study (vendor customer story) published 2026-08-05
  - LendingTree · Financial services
  - Functions: Customer support, Sales & marketing
  - AI does: Supervisor plans; workers explain loan concepts and fetch real-time rates and prequalification
  - People do: Support queue takes conversations outside allowed limits or asking for a human
  - Results: More than 97% of conversations completed with no human; Only about 3% asked for a human agent; 50%+ of recent conversations show transaction intent
  - Tools: Amazon Bedrock, Amazon Nova Pro, Amazon Nova Lite, LangGraph, Model Context Protocol (MCP), Amazon Bedrock Knowledge Bases, Amazon Bedrock Guardrails, Amazon OpenSearch Service
- [Linktree: $520,000 a Year of Moderation for $20,000](https://ax.yumes.net/cases/linktree-first-round/): How a team of under 200 people supports 70 million users
  - Source: First Round Applied Intelligence (media or VC deep dive) published 2025-08-28
  - Linktree · Software & internet · scale: Under 200 people; 70 million users; 80,000 new sign-ups a week
  - Functions: Software engineering, Customer support, Management & organization
  - AI does: Devin does repetitive code tasks; Lorikeet resolves support; Gemini classifies content
  - People do: Senior engineers scope and hand off tasks; planners prototype with Lovable, Figma Make
  - Results: Moderation cost cut from $520,000 to $20,000 a year; AI resolves 50% of support inquiries alone; Under 200 people serve 70 million users
  - Tools: Cursor, Devin, Lorikeet, Gemini 2.0 Flash, Lovable, Figma Make, GitHub Copilot, Claude Code
- [Konecta: The 10-20-70 Rule and Kolibri](https://ax.yumes.net/cases/konecta-bcg/): How an organization redesign and a single platform cut handling time
  - Source: BCG & BCG X client case (consulting client case)
  - Konecta · Professional services · scale: Operations in 28 countries; 25,000+ employees using AI daily
  - Functions: Customer support, Sales & marketing, Management & organization
  - AI does: Runs 80+ generative AI deployments on Kolibri that handle and assist customer inquiries
  - People do: Hybrid agent managers oversee AI-human handoffs; agents handle exceptions
  - Results: Average handling time down up to 35%; Conversion rate up to 40%; 80+ generative AI deployments in 12 months
  - Tools: Kolibri (in-house), BCG X Deep Customer Engagement AI
- [Glide Cuts Deployment from 23 Hours to 14 Minutes](https://ax.yumes.net/cases/glide-slipstream/): How an agent-ready process cut development time by 60%
  - Source: Glide Engineering Blog (company's own publication) published 2026-08-06
  - Glide · Software & internet · scale: 10 non-engineers made up 25% of headcount; engineers about 30%
  - Functions: Software engineering
  - AI does: Writes code, runs E2E tests, ships behind flags, monitors, repairs and rolls back
  - People do: Approve whether features do what was asked and whether fixes are safe
  - Results: Merge-to-production median from 23 hours to 14 minutes; PR cycle time (p75) from 21 hours to 5.4 hours; Agent-reviewed PR share from 0% to 90%
  - Tools: Slipstream (in-house)
- [Neo: An Overnight Migration and 20,000+ Resolutions a Month](https://ax.yumes.net/cases/neo-financial-intercom/): How linking Fin to financial systems let it solve 71% of chats on its own
  - Source: Intercom Fin customer story (vendor customer story)
  - Neo Financial · Financial services · Canada · scale: About 20-person support team, about 50,000 inquiries a month
  - Functions: Customer support
  - AI does: Looks up account status in core banking and explains declined or held transactions
  - People do: Agents take complex or legally required cases with full conversation context
  - Results: Fin resolves 71% of all chats with no agent; 20,000+ conversations closed monthly without human help; CSAT rose from about 50% to 86%
  - Tools: Intercom Fin, Fin Voice
- [Brex Cuts Card Dispute Handling from 3 Hours to 3 Seconds](https://ax.yumes.net/cases/brex-first-round/): Operations automation built on a shared tool standard and a three-level role redesign
  - Source: First Round Applied Intelligence (media or VC deep dive) published 2025-09-25
  - Brex · Financial services · scale: 1,000+ employees; about 350 in EPD; about 25-person systems engineering team
  - Functions: Back office & admin, Customer support, Management & organization
  - AI does: Drafts card dispute documents, runs KYC steps, scores 100% of support interactions
  - People do: L2 managers oversee agents and handle exceptions; L3 experts set policies
  - Results: Card dispute documents from about 3 hours to 3 seconds; 50%+ of support questions solved in first reply; Negative-news screening accuracy 88% versus 85% for humans
  - Tools: Retool, Model Context Protocol (MCP), Cursor, NotebookLM
- [bunq Drops the Central Router for 3 to 5 Primary Agents](https://ax.yumes.net/cases/bunq-bedrock/): Linking specialist agents as tools to handle 97% of customer support
  - Source: AWS Machine Learning Blog (vendor customer story) published 2026-01-21
  - bunq · Financial services · scale: 20 million users across Europe; 80-person cross-functional team
  - Functions: Customer support
  - AI does: Finn answers support requests; primary agents call specialist agents as tools
  - People do: Handle cases not fully automated; support staff and engineers refine agents daily
  - Results: Finn handles 97% of customer support; 70 to 82% fully automated with no human agent; Average response time 47 seconds
  - Tools: Amazon Bedrock, Anthropic Claude, Amazon SageMaker, Amazon OpenSearch Serverless, Amazon DynamoDB, Amazon ECS
- [MerciYanis Hands Its Backlog to AI Agents](https://ax.yumes.net/cases/merciyanis-agents/): How it reached a merge rate of about 95% by giving agents the backlog
  - Source: MerciYanis Engineering Blog (company's own publication) published 2026-04-20
  - MerciYanis · Software & internet · France · scale: 25-person company
  - Functions: Software engineering, Sales & marketing
  - AI does: Works real backlog tickets through a Handbook-driven /fix command and opens PRs
  - People do: Review PRs, give feedback, and decide which prototypes ship
  - Results: About 95% merge rate on agent tickets; Up to 32 times faster than junior engineer estimates; 14 tickets in 220 minutes for €84.05 in tokens
  - Tools: Claude Code, Opus 4.5, GPT-5.2
- [Kavak Gives Up on the Copilot](https://ax.yumes.net/cases/kavak-a16z/): How it put an AI agent in the middle of 90 to 95% of customer interactions
  - Source: a16z Podcast (media or VC deep dive) published 2026-02-18
  - Kavak · Retail & e-commerce · scale: About 10,000 employees at peak, now about 3,500; about 10,000 deals a month
  - Functions: Customer support, Back office & admin, Management & organization
  - AI does: Handles loan document review, first consultations and progress checks in customer funnels
  - People do: Mediate complex disputes and rule on non-standard warranty repair exceptions
  - Results: AI agent in 90 to 95% of customer interactions; Agent performance passed 1.5 times the best employees; Headcount fell from about 10,000 to about 3,500
- [Faire: One Engineer and a Swarm Finish 18 Months of Work](https://ax.yumes.net/cases/faire-cursor/): Dropping homegrown infrastructure for cloud agents, and the productivity it brought
  - Source: Cursor Customer Stories (vendor customer story) published 2026-05-26
  - Faire · Retail & e-commerce
  - Functions: Software engineering
  - AI does: Cloud agents run builds and tests, fix CI failures, triage bugs, migrate code
  - People do: One engineer orchestrates the agent swarm; engineers plan, scope and design work
  - Results: Weekly PR throughput doubled; 18-month MobX migration finished by one engineer with agents; 25+ automations launch agents 2,000+ times a week
  - Tools: Cursor Cloud Agents, Cursor Automations, Cursor plan mode, Swarm (in-house)
- [Personio Turns a 400-Person Sales Team into an AI Team](https://ax.yumes.net/cases/personio-gtm/): How it cut customer research to 15 minutes and doubled pipeline per rep
  - Source: SaaStr (Philip Lacor CRO) (media or VC deep dive) published 2026-01-07
  - Personio · Software & internet · scale: 1,500 employees, 15,000 customers, 400-person sales organization
  - Functions: Sales & marketing
  - AI does: Summarizes 10-20 systems into cross-sell calls, analyzes win/loss, books demos via Nia
  - People do: Dedicated manager reviews Nia conversations daily; reps act on recommendations
  - Results: Pre-call research cut from 2 hours a day to 15 minutes; Pipeline per Expansion SDR grew about 2 times; Nia booked 140 demo meetings in 7 days
  - Tools: Amazon Bedrock, Gong, Qualified, Salesforce, Snowflake
- [monday.com's AI Agents at Work](https://ax.yumes.net/cases/monday-bedrock/): How AI merges 19 of 20 pull requests with no human involved
  - Source: AWS ML Blog (vendor customer story) published 2026-07-22
  - monday.com · Software & internet · scale: Millions of paying users, hundreds of microservices, hundreds of builders
  - Functions: Software engineering
  - AI does: Agents take backlog tickets, write code, open PRs; Morphex auto-merges passing PRs
  - People do: Assign work to agents and review PRs that fall below the confidence bar
  - Results: 19 of 20 Morphex PRs auto-merge with no human; PR throughput per engineer up more than 50%; About 9 in 10 builders use AI tools monthly
  - Tools: Claude Agent SDK, Amazon Bedrock, Claude Code, Cursor, Model Context Protocol (MCP), Amazon EKS, Amazon EFS
- [How Vercel Put AI into Its Sales Team](https://ax.yumes.net/cases/vercel-gtm-agent/): How Vercel cut inbound sales from ten people to one and kept conversion
  - Source: SaaStr (primary) / drew.tech / Tunguz (company's own publication) published 2026-07
  - Vercel · Software & internet · scale: About 10 inbound SDRs before, 1.25 after
  - Functions: Sales & marketing
  - AI does: Enriches and researches inbound leads, sorts them into four categories, drafts replies
  - People do: One SDR reviews drafts in Slack and approves sending with one click
  - Results: Inbound SDR staffing cut from 10 to 1.25 people; Lead-to-opportunity conversion rate held steady; 32x ROI; about $5,000 a year in infrastructure and tokens
  - Tools: AI SDK, Workflow DevKit, Exa, Clearbit, ZoomInfo, Slack Bolt, Outreach, Vertex AI
- [Paychex Puts AI Inside Its Payroll Engine](https://ax.yumes.net/cases/paychex-bain/): Shadow-mode testing cut payroll turnaround time by 80%
  - Source: Bain client results (consulting client case) published 2026-05-11
  - Paychex · HR & recruiting · United States · scale: 800,000+ business clients, about 1 million payroll transactions a month
  - Functions: Back office & admin, Finance & accounting
  - AI does: Reads emailed payroll changes, checks them, and runs standard payroll transactions end to end
  - People do: Review ambiguous requests and complex exceptions; give customers advisory work
  - Results: Payroll turnaround time fell 80%; Human effort fell 30%; WISE: 20%+ of 350,000+ agent conversations resolved without humans
  - Tools: OpenAI, Paychex WISE

## AI Natives

- [Crescendo Bought a 3,000-Person BPO](https://ax.yumes.net/natives/crescendo/): Crescendo charges per resolved case, not per hour (2026-10-22)
  - Crescendo · Professional services · scale: About 20-person startup that bought PartnerHero's about 3,000 support agents and 200+ customers
  - Functions: Customer support
  - AI does: Handles first contact across channels, drafts agent replies, scores every conversation, proposes fixes
  - People do: Handle complex cases, approve system changes, deploy AI, respond to AI incidents
  - Results: Every conversation scored versus 1-3% sampled at traditional centers; Pricing from $2.99 per resolution (October 2025); One customer: AI resolved 70%, specialists 30%
  - Tools: Concierge, Expert Assist, Quality Agent
- [Periodic Labs: Automating the Reading of Results First](https://ax.yumes.net/natives/periodic-labs/): How lab results become training data for the company's own model (2026-10-22)
  - Periodic Labs · AI research · United States
  - Functions: R&D & science, Data & analytics
  - AI does: Periodic Neon analyzes XRD and microscope results; AI judge scores the model.
  - People do: Technicians record actual bench work; scientists supervise experiments and co-set research direction.
  - Results: AI judges agree with experts 74.6%, with consensus 84%; Human experts agreed with each other 77.2% of the time; 3.8 times higher success rate in Periodic's own environment
  - Tools: Periodic Neon, Kimi K2.6
- [Jooba: One Dedicated AI Agent for Every Role](https://ax.yumes.net/natives/jooba/): How per-role AI agents and people split the work of recruiting (2026-10-22)
  - Jooba · HR & recruiting · scale: 6 team members; helped 100+ companies hire
  - Functions: HR & recruiting
  - AI does: Custom per-role agents search millions of candidates and build candidate pipelines.
  - People do: Talent engineers run intake calls, review and approve pipelines; clients approve intros.
  - Results: Median 22 hours to first candidate introduction in 2026; Helped more than 100 companies hire; Claims eightfold lower cost per hire, 52% faster hiring
- [A CSO That Never Touches Data: Stanford's Experiment](https://ax.yumes.net/natives/stanford-virtual-biotech/): A study that models a drug R&D organization on 11 AI agents (2026-10-22)
  - Stanford University (James Zou lab, Virtual Biotech) · Healthcare & life sciences · scale: 11 agents; clinical trialist agent copied 37,075 times
  - Functions: R&D & science, Data & analytics
  - AI does: CSO agent plans and assigns; specialist agents analyze data; reviewer agent checks work.
  - People do: Pose the question, answer CSO clarifications, request follow-up analyses, audit final output.
  - Results: B7-H3 lung cancer analysis in under a day for $46; 1,839-hour clinical trial job finished in 6 hours; 89.7% agreement with human reviewers on primary endpoints
  - Tools: Claude Agent SDK, Claude Sonnet 4.5, Claude Haiku 4.5, Model Context Protocol (MCP)
- [Prism: Listing Each Candidate's Weaknesses Too](https://ax.yumes.net/natives/prism/): How Prism finds hiring criteria in past hires and candidate feedback (2026-10-22)
  - Prism · HR & recruiting · scale: Three- or four-person team
  - Functions: HR & recruiting
  - AI does: Infers hiring bar, runs parallel searches, sends tailored outreach, ranks shortlisted candidates.
  - People do: Staff run recruiting operations and client relationships; clients interview and make offers.
  - Results: Benchmark score 89.6 versus 68.2 best published product; Up to 50 ranked candidates per role; Fee of 15% of first-year salary per hire
- [Lila: Scientist Review Shrinks to a Safety Check](https://ax.yumes.net/natives/lila-sciences/): How Lila divides experiments among models, instruments, and people (2026-10-22)
  - Lila Sciences · AI research · scale: 22 employees on catalyst campaign; about 110 job postings
  - Functions: R&D & science
  - AI does: LLM over statistical models picks search strategy, catalysts to synthesize, and test conditions.
  - People do: Carry sample fixtures, check synthesis safety, verify surprising results, run instrument shifts.
  - Results: Tested 2,942 catalysts across 26 elements, 53 material systems; Second campaign took four weeks versus four months for first; Lead catalyst ran over 1,000 hours in strong acid
  - Tools: Lila Iris, AI Science Factory (AISF)
- [Norm Law: The Client Layer That Stays When Lawyers Leave](https://ax.yumes.net/natives/norm-law/): How Norm Ai builds the agents and Norm Law supervises them (2026-10-22)
  - Norm Law (with Norm Ai) · Legal · United States · scale: 35+ legal engineers (also stated as 30+), all former lawyers
  - Functions: Legal & compliance
  - AI does: Agents built from partner knowledge and client layers draft initial legal work.
  - People do: Legal engineers build and peer-review agents; licensed lawyers supervise and sign off.
  - Results: Agents complete first drafts; lawyers supervise and review outputs; Prices by outcome per deal instead of hourly billing; Client layer keeps client knowledge when lawyers leave
  - Tools: Norm Ai, LEAP (Legal Engineering Automation Platform)
- [Knowlee: Agents Make the Cards, Only People Move Them](https://ax.yumes.net/natives/knowlee/): How Knowlee works with agents and a review queue (2026-10-22)
  - Knowlee · Software & internet · Italy · scale: Founder-led small team (described as one-person company) running eight agents
  - Functions: Sales & marketing, IT & operations, Management & organization
  - AI does: Eight agents research prospects, draft content, coordinate status, monitor markets, route inquiries.
  - People do: Operator reviews queue, moves cards to Running, handles legal and strategic judgments.
  - Results: Eight agents claimed to replace a team of twelve to fifteen; Daily review queue takes 30 to 45 minutes; Agent costs below one junior employee's pay
  - Tools: Knowlee OS, Company Brain
- [Edison's Kosmos: 57.9% of Interpretations Held Up](https://ax.yumes.net/natives/edison-kosmos/): How Kosmos divides research work, and why people focus on verification (2026-10-22)
  - Edison Scientific · AI research · scale: About 30,000 users tried Kosmos; $70 million seed round
  - Functions: R&D & science, Data & analytics, Legal & compliance
  - AI does: Runs parallel literature search and data analysis, writes cited reports, drafts regulatory documents
  - People do: Set objectives, prepare data, steer mid-run, verify interpretations, make final decisions
  - Results: 79.4% of statements supported; interpretation only 57.9%; One run reads about 1,500 papers, runs 42,000 code lines; Kosmos wrote about 70% of a Pre-IND meeting request
  - Tools: Kosmos, Robin
- [Harper: Agents Read the Paperwork, People Close the Deal](https://ax.yumes.net/natives/harper/): How Harper divides the work of insurance brokerage between agents and people (2026-10-22)
  - Harper · Insurance · United States · scale: 25 employees on YC profile; about 1,000 new customers a month; 160+ insurers connected
  - Functions: Sales & marketing, Back office & admin, Customer support
  - AI does: Reads applications, routes submissions, follows up with underwriters, drafts follow-ups, supports intake calls
  - People do: Intake specialists verify facts; licensed account executives advise and close; veterans write evals
  - Results: Operations 90%+ AI-led, per company; 5,000+ businesses supported in first 13 months; Insurance in 24-48 hours, per company claim
- [CREAO: A Company of About 25, Coordinated by Agents](https://ax.yumes.net/natives/creao/): How agents coordinate and people decide, with no middle managers (2026-10-22)
  - CREAO · Software & internet · scale: About 25 employees, 10 engineers; no middle management
  - Functions: Software engineering, Management & organization, Sales & marketing
  - AI does: Proposes tasks, writes 99% of production code, reviews PRs, scores live output, coordinates
  - People do: Architects choose work; people check strategic risk, confirm diagnoses, approve publication
  - Results: AI writes 99% of production code, per company; Three to eight production deploys a day for 14 days; CTO's people-management time fell from 60% to below 10%
  - Tools: Super Agent, Anthropic, OpenAI, Google
- [Sakana's Manual Audit Caught What the Evaluator Missed](https://ax.yumes.net/natives/sakana-agentic-scientist/): How Sakana redesigned the division of labor between scientists and agents (2026-10-22)
  - Sakana AI · AI research · Japan
  - Functions: R&D & science
  - AI does: Paired agents form hypotheses, run experiments and ablations, and draft evidence-mapped papers.
  - People do: Write task specs, evaluators, and rubrics; audit outputs; edit and verify papers.
  - Results: Error Diffusion paper accepted as ALIFE 2026 oral presentation; People added about 17.5%, deleted 10.7% before submission; Manual audit caught shortcuts missed by evaluator and rubrics
  - Tools: The Agentic Scientist, Claude Code, Codex, OpenCode, Pi
- [Crosby: Training Agents on Lawyers' Edits](https://ax.yumes.net/natives/crosby/): How Crosby splits contract review with agents and charges per contract (2026-10-22)
  - Crosby · Legal · United States · scale: About 45 lawyers and about 40 engineering and operations staff (June 2026); 50+ attorneys listed
  - Functions: Legal & compliance
  - AI does: Routes requests, summarizes contracts, proposes redlines with explanatory comments
  - People do: Lawyers review every output, edit redlines, write prompts, compare redlines for evals
  - Results: Lawyer review time down 74% on one client's NDAs; About 80% less review time on another client's MSAs; Verifier trained on about 13,000 labeled redlines
  - Tools: Crosbrain, Verifier
- [Cognition Starts Work at the Pull Request](https://ax.yumes.net/natives/cognition/): How a team builds its own software with Devin, its coding agent (2026-10-22)
  - Cognition · Software & internet · scale: 350 employees after Windsurf deal (44 before); 15-person engineering team in May 2025
  - Functions: Software engineering
  - AI does: Devin triages bugs, writes fixes and tests, opens and reviews pull requests, audits merges
  - People do: Review pull requests, judge changes, decide what ships, handle visual design
  - Results: 89% of engineer-committed code committed by Devin (May 2026); 659 Devin pull requests merged in one week; $1 billion annualized revenue run rate
  - Tools: Devin, Devin Review, Windsurf, Sentry, Datadog, Linear
- [Recursive Built a Model in Under Four Months](https://ax.yumes.net/natives/recursive-flint/): How an AI research loop built a model, and lessons for R&D teams (2026-10-22)
  - Recursive Superintelligence (Recursive) · AI research · scale: About 20 staff as of April 2026
  - Functions: R&D & science, Software engineering
  - AI does: Eureka Machine proposes ideas, builds, runs experiments, verifies; designed and quantized Flint.
  - People do: Set goals, build checks against reward hacking, review results by hand.
  - Results: Built Flint from scratch in under four months; 4-bit model stayed within 0.7 points of 16-bit model; Flint runs over 125 tokens per second on M4 Pro
  - Tools: Eureka Machine

## Research

- [AI Infrastructure Capital Flows and Risk Transmission](https://ax.yumes.net/research/ai-economy-money-flow/): Debt-funded investment faces a 41% drop in effective token prices. (2026-10-26)
  - Topics: AI equity valuation gains; Debt-financed AI infrastructure capex; Off-balance-sheet lease commitments; Take-or-pay risk on model developers; Token price deflation versus demand; Loss transmission scenarios through 2030
- [The Three Pillars of Organizational Transformation in the Agentic AI Era](https://ax.yumes.net/research/agentic-organization/): Redefining Frontline Roles, Managerial Leadership and Performance Metrics (KPIs), and Enterprise Operating System Shifts (2026-10-02)
  - Topics: Agentic orchestration operating model; Frontline workers as agent orchestrators; Seniorized entry-level roles; Managers governing human-agent teams; KPI redesign after tokenmaxxing; Agentic factories and small pods

## Essays

- [Who Gets Paid When Software Is Cheap](https://ax.yumes.net/essays/developers-after-cheap-software/): More people build software than ever, and the money is moving to a few. (2026-11-05)
  - Topics: Cheap code, concentrated earnings; Spec-to-code developers losing ground; Domain experts choosing software; Entry-level developer hiring decline; Developers tied to customer metrics
- [Surviving the Era of Zero Junior Developer Hiring](https://ax.yumes.net/essays/cs-student-2029/): Prove that AI runs inside a real organization, not just that you can code. (2026-10-29)
  - Topics: Junior developer hiring collapse; AI task horizon extrapolation; Capability versus deployment gap; Forward deployed engineers; Student residency in small organizations; AI evaluation and error analysis
- [The Rise of AI-Native Startups](https://ax.yumes.net/essays/ai-native-startups/): The shift from selling tools to becoming the industry itself. (2026-10-22)
  - Topics: AI natives doing the work; Service as software; Outcome pricing versus billable hours; Incumbent internal resistance; Licenses and accountability as moats; Mobile-era O2O comparison
- [Incompetent Manager, Incompetent Human](https://ax.yumes.net/essays/incompetent-manager/): You've handed your work to AI. What kind of boss are you? (2026-10-07)
  - Topics: Delegating work to AI agents; AI-generated workslop; Verifying AI output; Accountability for delegated work; Cognitive offloading and ownership
- [A Day in the Life of a Driver Assigned 100 Buses](https://ax.yumes.net/essays/remote-bus-driver/): The unseen strain, responsibility, and fatigue left for frontline workers when AI takes over daily routines. (2026-10-02)
  - Topics: Human monitoring of automation; Ironies of automation; Moral crumple zone; Vigilance limits and fatigue; Designing human-AI handoffs