This is the result of running the toolkit on real logs from a month of work by Nei Maldaner (INEMA) with Claude Code and Codex. It shows how to read your own ladder and where it points.
From Oct 13, 2026, to Nov 12, 2026. The toolkit read 9,959 Claude Code sessions and 3,233 Codex sessions: 1,366 conversations with Nei and 11,269 sessions that ran on their own (scripts, timers, bots). AI performed 111,442 actions (commands, edits, reads). A low-cost model (codex gpt-6-luna (low)) only grouped requests into tasks; each session’s level is based on recorded behavior, not opinion.
The bar shows the total hours of AI-assisted work for the month, divided across the four levels. Half are already on the upper rungs of the ladder.
Each bar represents a task: its length shows hours per week, and the colors show the level at which it was done. ⚠ = critical task (public-facing or involves a system), which the toolkit keeps at AL4.
| Task | h/wk* | Level | AI leads | Runs on its own | Where it’s headed |
|---|---|---|---|---|---|
| Produce videos and series critical | 75.9 | AL4 | 64% | 18% | AL4 ceiling (critical) · 95% at AL4+ in Mar 2027 |
| Other tasks | 46.0 | AL3 | 43% | 31% | 50% at AL5 in Dec 2026 |
| Create projects and toolkits | 19.0 | AL4 | 53% | 0% | AL5 not yet measured |
| Research and validate content | 13.8 | AL3 | 35% | 7% | 50% at AL5 in Mar 2027 |
| Maintain systems and servers critical | 8.4 | AL3 | 25% | 0% | AL4 ceiling (critical) · 95% at AL4+ in Jun 2027 |
| Translate content | 6.5 | AL5 | 65% | 54% | Already over 50% at AL5 |
| Analyze metrics and operations critical | 5.7 | AL3 | 36% | 10% | AL4 ceiling (critical) · 95% at AL4+ in May 2027 |
| Create courses and training | 4.7 | AL4 | 74% | 0% | AL5 not yet measured |
| Plan products and campaigns critical | 3.2 | AL3 | 15% | 0% | AL4 ceiling (critical) · 95% at AL4+ in Jul 2027 |
| Publish and promote content critical | 1.7 | AL3 | 33% | 0% | AL4 ceiling (critical) · 95% at AL4+ in May 2027 |
* Hours of work WITH AI: they add together parallel sessions and machine time (so the total, 184.9 h/week, exceeds a real week). They’re useful for comparing tasks, not for counting a person’s time. Work done without AI (recording, meetings, calls) doesn’t appear in the logs.
In its Nov 2026 report, AI led 26% of Anthropic’s research, and no task was fully automatic. In Nei’s case, AI leads twice that share, and nearly one-fifth already runs on its own, mainly translations and video workflows.
Illustrative comparison: a content creator’s routine isn’t the same as AI model research. The point is the pace, not an exact yardstick.
The curve uses the same rate projected by the AI 2027 authors for automated research (from 50% in Jun 2027 to 95% in Nov 2027), applied to Nei’s starting point. Amber line: share led by AI (AL4+AL5). Orange line: share that runs on its own (AL5).
| Month | Date | AI leads | Runs on its own |
|---|---|---|---|
| 0 | Nov 2026 | 52.2% | 17.9% |
| 1 | Dec 2026 | 66.2% | 28.1% |
| 2 | Jan 2027 | 77.9% | 41.3% |
| 3 | Feb 2027 | 86.4% | 55.8% |
| 6 | May 2027 | 97.3% | 88.0% |
| 9 | Aug 2027 | 99.5% | 97.7% |
| 12 | Nov 2027 | 99.9% | 99.6% |
This is an extrapolation, not a prophecy. It assumes Nei’s routine accelerates at the same pace as the lab and ignores the ceiling for critical tasks: in practice, publishing and changing servers will still require human review, so total AL5 won’t reach 100%.
The numbers show that Nei no longer does most of the work himself: he directs it. In one month, there were 11,269 automated sessions and only 1,366 conversations. Daily work becomes five things AI doesn’t do for him.
If you use Claude Code or Codex, two commands generate the same report for you. The first (de-logs.py) reads the records only on your computer, without internet access. The second (escada-logs.py) labels tasks in one of three ways: --motor codex, the default, sends the beginning of your requests and the task and hours table to OpenAI through your Codex subscription to write the summary (the kit removes known keys, emails, and long numbers; the rest of the text is sent); --motor ollama uses a model on your computer; --motor palavras uses keywords and sends nothing. Confidential data (client work, confidential information): use --motor palavras or ollama, and run the commands yourself in the terminal, because if you ask your AI to run them and summarize the results, the report passes through its provider. Claude Code deletes records older than 30 days (increase cleanupPeriodDays to measure a longer period); if the period is missing, de-logs.py will let you know.
python3 kit/scripts/de-logs.py --dias 30 # reads the logs → sessoes.json python3 kit/scripts/escada-logs.py --motor codex # or --motor ollama / palavras
Technical details of the toolkit ↗ · Back to the Relógio 2027 guide