AI results at work
An architect guided Claude Opus 5.5 to permit-quality plans for a simple house. Microsoft-backed research found 6 of 202 big US firms had four key agent practices. A Manchester report cites AI lifting coding about 240%, but releases only about 30%.
The results of AI at work inside organizations, such as time saved, output, quality and cost per result, from company reports and wider studies.
November 2026
9–15 November 2026
14 stories in the AI newsletters I read.
The facts
Airbnb said its AI assistant cut support cost per booking by more than 17%.1
Shopify said a shared generative model for search and recommendations lifted its conversion rate 9.85%.2
2–8 November 2026
4 stories in the AI newsletters I read.
The facts
An architect guided Claude Opus 5.5 to permit-quality plans for a simple house.1
Cisco aims to refactor 98% of its legacy code in 24 months with unlimited AI tokens.2
The stories
- A study reported by CNET finds writers who only approved AI decisions produced worse work than those who stayed hands-on.
- Snap says in its filing that engineering teams more than doubled merged pull requests year over year while defects per merged pull request fell over 71%.
- Airbnb's AI assistant cuts support cost per booking by more than 17% and resolves about 40% of Q3 issues with no human agent.
- Customers Bancorp says it doubled its AI agents in the quarter and puts hours saved at 76,000, up nearly 65%, worth over $15 million of annualized capacity.
- Santander says data and AI work made it €194 million this year and targets more than €1 billion from 2025 to 2028.
- ZipRecruiter says its employer response rate per application tripled from a year earlier, and its R&D spend fell from 27% to 20% of revenue on reduced headcount.
- Etsy says in its shareholder letter that personalization improvements alone contributed approximately two percentage points to its annual sales growth rate.
- PennyMac says automation enables operating leverage with reduced variable hiring; its AI voice agent handled more than 15,000 transfers and 1,600 rate locks in Q3.
- Cognizant reports over 50% of its code is now AI-assisted, alongside rising revenue per employee.
- MVB Bank says it has deployed 16 Digis (digital workers) and plans to reach 36 across credit, anti-money-laundering and fintech work.
- CareCloud says AI cut manual claim touches by 44% in its main division, letting teams handle more volume without adding headcount.
- Groupon expects its 2026 restructuring to save $20 million to $25 million a year and moves staff onto FoundryAI; it relaunches LivingSocial as an agent-first marketplace.
- Cloudflare describes a hybrid agent design where fast decision models pick tools and a large reasoning model only diagnoses, cutting six LLM calls to one and about 40K tokens to 1.8K.
- Shopify deploys a shared generative discovery model for search and recommendation using semantic-ID tokens, which lifts conversion rate 9.85%.
- Perplexity says its Computer agent raised Get Maine Lobster's return on ad spend by 63%.
- David Fowler reports a one-day run produced about 90 performance PRs to ASP.NET Core using GitHub's new cloud agents and Opus 5.5, for about $230 in tokens.
- Stephen Gallagher tests 13 models on house blueprints; Claude Opus 5.5 produces a permit-quality set in 12 hours for a few hundred dollars.
- Cisco's Jeetu Patel gives a 27-year-old unlimited AI tokens to refactor 98% of its legacy code in two years.
- An EAPJ article cites Gartner's 2027 CEO Survey: 44% of CEOs struggle to show AI financial returns, 75% plan more technology investment, 63% call radical operating-model redesign essential.
October 2026
26 October – 1 November 2026
8 stories in the AI newsletters I read.
The facts
Microsoft-backed research found 6 of 202 big US firms had four key agent practices.1
A Manchester report cites AI lifting coding about 240%, but releases only about 30%.2
The facts
The New York Times reports a poll of nearly 6,000 senior executives: about 90 percent saw no AI impact on jobs or productivity, while 69 percent said their firms use AI.1
Anthropic says Claude writes about 90% of Claude Code's code. Merged PRs more than doubled from June to September, and the share reverted fell over 40%. It credits a change in code review.2
Google says that in its Antigravity coding platform, its engineers using Gemini 4 Argon made over 35% more agentic code submissions, with rollbacks cut by nearly a third.3
The stories
- The University of Manchester's Productivity Institute, in a report commissioned by Bytes Technology Group, says up-front costs and task-versus-job gaps explain missing AI productivity gains.
- Gartner cites 86% of CIOs saying AI risk grows faster than value, four times faster on average.
- McKinsey says in its CPO survey that 80% plan to raise procurement AI investment, but only 2% report AI embedded at scale and 55% remain in pilot mode.
- Anthropic says the Messages API now runs fully on Rust, mostly written by Claude, with roughly 40-50% CPU savings and 10-20% lower time to first token.
- Microsoft and Keystone publish a Harvard Business School working paper finding only 6 of 202 surveyed US enterprises (3%) have all four practices connecting AI operations.
- McKinsey publishes lessons from a year of agentic AI in financial-crime prevention, saying a source-of-wealth check that can take up to 35 hours manually runs in about half the time.
- McKinsey reports more than 40% less human processing time in agentic transaction-monitoring investigations, with Commerzbank's Hans-Georg Beyer interviewed on a proof of concept.
- Latent Space's Kill My SaaS draws 72 submissions to replace a conference speaker-management tool quoted at $43,000 a year; the replacement cost about $1,000.
- Gartner research on the 2027 CIO and Technology Executive Agenda, as cited by EAPJ, finds only 13% of CIOs report significant or transformational value from AI.
- Anthropic says Claude writes about 90% of Claude Code's code; merged PRs doubled from June to September while the reverted share fell more than 40%.
- Meta says medical experts used a Muse scaffold over 5M licensed, de-identified EHRs to produce 10 manuscripts in two to four weeks.
- Google says Argon helped separate core webpage handling from a six-million-line legacy Chromium folder, cutting months of work to weeks; it reports 35% more agentic code submissions (internal claims).
- Anthropic reports daily merges growing from about 2,500 in May to more than 30,000 in September, then migrating its monorepo off Git in five months.
- Tom's Hardware reports AI-assisted patches cut Linux kernel build times from 22 seconds to 15.
- The New York Times reports workers spend 37% of AI's time savings fixing its output.
- Coinbase says a Qwen3.5-9B fine-tune for fraud prevention beats frontier models at lower cost and latency.
- Samaya says frontier AI models out-forecast human analysts on earnings prediction.
- Epoch launches Automation Reports testing models on its own open-ended work; Claude Fable 5.1 and GPT-6 Astra lead, and neither comes close to full automation.
- Airbnb's CTO says on a podcast that 60% of its code is AI-authored; Q2 results say AI resolves nearly 45% of support tickets.
September 2026
Fred Schott says agents now triage Astro's issues, and Endura's CEO says AI agents screened about 500 disease targets.
4 stories in the AI newsletters I read.
The facts
Fred Schott says agents now triage Astro's issues.1
Endura's CEO says AI agents screened about 500 disease targets.2
All 4 stories
- Endura Therapeutics' CEO says LLM research agents triaged about 500 disease targets, work he puts at about a person-year at the old rate.
- A post on X claims AI-enabled drugs reach first-in-human trials up to 80% faster.
- Astro creator Fred Schott says agents now handle issue triage, reproduction and fix checks, and the team has regained control of its backlog.
- Cognichip launches ACI Enterprise, a chip design copilot; in one reported run, one engineer did in 10 days what a team takes 4-5 months.
August 2026
In one survey, only 30% of marketers report significant, measurable results from AI.
2 stories in the AI newsletters I read.
The facts
In one survey, only 30% of marketers report significant, measurable results from AI.1
July 2026
Netflix says about 300 titles used generative AI. A study of one company's AI mandate found pull requests per engineer doubled.
4 stories in the AI newsletters I read.
The facts
Netflix says about 300 titles on its service used generative AI.1
A study of one company's AI mandate found pull requests per engineer doubled.2
All 4 stories
- A Danish study finds AI saves about 2.8% of total work time on average, without automatic business value.
- Netflix says roughly 300 titles have used generative AI, including 17 minutes of documentary footage made twice as fast at half the cost.
- A study of nearly 200,000 pull requests finds AI nearly doubled coding output while human review share fell from 89% to 68%.
- OpenAI launches Presence, an enterprise agent platform, and says its voice agents already resolve 75% of its own support line.
June 2026
Anthropic says Claude now writes over 80% of its merged code.
6 stories in the AI newsletters I read.
The facts
Anthropic says Claude now writes over 80% of its merged code.1
All 6 stories
- Vendor Glean's survey says 87% of digital workers use AI and save about 11 hours a week, but spend 6.4 hours checking it.
- Anthropic says Claude writes over 80% of its merged code and internal task success rose from 26% to 76% in six months.
- Palantir's AI Sepsis Hub at Tampa General halves sepsis deaths, saving nearly 900 lives.
- NHS England is giving 500,000 staff Copilot access, which it expects to free up 43 minutes a day.
- Stripe reportedly used Fable 5 to run a 50-million-line Ruby migration in one day, work that would take a team two months.
- Recidiviz ties its AI dashboards across 19 state prisons to a 16% drop in recidivism.
May 2026
The stories
October 2025
September 2025
The board, once a week, by email.
It comes every Monday.
By subscribing, you agree to get the board by email.See the Privacy Notice.