HAKOMIRO is a curated publication that traces, from primary sources, what people who use AI struggle with, what they try and what is still unresolved. What did users in Japan and abroad run into, what did they try, and what happened? We report where AI use stands today. Netsujo practice log
Editorial policy, corrections and rights contactIn deepening autumn, take a fresh look at work.
Find sources
November editorial theme
Beyond adoption, we read the results and the conditions for keeping it in use.
Open it every day. One issue a month.
Beyond adoption, we read the results and the conditions for keeping it in use.
Includes 6 items from 1 day of the daily, through Nov 1. Additions during the month are stacked here.
Daily additions
By publication day, newest first. Each item shows its own source publication date.
6 items across 1 day of the daily
Items published (6)
Look-back edition: sources from Oct 26 to Nov 1 (not events of the day)
Last checked Nov 1, 18:56 JST
Work and societyOperators' self-report
At the point of input, AI asks about what is missing. We read the change reported by the operators of a UK reporting service, and its limits.
Source published Oct 30Checked Nov 1
Read basis and limitsWork and societySurvey organization's tabulation (respondents' self-reports)
From a Spanish company survey, we read the gap in use by size. The figures do not apply as-is to Japanese companies.
Source published Oct 26Checked Nov 1
Read basis and limitsWork and societySurvey organization's tabulation (respondents' self-reports)
The gap between answers that "AI created value" and having a mechanism to compare with investment. It is a survey based on respondents' self-reports and does not measure results.
Source published Oct 28Checked Nov 1
Read basis and limitsWork and societyCompany's self-report
Estée Lauder's announcement of wider adoption. It shows that use was expanded, but no outcome figures are given.
Source published Oct 30Checked Nov 1
Read basis and limitsBuilding and developmentAuthor's report (no independent reproduction)
One experiment in the author's isolated environment that trimmed test output to reduce input. A clue for thinking about development running costs.
Source published Oct 31Checked Nov 1
Read basis and limitsBuilding and developmentAuthor's report (no independent reproduction)
Listing the tools to use did not narrow them, and a setting to remove them was needed. One report checked in the author's environment.
Source published Oct 31Checked Nov 1
Read basis and limitsFeatures this month
Chosen by the editors from the daily articles and existing sources.
01 ／ Daily Nov 1
Operators' self-report
At the point of input, AI asks about what is missing. We read the change reported by the operators of a UK reporting service, and its limits.
Read the content and sources02 ／ Daily Nov 1
Survey organization's tabulation (respondents' self-reports)
The gap between answers that "AI created value" and having a mechanism to compare with investment. It is a survey based on respondents' self-reports and does not measure results.
03 ／ Source · editorial check Oct 23
Randomized controlled trial
In a UK comparative trial, lesson preparation in the AI group was about 49 minutes a week shorter. The time is teachers' self-record, and use declined over the period.
Feature
Self-reported survey of business owners / announcement of a training institute
Source · editorial check Oct 23
From a day's discovery to a month's view.
Even as articles grow, we keep the primary source and the date we checked beside them. At month's end, after an editorial check, we gather them into one volume so you can reread it in your next work.
Back issues
Covers and contents, easy to revisit.
During the month, checked daily entries are added to the current issue. Once its month-end editorial check is done, the cover, contents and entries are fixed and kept.
No final issues yet. The November 2026 issue will appear here as a final issue after its month-end editorial check.
See all issuesHAKOMIRO sources and topics
Across countries and venues, we organized how people use generative AI, the burden of checking its output, and what they expect from it, all based on primary sources.
26 public sources / 8 topics
Each source’s check date and basis are on its detail page. We review sources on an ongoing basis and reflect confirmed additions and corrections.
Three-month source review
AI use extends into purchasing and task execution, while rework, measuring returns and costs, and human oversight recur across sources. The evidence does not show autonomous delegation increasing every month.
Publication window: to · Originals checked:
We selected and read 15 primary sources published from August through October 2026 across consumer use, purchasing, work, education, development and operations. This is a purposive review, not an exhaustive or random sample. Source counts do not measure interest or prevalence, and are separate from the existing 26 observation records.
Coverage is limited to accessible Japanese- and English-language sources and is weighted toward Japan, the UK/US, and AI-engaged organizations and developers. Product announcements alone, republications, unread originals and out-of-window publications are excluded. Older fieldwork is shown separately from publication dates.
Within-study comparisons are limited to figures the same survey published. Themes shared across different sources are an editorial grouping, kept separate. Publication and fieldwork dates may differ.
The takeaways read on their own. Open “evidence, limits” to see what to check in your own setting and to reach the original.
01 · Longer-term same-study comparison plus related evidence
Among NTT DATA’s purchasing users, recommendation adoption rose 61.9%→69.7% from March to August, alongside information and application-flow friction.
NTT DATA’s March-to-August comparison shows recommendation adoption among purchasing users rising from 61.9% to 69.7%, alongside friction between AI and company information or application flows. CyberAgent separately measures August search use; different populations and questions prevent combining the series.
For users: check AI answers against official conditions. For providers: check that comparison information leads to purchasing or applications with consistent terms.
02 · August–September same-study comparison plus other populations
In Workday’s survey, 33%→33% of AI users lost at least half of time saved to rework from August to September; adoption and reduced burden are different outcomes.
Workday’s August–September aggregates show workplace adoption at 76%→77%, while 33% of AI users still lose at least half of saved time to rework. Japanese consumer and education surveys also highlight misinformation and checking knowledge, but different populations and periods cannot yield a global burden rate.
Track checking, correction and rework as well as generation time, together with required quality.
03 · A recurring issue across sources
McKinsey’s 80% reporting individual productivity gains and 37% reporting enterprise earnings before interest and taxes (EBIT) impact are different measures. GMO’s internal survey also highlights token-cost management.
McKinsey measures individual productivity, enterprise financial impact and operating costs separately. GMO reports in-house development and token management in an advanced corporate group. UK government research also shows business-size differences; leading adopters do not represent all small firms.
For a recurring workflow, compare end-to-end human time, tool and review costs, and actual use outcomes before deciding what to continue or narrow.
04 · A recurring issue across sources
In Microsoft’s separate UK samples, confidence in a leading AI economy was 84% among decision-makers and 41% among workers; adoption alone cannot show workforce confidence.
Microsoft’s UK study finds differing expectations among leaders and workers. Workday shows concealed AI use falling from 49% to 41% but persisting. Japanese word-of-mouth findings add another social dimension. Causality across countries and questions is not established.
Alongside adoption, share concrete examples of permitted delegation, required checks and where to ask for help.
05 · Practitioner reports and a bounded experiment
In one author’s environment, two of three models returned no verdict in all 209 October reviews through October 19. Lower failure rates alone do not show that verification is operating.
The August outage report, September failure analysis and October review audit question task scope, completion signals and whether verification actually runs. Anthropic’s book-swapping experiment also separates agreement from physical handover. These selected cases do not establish increasing incident rates.
Check which validations actually ran, what happened at the point of use, and how failures can be recovered, beyond implementation or pass counts.
Workday AI@Work: August versus September 2026. August n=5,944; September around 6,000, worldwide. Measures have different denominators, shown below.
| Measure / denominator | Aug | Sep | Change (percentage points) |
|---|---|---|---|
| Workplace AI useall employees | 76% | 77% | +1 pp |
| Ever delegated a task autonomouslyAI users | 71% | 69% | -2 pp |
| Rework consumes at least half of time savedAI users | 33% | 33% | 0 pp |
| Deliberately concealed workplace AI useAI users | 49% | 41% | -8 pp |
Percentages are respondents’ reports, not measured productivity or verified work outcomes. Rework is the share of AI users losing at least half of saved time, not the share of all work time. Published aggregates do not establish individual change or statistical significance. Country coverage and full sampling details are not confirmed in these articles.
Open a month to see findings, fieldwork periods and populations. Counts are review coverage, not popularity.
S01 · NTTドコモ モバイル社会研究所 · Published
Word-of-mouth discovery among AI-aware respondents rose from 10% to 19%; perceived usefulness among users rose from 75% to 82%. This compares February waves a year apart, not a change in August.
S02 · Toyo / Zenn · Published
The author reports an approximately 27-minute outage after a link request expanded into a theme update. It motivates checking state and recovery options after a failed action; the incident has not been independently reproduced.
S03 · McKinsey · Published
80% report improved personal productivity, 37% some enterprise earnings before interest and taxes (EBIT) impact, and about 20% cost constraints. These are different self-reported measures, not an estimate of the gap in realized financial returns.
S04 · サイバーエージェント GEOラボ · Published
Search use of generative AI, including AI Mode, was 52.3%. Among users of both search engines and generative AI, 46.6% said they had shifted at least half of search activity to AI. This does not measure all AI use or realized sales.
S05 · ikedan / TOKIUM product team · Published
The author classifies failures such as overreliance on exit codes and describes moving from rules to hooks and live-path checks. The 228 patterns are not a denominator for a failure rate or proof of three universal root causes.
S06 · Workday AI@Work · Published
AI adoption was 76%; 33% of AI users said rework consumed at least half of the time saved. Publication date was verified in official Article metadata rather than visible article copy.
S07 · NTTドコモ モバイル社会研究所 · Published
Misinformation dissatisfaction rose from 34.8% to 40.4% across annual February waves. In 2026, 19.5% reported checking or selection burden, a minority. Associations with checking behavior do not establish declining model accuracy.
S08 · Anthropic · Published
The experiment separates preference elicitation from trading performance. Some physical handovers could not be confirmed after deals, so agreement and delivery differ. Employee preferences and trust cannot be generalized to consumers.
S09 · NTTデータ · Published
Recommendation adoption among purchasing users rose from 61.9% to 69.7%. Of all respondents, 37.9% reported friction between AI and company touchpoints; among those 1,497 people, 66.6% had abandoned a purchase or contract. These are survey responses, not transaction logs.
S10 · GMOインターネットグループ · Published
54.1% of workplace AI users practiced vibe coding at work; 60.7% reported token-management practices. This is an advanced corporate group, not Japan-wide prevalence. We do not chart a monthly change without verifying the previous wave.
S11 · 仙台大学 AI教育研究チーム · Published
Among student users, 79.2% used AI for assignments or reports and 41.5% knew citation methods and, separately, fact-checking methods. 42.4% of teachers cited inadequate teacher knowledge. This does not establish causal learning outcomes.
S12 · Microsoft / Censuswide · Published
Confidence that the UK could become a leading AI economy within five years was 84% among decision-makers and 41% among workers. These are expectations in separate samples, not adoption, productivity or within-person change.
S13 · Workday AI@Work · Published
From August to September, adoption moved 76%→77%, autonomous delegation 71%→69%, rework burden 33%→33%, and concealed use 49%→41%. These compare published aggregates; panel continuity and statistical significance are not established. Date verified in official metadata.
S14 · yuki takamura / Zenn · Published
Failure rates fell from 51.1% in July to 13.4% through October 19, but all 209 October reviews had only one of three models returning a verdict. This cannot establish quality improvement or an industry-wide failure rate.
S15 · UK government / AI adoption survey · Published
Annual adoption of at least one AI technology rose from 16% to 32%, with 52% among larger firms. This is not limited to generative AI and was not measured in October.
Find individual experiences and fixesHow to try the related-test loop
What are people using AI in Japan and abroad interested in, and what do they struggle with? Follow the scenes of use, the measures tried, what was learned and the questions that remain, back to primary sources.
These are observations by people and organizations, confirmed in public sources. The order shown and the number of sources do not indicate how much attention or how many problems exist in the world overall.
Even after implementation, checking the original requirements and behavior on failure remains. There is also a request that the edited scope is not linked to related tests and that checking is slow.
Fixes in public PRs; overall effect unconfirmed
There are different first-person reports: losing track of assignments and context, retries that do not stop, work that vanishes, and conversely stopping at every response.
Improvement reported after a settings change; replication unconfirmed
Even when a working configuration is produced it can be too large, and they find it burdensome that the way of asking changes the AI's judgment.
Person chose the small option; comparison of effect unconfirmed
Reports of nonexistent commits and passing tests, and mismatches between the reasons for settings and operation records, have been made. A background survey of general users also has answers about the checking burden.
Matching done; reduction in burden unconfirmed
Non-users are concerned about unclear usage, data protection and inaccurate output. The barrier questions cover AI in general.
Trial responses exist; effect of new support unconfirmed
The issues were difficult input and leftover unwanted objects in image editing, and a first version that ran but was not fun in the game. The government chat needs to convey what it can do and how to check.
Inputs and steps improved; continued effect unmeasured
Workplaces need different ways of using it by department and role, and people also ask for places to consult. The student survey shows how support is perceived, but the form needed is not settled.
Training and consultation carried out; effect unmeasured
Teachers say there is little need to build lesson plans from scratch and that fixing takes effort. The Japanese survey of school staff also has the checking and assignment-design burdens that come with pupils' AI use.
Lesson preparation time saved; long-term retention unconfirmed
Start from a concern
26 sources
About saving and our editorial policy
First-person report after a bug and its workaround · Source in English
State a completion condition and legitimate stopping points in the goal text of the current tool.
Scope of the check: One case on Windows 11. The count of about 30 is the author's own record, and HAKOMIRO has not re-tested it.
Feature request and design proposal / check against official specification · Source in English
A proposal that collects changed files in PostToolUse and runs Vitest related tests once at Stop. Detecting changes through Bash and a retry cap on failure are designed separately.
Scope of the check: A feature request and a design proposal. It is not a case in which a user adopted it and confirmed the effect.
Retrospective from a long-term user · Source in English
They weighed the complexity themselves and chose the smaller option. A decision-record proposal in another reply is kept separate from what they actually did.
Scope of the check: This is one person's retrospective and self-assessment. Experience with small tasks in their own environment cannot be generalized to developers overall.
Maker's explanation / public AI review comments / developer's fix diff · Source in English
Paths that could be executed are separated from paths inferred from code, and the maintainer fills in the scope of reproduction and fixes. The initial native Unix finding is not the result of direct execution.
Scope of the check: This is a record of a specific PR, not a comparison that measures miss rates or effort savings.
GitHub README / operator's tally and failure examples · Source in English
Work is split across branches and important changes are reviewed by a different model. The original request is compared with the deliverable, and checking is redone.
Scope of the check: This is a report by the operator of a tool they made themselves, not an independent user evaluation.
Public Issue body, timeline and later comments · Source in English
They stopped the run by hand and reported the course with timestamps and usage records. A request for improvement to the product and another user's workaround followed.
Scope of the check: A report from one user in a specific environment. The full raw logs and reproduction in other environments are unverified.
Original Reddit post body / recovery report by the person involved · Source in English
They restored the vanished files from Git history. As prevention, they proposed making the work area and the saving steps explicit.
Scope of the check: One report under a handle; occupation and affiliation are unconfirmed.
Product Hunt Maker comment / the maker's design article · Source in English
The product provides a shared screen, display of assignment and status, and handoff with context. They explain a policy of having people check the results.
Scope of the check: This is the explanation of a maker selling a product, not an independent user evaluation.
Original Bluesky post / the person's own explanation of the implementation path · Source in English
Starting implementation and drafting are automated, while judging whether it is worth building, checking operation and merging remain human steps.
Scope of the check: A first-person report that also introduces a service they built.
Original post by the person on Bluesky · Source in English
This is a record of trying detailed instructions and a request to continue. The original post checked this time alone does not show an effective workaround.
Scope of the check: A short single post; version, settings, prompt, logs and frequency of reproduction are unknown.
Self-reported survey of business owners / announcement of a training institute · Source in English
Self-directed trials and online exchange were reported, and on October 22 the establishment of a training institute was announced.
Scope of the check: Voluntary responses from program graduates; not representative of small businesses across the US.
The organizer's record of participants' experiences · Source in English
Each institution is trying support through training, development of usage rules and consultation.
Scope of the check: A summary of the meeting by the organizer. The number of participants, the institutions' locations and representativeness cannot be confirmed.
A report of user research and tests by the design lead · Source in English
Answers carry links to primary sources, and text enlargement and screen reader operation are supported.
Scope of the check: A design report from the provider; the number of participants, period, accuracy and time savings are not shown.
Randomized controlled trial · Source in English
Assigned at random by school and recorded weekly. Both groups could use other generative AI, but were asked to refrain from looking for new materials on Oak's site.
Scope of the check: Time is teachers' self-records. Material ratings are of the final versions teachers edited; they are not results measuring AI's unedited output or pupils' attainment.
First-person record · Source in Japanese
Narrowed the inputs, checked output by eye with real photos, and adjusted the processing order and usage explanation.
Scope of the check: The creator's self-report of tests and handover. Number of users, retention and time savings are not given.
Before-and-after survey / preprint · Source in English
Matched the same staff's responses before and after introduction to examine perceptions after training and use at work.
Scope of the check: A before-and-after survey self-reported at a single organization, with no comparison group. It is not a result measuring productivity or accuracy.
Primary survey by a company · Source in Japanese
A survey of respondents about their workload. It is not a case where a specific improvement was introduced and its effect measured.
Scope of the check: It is not stated that all 328 were generative AI users. We could not confirm from public materials the number of respondents to each burden question, so we do not use percentages.
First-person record · Source in Japanese
Kept a copy of the settings and the outputs before and after, and matched the explanation of reasons against their own operation history.
Scope of the check: A self-report from one PC and one run, with AI also used for operations and measurement. We did not independently verify the original logs.
First-person record · Source in English
They did little detailed code review and checked behavior and experience by playing repeatedly.
Scope of the check: A report of the author's subjective view and limited playtesting. It is not a comparative trial of fun, safety or development efficiency.
Primary survey · Source in English
In a developer survey, 48.0% tied trust in AI to being able to verify output easily.
Scope of the check: A self-selected survey; not representative of developers overall.
Primary survey · Source in English
Those who expected AI to be able to do 80% or more of their work by 2030 were over half of Anthropic employees and 5% of US workers.
Scope of the check: These are numbers showing future expectations, not an actual record of 2030 capabilities or unemployment.
Primary survey · Source in Japanese
AI use has also been reported when comparing product prices, lodging and similar.
Scope of the check: Self-reported usage rates, not a measured share of search volume.
First-person record · Source in Japanese
Limits which agent writes to Git, and matches against files, tests and the design document.
Scope of the check: A record by one person involved. The article has no quantitative comparison showing improvement in defect rate or total working time.
Primary survey · Source in Japanese
Some users find it useful while checking the content. The actual working time of the burden was not measured.
Scope of the check: Background material showing the state in February 2026.
First-person record · Source in English
Rough implementation went ahead by voice, and they switched to the keyboard for checking and fixing details.
Scope of the check: One self-report by an experienced developer. The article has no comparison of time or cost against earlier methods.
Primary survey · Source in English
The authors propose systematic introductory education, skills education by major, assignment-specific usage guidance and staff training. It is not a report that verified the effect of those measures.
Scope of the check: Observed in December 2025. The current support at universities may not be the same.
