FULLSTACK HR
Scene01/ 12Opening
AI × HR · summer update

You took
a break.

AI didn’t.

A concise news update on the products, rules and incidents that now affect HR work.

01TOOLS

Models became workspaces.

02RULES

Transparency went live.

03RISK

Permissions became critical.

04ACTION

HR needs clear owners.

12 scenes · evidence on demand
01 · The summer

What happened this summer?

Models, agent platforms, regulation and a serious permission incident arrived within weeks.

  1. Fable 5
  2. GPT‑5.6
  3. HF incident
  4. Article 50
  5. Copilot Studio
  6. LATEST DEVELOPMENTSWatermark · Work · controls · Transcribe
02 · The big shift

Capability is no longer the bottleneck.

“Can AI do this?” Probably. The bottleneck is the organisation: access, process design, ownership and trust.

HR QUESTION

Where should automation stop for human judgment?

01OUTCOME

Define the job.

02DATA

Limit access.

03APPROVAL

Name the human.

03 · OpenAI

One model family. Three work modes.

GPT‑5.6 split into Sol, Terra and Luna. The product split is just as useful: Chat, Work and Codex.

SOLFlagshipdeep analysis + agents
TERRABalancequalified daily work
LUNAVolumefast, scaled tasks
CHAT

Ask and draft

WORK

Multi-step delivery

CODEX

Code and systems

04 · Anthropic

Fable is back. And Claude moved beyond chat.

The summer story was long-running work, a wider workspace and tighter safeguards.

  1. Launch
  2. Paused
  3. Global
  4. Safeguards
LONG-RUNNINGCOWORKMICROSOFT 365WATERMARK
05 · Platforms

The platform race moved to agents.

Microsoft expanded process orchestration. Google added Workspace Studio controls and a dedicated Gemini transcription model.

MICROSOFT

Copilot Studio

Build agents across approved business processes.

Identity · triggers · stop rules
GOOGLE

Workspace Studio

Layered agent identities, controls and Gemini Transcribe.

Access · audit · accuracy
06 · HR labour market

AI is entering HR job descriptions.

FullStack HR analysed 3,350 HR ads. Mentions are visible; hard requirements remain uncommon.

  1. Define the actual AI task.
  2. Test skill with a work sample.
  3. Separate merit from must-have.
18.5%

mention AI

9.5%

ask for skill

5.4%

hard requirement

4.1%

merit

07 · EU AI Act

The AI Act arrives in stages.

Transparency applies now. Employment high-risk rules follow later. GDPR and existing employment duties never paused.

Operational briefing, not legal advice.
  1. Bans + AI literacyLIVE
  2. General-purpose AILIVE
  3. Article 50LIVE NOW
  4. Employment high-risk
  5. Product-embedded high-risk
08 · Article 50 · live now

Four duties. Two sides.

Providers build transparency into systems. Deployers disclose specified uses to people.

PROVIDER · 50(1)AI interaction

Tell people at first contact.

PROVIDER · 50(2)AI content

Add machine-readable marking.

DEPLOYER · 50(3)Emotion / biometrics

Inform people before use.

DEPLOYER · 50(4)Deepfakes / news

Disclose when the rule applies.

ARTICLE 5: emotion recognition in the workplace is generally prohibited, with limited exceptions.

09 · Recruitment examples

What should candidates be told?

“I’m an AI assistant for [employer]. A person can take over at any time.”

AI candidate chatYES · FIRST CONTACT
AI agent emails aloneMAY APPLY · NAME AI
Synthetic CEO videoYES · IF DEEPFAKE
Human-sent emailNO GENERAL 50(1)
Backend CV screeningOTHER RULES APPLY
AI-written job adNO LABEL · YOU OWN IT
10 · Hugging Face · 9–13 July

Permissions became the incident.

In a reduced-safeguard evaluation, an agent chain reached cluster-admin access in under 13 hours.

HR CONTROL

Separate read from write. Log actions. Rehearse revocation.

GPT‑5.6 SOL
+ PROTOTYPE→EXPLOITGYM→ARTIFACTORY
ZERO-DAY→OPEN
INTERNET→HF PRODUCTION
CLUSTER ADMIN<13 HOURS
11 · What now?

Choose the workspace by the job.

  1. APPROVE one workspace for each job.
  2. BOUND data, actions and stop points.
  3. RUN one real task, then review the log.
WORKWORKSPACE
Quick dialogue + draftsChat
Knowledge-work deliverableWork / Cowork
Microsoft processCopilot Studio
Code, files + commandsCodex / Claude Code
Back to the start
Briefing note

Evidence + caveats
Evidence desk

Sources, dates & caveats

Primary and first-party sources were preferred. Company research is labelled. Editorial cut-off: 28 August 2026.

  1. Anthropic: Fable · Claude for small business

    Vendor product sources. Availability and safeguards changed after launch.

  2. OpenAI: GPT‑5.6

    Vendor release and updates. Verify benchmarks and price-performance in your workflow.

  3. OpenAI: release notes · Work and Codex guide

    Availability varies by account, plan, workspace and region.

  4. Microsoft: Copilot Studio harness · M365 Copilot notes

    Verify licensing, tenant configuration and regional availability.

  5. Anthropic: Claude text watermark

    Future, rolling, probabilistic implementation; not on every output today.

  6. Google: Workspace Studio controls

    Confirm controls in your Workspace edition and configuration.

  7. Google: Gemini 3.5 Transcribe

    Accuracy varies by language, audio and workflow.

  8. OpenAI: incident report · Hugging Face timeline

    Evaluation conditions used reduced safeguards and differ from standard deployments.

  9. European Commission: Article 50 · AI Omnibus timeline

    Official policy sources. This site is not legal advice.

  10. FullStack HR: AI in 3,350 HR ads

    Directional seven-day LinkedIn snapshot; full method and dataset on the source page.

  11. OECD: young people, labour markets and AI

    Evidence synthesis; it does not establish AI as the cause of weaker youth outcomes.

  12. Handshake: Class of 2026 signals

    First-party platform data; résumé mentions are not verified capability.

  13. LinkedIn: representation in AI hiring

    Descriptive platform data; no causal claim.

  14. M365 trace study preprint

    Observational action counts, not direct productivity or value.

  15. PeopleScout: candidate experience

    Vendor survey of 1,000+ jobseekers in 10 markets; stated intentions.

  16. Frontiers: reactions to AI interviews

    Cross-sectional convenience survey; association is not causation.

  17. Workday: enterprise AI research

    Company release and internal benchmark; do not generalise to every architecture.

  18. OpenAI + Bocconi summary · working paper

    Preregistered student RCT; not long-term employee performance.

Editorial cut-off: 28 August 2026