The Week in AI Adoption: Supervision becomes the scarce resource, May 16
Nineteen posts, three refreshed insights dashboards, one shared verdict. The bottleneck moved from deployment to supervision, and the bill is landing this quarter.
The week in one glance
- The three Cerevisor insights dashboards refreshed on May 14, and they all point at the same shift: deployment was the old scarcity, supervision is the new one.
- On the engineering side, Microsoft cancelling internal Claude Code licenses and Jellyfish reporting a 64 percent velocity claim arrived in the same 48 hours. The board needs a different number.
- Inside the human, attention researchers gave the new tiredness a name: vigilance fatigue. It is the cost of being the supervisor when the agent does the typing.
Start here: the three insights dashboards just refreshed
Before the weekly posts, three pages on the site quietly updated on May 14, and each one compresses several weeks of coverage into a single argument. They are the fastest way to brief a board, a hiring manager, or yourself on where the field actually moved.
The headline: adoption is no longer the moat, supervision is. 88 percent of companies now use AI in at least one function, 96 percent run agents, and 20 percent capture 74 percent of the value. The gap is not models, money, or talent. It is whether one named person can tell the board on Friday what every AI in the building did this week.
Adoption of coding agents hit 84 percent. Trust in the output sits at 29 percent. Fifty-five points of daylight between using the tool and believing the tool, and the renewal contract is splitting into platform, verifier, and governance line items because of it.
The fortnight's research converges on one mechanism. When typing, drafting, and first-pass synthesis move into the model, the human residue is judgment, scope-keeping, and integration. The inner life is reorganizing around supervision faster than the language to describe it.
Theme of the week
Three pipelines, nineteen posts, one verdict. Supervision is the scarce resource. The boardroom version is the governance ownership gap. The engineering version is the trust-versus-adoption gap. The human version is vigilance fatigue. The same structural shift shows up at three altitudes in the same week, and the insights dashboards crystallize it: the old question (are we using AI) is settled, and the new question (who is accountable for what the AI did this week) is the one that decides the next ROI report, the next renewal, and how the supervisor feels at 6pm.
What we published
AI adoption this week
Goldman named the FOMO cycle and ServiceNow showed Wall Street the math. I wrote the ROI report that holds both rooms.
AWS, Anthropic, and Cognizant all moved in 72 hours. Five contract commitments your board should require before signing any agentic deal this quarter.
Cloudflare cut 1,100 people and credited 100x AI gains. The April jobs report showed 16 months of white-collar contraction. Most CEOs are running the layoff before the redesign.
Three AI agents reviewed an 8.5 million dollar federal proposal in the time humans take weeks for. Four questions, one weekend, one page.
Two regulators slipped their AI deadlines. The courts and procurement teams did not. What audit committees should be asking before Q3.
OpenAI raised over 4 billion for a Deployment Company the same week IBM Think 2026 made the operating layer the moat. The five-line audit a Series B board should expect.
AI coding agents this week
Coinbase cut 14 percent of staff. Anthropic put a 23,000-engineer customer on stage promising 90 percent autonomous coding. Neither number survives a sharp board director.
TrustFall hit four coding-agent CLIs with one folder-trust prompt. The register flipped from a vendor list to a convention list.
Jellyfish surveyed 636 engineering professionals and dropped two different 64 percent numbers. Only one is making the slide.
Background agents shipped code while engineers slept. The 8am question is which named person owns the triage.
Cursor retired a seat-priced SKU and slid into Microsoft Teams. GitHub Copilot CLI shipped autopilot. The renewal conversation flipped.
94 percent of engineering leaders admit their AI productivity metrics miss the work that matters. The 14-day pilot shape that turns the admission into a procurement page.
Microsoft cancelled thousands of internal Claude Code licenses the same week Jellyfish reported 64 percent of leaders claiming a 25 percent velocity gain. The number an executive can actually defend.
Self-awareness in the age of AI this week
A new Behavioral Sciences study traces a quiet path from heavy AI use to eroded self-efficacy. The slip is small, and it shapes the next task.
Two peer-reviewed studies find the felt sense of being capable shapes how teams use AI, and heavy passive use erodes it in persistent ways. The lever sits upstream of the dashboard.
A 2026 study finds AI assistants report near-ceiling confidence even when accuracy slips. For builders, that gap changes what is worth surfacing to users.
New research argues AI tools are now part of how a leader's sense of self at work gets re-formed week by week.
Two peer-reviewed experiments find the most automated AI assistance produces the least exploration and the lowest learning, even though people prefer it.
Attention researchers have a 75-year-old name for the kind of tired that follows AI supervision: vigilance fatigue. A new Human Factors paper moves the field into the modern workday.
Signals to implications
Signal. Microsoft cancelled thousands of its own Claude Code licenses on the same week Jellyfish reported 64 percent of engineering leaders claiming a 25 percent velocity gain from AI coding.
Implication. Strip the self-reported velocity number from the May board deck. Replace it with a defended cycle-time delta on a named cohort of repos. [Exec + Eng]
Source: Microsoft just cancelled its own Claude Code licenses
Signal. Background coding agents now ship code while engineers sleep, and the 8am review queue is the new bottleneck.
Implication. Put a named person on the org chart as the triage owner for the morning queue before Q3 hiring planning closes. [Eng Leader]
Source: The morning review queue is now an org-chart question
Signal. 88 percent of companies use AI in at least one function, but 20 percent capture 74 percent of the value, and the gap is who has full visibility into AI use.
Implication. Add one slide to the Q3 board pack naming the single accountable person for AI inventory across the org. [Exec]
Signal. A 2026 study finds AI assistants report near-ceiling confidence even when their accuracy drops, while human confidence still tracks whether the answer is right.
Implication. If you build AI features, surface a calibration signal, not a confidence score. Users borrow the model's confidence by default. [Founder]
Signal. Attention researchers named the new tiredness of overseeing AI agents: vigilance fatigue, a 75-year-old construct from radar operators that now describes the workday of most leaders.
Implication. Stop scheduling oversight back-to-back. The cost is real, the recovery is slow, and the calendar is the only place you can defend it. [Self-aware Worker]
The contrarian take
The popular reading of this week is that AI is finally working, because the productivity claims got bigger. The honest reading is the opposite. The 100x story behind Cloudflare's headcount cut, the 64 percent claim that Microsoft contradicted with its own cancellations, and the 55-point trust gap on the harness insights page are not three stories. They are one story about a missing supervisor, told in three rooms. The companies that will outperform in Q3 are the ones that staff that supervisor before the renewal, not after, and the ones that price the vigilance cost into the role from day one.
Next week
If this recap was useful, the newsletter delivers it straight to your inbox every Monday. Subscribe here.