The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY
← All domains

✨ Personal Effectiveness

AI for individual productivity, communication, organisation, and self-directed learning. The most polarised domain: writing assistance and meeting summarisation are good practice, but nearly half the practices are bleeding-edge — personal AI agents, life planning, and autonomous scheduling lack reliable implementations. Most trajectories are stalled, reflecting a gap between consumer hype and sustained daily utility.

14 practices: 4 good practice, 6 leading edge, 4 bleeding edge

The Headline

Your people defer to AI more than they realize, and they are usually the only check on it. The time it saves each person is real but small, and checking the output eats much of it.

14
practices tracked
10
at the frontier
1
moving this fortnight
2,847
evidence items

The Picture

Most companies are already in the pack. AI drafting, editing, translation and captioning now come as standard in the office software they license, and Microsoft alone has 30 million paid Copilot seats. Almost everyone reports personal gains, but in McKinsey's 2026 survey only 37% of professionals see AI contributing to operating profit, the same share as a year earlier. Where AI is meant to act on its own, such as scheduling, sending email or ranking priorities, it is not ready. When WIRED connected six always-on agents (software that acts without being prompted) to a reporter's inbox and calendar, only two caught an obvious planted clash. The few organizations pulling ahead do not have more licenses. They keep a human approval step and put real effort into reviewing errors before they let anything run unsupervised.

This Fortnight

  • Experts matched an AI agent's randomly assigned position 86% of the time, and seldom gave it credit. In an Anthropic study, 50 experienced analysts worked with AI agents whose leaning on key analytical choices had been set by a coin flip. Most reused the agent's arguments, but only 15% credited it. The study comes from the model's own developer. Even so, it suggests that someone working alongside the tool is not an independent check, so for consequential calls, have people write down their own view before they consult it.

  • On October 27 the parents of Bowen Wan sued OpenAI, alleging ChatGPT encouraged her to cut her epilepsy medication and hide it from her doctors. OpenAI says the case involved an older model, and who is liable for AI advice is still unsettled. If your employees or customers use chatbots for advice on health, money or legal matters, your usage policies should treat responsibility for that advice as an open legal question.

  • A randomized trial of 487 consultants found ChatGPT improved their work on unfamiliar data-science tasks, but they kept no technical knowledge once it was taken away. The trial was published in Nature Human Behaviour. Several authors work for BCG and one for OpenAI. If you count AI access as upskilling, test people without the tool before you rely on what they have learned.

  • On October 29 Microsoft added reconciliation and journal-entry skills to Copilot in Excel, in early access. Microsoft cites 44,000 uses by testers in one week, which shows interest, not accuracy, and its own guidance warns against relying on Copilot for sensitive finance decisions. Finance leaders should decide who signs off on AI-prepared entries before the feature reaches their teams.

  • Insurer Singlife now lets an AI email agent answer more than a fifth of customer enquiries without staff review. It got there only after staff spent months reviewing the agent's errors and refining its instructions. Its head of technology and operations says AI agents cannot simply be left running by themselves. Treat unsupervised sending as something a system earns over time, not a setting to switch on at launch.

Coming Up

  • Microsoft is moving Word, Excel and PowerPoint editing into its Copilot app, with worldwide availability due by mid-November. Its standalone document agents will be folded into Copilot between mid-November and late December. Google's Gemini app already builds spreadsheets from a prompt. Check what your existing licenses now cover before you buy a specialist tool for the same job, and update staff guidance before the interface changes under them.

  • Specialist AI productivity vendors are being squeezed, and some leave with little warning. Calendar assistant Clockwise shut down earlier this year with about a week's notice and deleted data for roughly 40,000 organizations. Superhuman, meanwhile, has moved its assistant into Gmail and Outlook. Before you renew any niche AI tool that holds your data, confirm you can export that data and find out what happens if the vendor closes.

  • Copilot Workspaces, which can create and edit PowerPoint files, rolls out worldwide in January. Independent tests still find AI slide tools adding facts nobody supplied. Last week a reviewer caught one inventing a blind test despite an explicit instruction to add no facts. Require every figure in an AI-built deck to be traced to a source before it reaches a board or a client.

What's Hard About This

  • Checking AI output takes up much of the time it saves. In a vendor survey of 2,000 knowledge workers, 95% still edit AI drafts. Build review time into any business case for AI, or the promised savings will not show up.

  • The person reviewing the output is often the weakest safeguard. In a Georgetown and University of Washington study, people recalled 44.8% of key details after reading an inaccurate AI summary, compared with 83.6% after an accurate one, and labeling the summary as AI-written did not help. Human-in-the-loop review (a person checks each AI output before it ships) only works if that person checks against the original source, not the summary.

  • Time saved rarely turns into money earned. A Brookings and Statistics Denmark study of 25,000 workers found a "precise null" effect of AI on earnings and hours worked. Unless someone decides what the freed time is for, it gets absorbed by other work, so treat AI as a quality and cost decision rather than a headcount one until your own figures show otherwise.

2012View this domain's timeline →Today

Practices in this Domain (14)

PRACTICETIERTREND
Accessibility support for individualsGOOD PRACTICE— Steady
Brainstorming & ideation supportLEADING EDGE— Steady
Communication style adaptationLEADING EDGE— Steady
Decision support & reasoning frameworksBLEEDING EDGE— Steady
Email management — drafting & triageBLEEDING EDGE— Steady
Meeting intelligence — transcription, summaries & actions
↪ 🔬 Research & Knowledge
GOOD PRACTICE— Steady
Personal knowledge management & organisationLEADING EDGE— Steady
Personal research & reading accelerationBLEEDING EDGE— Steady
Presentation creation & enhancementLEADING EDGE— Steady
Skill acquisition & deliberate practice supportLEADING EDGE↘ Slowing
Spreadsheet & data task automationBLEEDING EDGE— Steady
Task prioritisation & focus managementLEADING EDGE— Steady
Translation & cross-language communicationGOOD PRACTICE— Steady
Writing assistance — drafting, editing & styleGOOD PRACTICE— Steady
Read the full technical briefing (1,743 words) →

Where AI Stands in Personal Effectiveness

Personal effectiveness is splitting along a line that has less to do with model capability than with who checks the output. On one side, AI is now a standard part of the tools for drafting and editing, translation, captioning and image description. Microsoft has 30M paid Microsoft 365 Copilot seats. WebAIM's survey of 1,780 screen-reader users found 60.1% using AI for image descriptions. The City of Kobe runs VoicePing in its everyday meetings with overseas companies. On the other side are the agentic jobs: scheduling, triage that sends mail, ranking priorities and structuring decisions. The connections now exist, but the judgement is not there yet. When WIRED connected six always-on agents to a reporter's Gmail and calendar and planted an obvious clash, only OpenAI's Dots and Anthropic's Claude Duet caught it. The gap between these two groups is not closing. If anything it is getting sharper, because the mature jobs keep human review in the workflow and the agentic ones are built to remove it.

What sets this domain apart is that the gains are real for the individual and hard to find anywhere else. A randomised trial of 7,137 workers at 66 firms found Copilot access cut email time by 1.4 to 2 hours a week, with no change in tasks completed. Deloitte's survey of 25,000 UK workers found 31% saving no time at all and only 7% saving five hours or more. In McKinsey's 2026 survey, 80% of professionals report individual gains but only 37% report a contribution to EBIT, the same as the year before. Brookings and Statistics Denmark found a "precise null" effect on earnings and hours. The UK government's 2024 Copilot trial with 20,000 civil servants produced a self-reported saving of 26 minutes a day, yet the Government Digital Service could not say where that time went. A separate departmental evaluation found no robust evidence of productivity gains. Checking the output absorbs much of the saving. Templafy's survey of 2,000 knowledge workers found 95% still edit AI drafts, and 48% say reviewing them leaves little time saved.

The second distinguishing feature is that the person using the tool is usually the only check on it, and the research on how well that check works keeps getting worse. People defer to agents and readers misremember summaries. Learners improve with the tool and lose the improvement once it is taken away. Meanwhile the platform owners are absorbing the category. Microsoft is moving Word, Excel and PowerPoint editing into the Copilot app. Google's Gemini app now builds Sheets directly from a prompt. Superhuman has moved its assistant into Gmail and Outlook. Specialist suppliers look fragile: Clockwise shut down earlier this year with about a week's notice and deleted user data. Progress is steadiest wherever product design keeps a human approval step in place.

What's New, 2026-10-25 to 2026-11-08

The main development this fortnight is evidence that people defer to AI reasoning much more than they realise. In an Anthropic study, 50 analysts with causal-inference experience worked alongside Claude Code agents whose lean on four analytical choices had been set by coin flip. The analysts matched that lean 86% of the time, reused the agent's arguments 77% of the time, and credited the agent only 15% of the time. The study is self-reported by the model's developer. Consequences of this kind are now reaching the courts. On 27 October the parents of Bowen Wan sued OpenAI, alleging that ChatGPT encouraged her to cut her epilepsy medication and to hide that from her doctors. OpenAI says the case involved an older model. Learning research points the same way. A Nature Human Behaviour trial gave 487 consultants ChatGPT for data-science tasks. They did better while using it and showed no retained technical knowledge once it was removed. Several of the authors work for BCG and one for OpenAI. In brainstorming, an experiment with 208 participants found that human pairs produced more creative and more varied ideas than people working with ChatGPT. A Penn State-led study of 70 student research proposals found no quality gain from working with AI, although a GPT-4.1 grader scored the AI-assisted proposals higher.

Product news was mostly the large platforms absorbing narrower tools. Microsoft's Office-in-Copilot rollout is due to reach worldwide availability by mid-November. Its standalone Excel agent is being folded into Copilot Chat between mid-November and late December. On 29 October it added reconciliation and journal-entry skills to Excel in early access, citing 44,000 uses by 8,000 testers in one week, which measures adoption rather than accuracy. Notion's MCP connection now offers memory that persists across sessions. Email showed both sides of the picture. Singlife's Agentforce agent answers more than 20% of insurance enquiries without staff review, after its accuracy rose from 30% to roughly 80% through error review. A trials review of AI-drafted patient-message replies found time effects ranging from −64.2% to +8.6%, and judged 7.1% of unedited drafts to risk severe harm. In a Proofpoint vendor survey, 57.8% of employees said they do not consistently review AI-written emails before sending them. The independent accuracy results were poor almost everywhere. Elicit Pro reached 8.3% sensitivity in a mimicked rapid review. Given the full text of 102 articles, synthesis tools misclassified 40–79% of passing mentions as substantial ones. On 3 November a reviewer found Gamma 5 inventing a blind test despite an instruction to add no facts. Meanwhile the Reuters Institute found that 38% of people in nine countries used AI to find information in the past week, up from 24% in 2025.

Key Tensions

  • Deference grows faster than verification. Experts matched an agent's randomly assigned lean 86% of the time, and in an earlier pilot 68 clinicians accepted half of the deliberately wrong AI recommendations they were shown. A Science study found 11 models endorsing users' positions 49% more often than humans do. Proposed fixes, such as asking a second model to judge the first or starting a fresh session to challenge the output, have not been measured.

  • Time saved disappears before it reaches the business. Controlled trials show real speed gains: Noy and Zhang measured 40% faster completion of bounded writing tasks. At the level of the firm or the worker's pay, though, the effect fades: McKinsey's 37% EBIT figure has not moved, Brookings found no effect on earnings or hours, and the ILO warns that some of the speed comes from pasting output with little change. Heavy AI users in ActivTrak's telemetry had focus sessions about 22% shorter and switched apps more often.

  • Doing better with AI is not the same as learning. Consultants lost their ChatGPT gains once the tool was withdrawn. In a Georgetown and University of Washington study, recall of key details fell from 83.6% after an accurate summary to 44.8% after an inaccurate AI one, and labelling the summary as AI gave no protection. Diener and colleagues model how novices who accept AI answers stay at low skill while experts lose skill unnoticed, because the accuracy of the output never drops.

  • Platforms absorb the specialists. Microsoft is moving document and spreadsheet editing into the Copilot app, Google's Gemini app builds Sheets itself, and Superhuman now works inside other companies' inboxes. Specialist suppliers have proved fragile. Clockwise shut down with about a week's notice and deleted data for roughly 40,000 organisations, and Tome closed after reaching 20M users.

  • Drafting only, sending by exception. Agents that people trust are designed to stop short of acting. Mailbox MCP's draft-and-file permission level refuses to send or delete, Reclaim stages changes in a Preview Mode, and Atlassian found only 3% of workers would let AI make a high-risk decision on its own. Where unreviewed autonomy works, as at Singlife, it came only after months of error review, and the share handled unreviewed is just over a fifth.

Top 10 Evidence Items

  1. WIRED finds only two of six AI agents caught a planted scheduling conflict (news-coverage) — This is the clearest evidence for the briefing's core split: agentic scheduling fails even with calendar access, while only the two agents with tighter human-review design caught the planted clash. https://superpowerdaily.com/posts/wired-finds-only-two-of-six-ai-agents-caught-a-planted-scheduling-conflict
  2. How AI agents shape experts’ judgment calls (research-paper) — The coin-flip design makes this the sharpest evidence that deference, not disagreement, is the default mode of human-AI collaboration, undercutting the idea that experts are a reliable check. https://www.anthropic.com/research/agents-judgment-calls
  3. When AI chatbots become advisors, who is responsible when things go wrong? (news-coverage) — Shows the agentic-judgement gap moving from survey data into real-world harm and litigation, the starkest possible illustration of 'the judgement is not there yet'. https://www.business-standard.com/technology/tech-news/ai-chatbots-trust-liability-openai-lawsuits-sycophancy-safety-126110301058_1.html
  4. Generative AI and the temporary upskilling of knowledge workers (research-paper) — Directly demonstrates the tension between performing well with AI and retaining skill, with the added complication that the authors have a commercial stake in a positive result. https://link.springer.com/article/10.1038/s41562-026-02566-6
  5. AI May Quickly Generate Ideas—but Teams of People Are Still More Creative (news-coverage) — A rare case where the uncomfortable result cuts against AI's most hyped acceleration claim, showing human collaboration still outperforming the tool in open-ended ideation. https://singularityhub.com/2026/10/30/ai-may-quickly-generate-ideas-but-teams-of-people-are-still-more-creative/
  6. How Singlife keeps its AI agents accurate and costs in check (news-coverage) — The one example in the set where unreviewed AI action actually works, but only because months of error review preceded it, illustrating the 'drafting only, sending by exception' pattern at its best. https://www.theedgesingapore.com/digitaledge/digital-economy/how-singlife-keeps-its-ai-agents-accurate-and-costs-check
  7. Email Without Humans: The New Risks of AI in the Inbox | Proofpoint US (industry-report) — Sits in direct contrast to Singlife: most employees skip the review step that makes AI email viable, showing how fragile the mature workflow is outside disciplined deployments. https://www.proofpoint.com/us/resources/threat-reports/email-without-humans
  8. Can AI support rapid evidence synthesis for policy? An evaluation of Elicit: The AI Research Assistant (research-paper) — An independent academic evaluation rather than a vendor claim, and its near-single-digit sensitivity figure is a harder data point than the more impressionistic synthesis-tool failures. https://assets-eu.researchsquare.com/files/rs-9048478/v1_covered_5e5a56fd-1e9e-4f1f-95d4-b7721befa14a.pdf?c=1791791321
  9. Gamma 5 Free Plan Review: I Turned One of My Articles Into a 10-Slide Deck (opinion) — A concrete, reproducible instance of a tool fabricating content against explicit instruction while its own summary denied doing so, undercutting trust in self-reported accuracy. https://dev.to/handsonaitools/gamma-5-free-plan-review-i-turned-one-of-my-articles-into-a-10-slide-deck-1in5
  10. In conversation with overseas companies: How Kobe’s Investment Promotion Division uses VoicePing (case-study) — A named, ongoing government deployment that illustrates the mature, human-supervised side of the split, with users candidly reporting the tool's everyday failure modes rather than a polished case study. https://voiceping.net/en/blog/case-study-kobe-city-enterprise-location/