Issue #11 May 16, 2026

Claude Starts Dreaming, Anthropic Opens the Legal Stack — and How Enterprise HR Is Quietly Redesigning the Enterprise

AI Brief, Curated

Editor's Take

Two short stories from Anthropic this week, both worth attention. The first is the "dreaming" feature for Claude Managed Agents — a scheduled, between-session process that reviews the agent's memory, extracts patterns, prunes stale entries, and lets the agent self-improve overnight. The headline-grabbing piece is the framing (Anthropic chose to call it "dreaming," which is the kind of marketing that does the model researchers no favors); the substance is more modest but still meaningful. The second is anthropics/claude-for-legal, the legal-vertical analogue of the financial-services repo from last week's deep dive — twelve practice-area plugins, 80+ workflow agents, Apache 2.0, and the same "AI drafts, humans sign off" structural design. Anthropic now has two of the most prestigious verticals open-sourced in a fortnight. The Org section today is the long-form deep dive that I have been working on for several weeks: how enterprise HR is using AI to redesign HR workflows, the HR operating model, and the broader enterprise organization — all at once.

This Week's Stories

Claude Starts Dreaming: Self-Improving Agents Between Sessions

Anthropic / VentureBeat / The New Stack Industry

Anthropic announced "dreaming" for Claude Managed Agents on May 6 at its Code with Claude conference. Despite the evocative name, the mechanism is straightforward: a scheduled background process reviews an agent's prior session logs and memory store, merges duplicates, removes stale entries, and surfaces recurring patterns (repeated mistakes, team preferences, recurring task structures) so the agent's persistent context improves over time. Importantly, this is not changes to the model itself — it's curated memory updates to the agent's notes. You can have it run automatically or require human review before changes land. Early adopter Harvey (legal AI) reports task completion rates increased roughly 6× after implementation; Anthropic shipped two companion features the same day — "outcomes" (a self-grading loop where an evaluator scores agent outputs against a rubric) and multiagent orchestration (a lead agent fans work out to specialist sub-agents). Dreaming remains in research preview; access by request.

The implications are worth sitting with. Most agent systems today have terrible long-term memory hygiene — they accumulate context that contradicts itself, stale references that mislead, and patterns the agent should have learned that simply never get encoded. Dreaming addresses that at the agent layer rather than the model layer, which is the practical place to address it. If you are building or operating Claude-based agents at scale, this is the feature to watch. The risk is the obvious one: between-session memory updates without good guardrails are exactly the surface where an agent drifts in directions you cannot see. Anthropic's optional human-review pathway is the right design for production use; treat it as required, not optional, in regulated contexts.

Sources: VentureBeat · The New Stack · 9to5Mac · Techzine

Anthropic Opens Claude for Legal: 12 Practice Areas, 80+ Agents, Apache 2.0

Anthropic / GitHub / Artificial Lawyer Industry

Two weeks after publishing the financial-services repo, Anthropic shipped anthropics/claude-for-legal on May 12 — the legal-vertical analogue. The structure mirrors what they shipped for finance: 12 plugins covering specific practice areas (in-house commercial, privacy, corporate, employment, litigation, IP, regulatory, AI governance, plus law-school clinics and student learning), 80+ specialized agents (Vendor Agreement Reviewer, DSAR Responder, Claim Chart Builder, and many more), roughly 20 MCP connectors into legal-specific systems (Ironclad, Everlaw, CourtListener, Westlaw via Thomson Reuters CoCounsel), and a Managed Agents API for headless background routines. 882 stars and 165 forks in the first 24 hours. Apache 2.0 license, which means commercial use, modification, redistribution, and deployment with the firm's own API key are all permitted.

The strategic read is the same as last week's: Anthropic is converting workflow IP that took twenty years to accumulate inside top firms into a public reference implementation. The data still requires licensed seats (Westlaw, Ironclad, etc.) — the workflow does not. The "AI drafts, humans sign off" structural pattern carries over verbatim: none of the agents file documents, bind risk, or commit a firm to a position. Every output is staged for review by a qualified professional. This is precisely the design that makes adoption legally tractable in a regulated profession with bar-association exposure. With finance done two weeks ago and legal done this week, the read on Anthropic's next twelve months is clear: expect healthcare, audit/tax, and at least one more regulated vertical before year-end. The vertical-AI moat is being open-sourced, vertical by vertical, with the data connectors as the only remaining gating layer.

Sources: GitHub: anthropics/claude-for-legal · PYMNTS: Lawyers as Cowork's top users · Artificial Lawyer · Earlier knowledge-work-plugins/legal repo
Org Design & Changes, Curated

This Week's Essay

How Enterprise HR Is Using AI to Redesign Itself — and the Enterprise Around It

Heads-up: This is a long deep-dive (~4,500 words / ~18-minute read) with an interactive companion dashboard immediately below the TL;DR. Synthesizes findings from SHRM 2025–2026, Gartner 2025–2026, Mercer GTT 2026, WEF Future of Jobs 2025, Bain, Deloitte, Microsoft WTI, SHL, and 20+ named enterprise case studies.
Disclaimer: All opinions in this essay are my own. They do not represent the company's position. This is meant as an academic discussion of HR transformation in the AI era.

TL;DR

Enterprise HR has moved from AI experimentation to operating-model redesign in 2024–2026: SHRM's 2026 State of AI in HR report finds 39% of HR functions have AI deployed (rising to 60% in extra-large organizations versus 33% small / 35% midsize), and Gartner reports 61% of HR leaders are in advanced GenAI implementation stages with 82% planning agentic AI within 12 months — yet 88% of HR leaders concede they have not yet realized significant business value (Gartner, October 2025), making how HR redesigns work, not whether it adopts AI, the central CHRO question.

The biggest gains are coming not from automating tasks but from reinventing three things at once — HR workflows, the HR operating model, and the broader enterprise organization: IBM reports AskHR cut HR operating cost ~40%, with 94% containment, 75% ticket reduction, and 11.5M interactions in 2024; Mastercard's "Unlocked" talent marketplace (Gloat) reaches ~90% of the workforce with 500,000+ project hours unlocked; Standard Chartered reports ~$49K saved per redeployed role; and JPMorgan, Citi (mandatory AI prompt training for 175,000 employees) and Bank of America (Erica for Employees reaching the bulk of the bank's ~213,000 employees) show finance leading at scale.

The decisive variable is governance and change management, not technology: with the EU AI Act's high-risk-employment obligations expected to apply from August 2, 2026 and NYC Local Law 144 enforcement now active, CHROs who pair AI deployment with bias auditing, human-in-the-loop design, skills-based redeployment and explicit "what to do with time saved" guidance are pulling ahead — only 7% of HR leaders currently provide that guidance, according to Gartner.

Interactive Dashboard

The companion dashboard below is the data layer underneath this essay. It is navigable by tab — Overview, Adoption, Workflows, Case Studies, Operating Model, Skills, Governance, and the seven-stage CHRO Roadmap. The essay below the dashboard is the argument; the dashboard is the evidence.

Dashboard loading slowly? Open the dashboard in a new tab →

Key Findings

1. Adoption is now mainstream in HR but heavily skewed by company size, ownership and industry. SHRM's 2025 Talent Trends found 43% of organizations use AI in HR, up from 26% in 2024; SHRM's December 2025 State of AI in HR 2026 puts current HR-function AI adoption at 39%, with 60% in extra-large enterprises versus 33–35% in small/mid. Publicly traded for-profits lead at 58%; federal government trails at 19%. Gartner's January 2025 data shows 61% of HR leaders in advanced GenAI implementation (up from 19% in 2023) and 82% planning agentic AI within 12 months. Mercer's Global Talent Trends 2026 (n≈12,000) finds 98% of executives planning org-design changes within two years and 65% expecting 11–30% of their workforce to be redeployed or reskilled due to AI.

2. The most mature HR use cases are recruiting, HR service delivery, and learning; experimental territory is performance, succession and sentiment. SHRM 2025: 66% use AI to generate job descriptions, 44% to screen resumes, and 89% of HR professionals using AI in recruiting say it saves time. L&D: 41% report more effective programs, 39% lower costs. Least adopted areas (SHRM 2026): C-suite/board relations, ESG/ethics, future of work, DEI, labor relations, and broader talent management — precisely the domains where AI's judgement, bias and explainability risks are highest.

3. The pilot-to-production gap is the dominant failure mode. Gartner's October 28, 2025 release on a July 2025 survey of 114 HR leaders: 88% say their organization has not realized significant business value from AI tools, and only 7% provide employees any guidance on how to use time saved. Only 17% of HR pros describe their AI implementation as "highly successful" (SHRM), and those that followed change-management best practices were 2.6× more likely to report success. The lesson: technology is not the constraint; redirection of freed capacity is.

4. Workflow redesign falls into three distinct patterns. Automation (Tier-0/1 HR service queries, letter generation, scheduling) is being industrialized by IBM AskHR (94% containment, 11.5M interactions in 2024, 75% ticket reduction since 2016), ServiceNow's HRSD agents, Workday + Paradox (acquisition completed October 1, 2025), and SAP Joule. Augmentation (recruiter screening, manager coaching, performance review drafting, policy interpretation) shows the largest productivity multiplier: Bain's November 2024 release "Generative AI Can Make HR More 'Human,' Not Less" finds AI could save around 15–20% in HR labor time on average, with HR operations teams saving up to 35%. Reinvention — talent marketplaces, skills inference, dynamic workforce planning, ambient employee listening — is rarer but where competitive advantage compounds.

5. The HR operating model is moving from Ulrich three-pillar to a "product + platform + AI-agent" structure. Gartner argues AI solutions are poised to augment 100% and perform up to 50% of HR's current tasks. Bain forecasts COEs becoming "hubs of innovation," HRBPs evolving into business advisors and culture designers, and shared services becoming "frictionless" via AI agents. TI People's research with 15 European enterprises (BASF, Deutsche Bahn, Novartis, thyssenkrupp, Symrise, Brose, GXO Logistics, others) confirms HR is being squeezed simultaneously from the top (demand for integrated business solutions) and bottom (automation of expert tasks).

6. CHROs are increasingly leading enterprise-wide AI transformation, not just HR's own AI adoption. Standard Chartered's Tanuj Kapilashrami was promoted in April 2024 from CHRO to Chief Strategy & Talent Officer, adding Corporate Strategy, Bank-wide Transformation, Brand, Real Estate and Supply Chain to her remit — engaging 38,000 colleagues (40% of the workforce) in future-skills learning in 2024 and saving an estimated $49,000 per redeployed role. Michael Fraccaro at Mastercard built "Unlocked" (Gloat-powered talent marketplace covering ~35,000 employees, $21M productivity in year one, 500,000+ project hours, 90% workforce coverage, 9-point gain in career-development pulse score). At Johnson & Johnson, Peter Fasolo moved roughly two-thirds of HR headcount into a global services org to enable LLM and ML deployment, achieving ~80% internal fill rates for senior management. At JPMorgan, Jamie Dimon publicly said in February 2026 that AI "has displaced people and we offer them other jobs" — backed by "huge redeployment plans" for the 318,512-employee workforce.

7. Cross-industry case-study evidence is now substantial.

8. Skills transformation is now the largest enterprise-AI work-stream, and HR is its owner. WEF Future of Jobs 2025 (1,000+ employers, 14M+ workers): 86% expect AI to transform their business by 2030; 170M jobs created, 92M displaced (net +78M); 39% of core skills will change by 2030; 85% of employers plan internal upskilling; 77% plan to reskill workers to work alongside AI. Mercer 2026: 63% of employees would trade a 10% pay raise for AI/digital upskilling opportunities; 77% of investors are more likely to invest in companies committed to AI education.

9. Governance is no longer optional. From August 2, 2026, AI systems used in employment decisions are "high-risk" under the EU AI Act (Annex III, Category 4) — requiring documented risk management, data governance, technical documentation, transparency, human oversight, post-market monitoring, and conformity assessment. Penalties: €15M or 3% of global turnover for high-risk breaches; €35M / 7% for prohibited practices. NYC Local Law 144 requires annual independent bias audits, public posting and candidate notification. UK Data (Use & Access) Act 2024 introduces automated-decision rules; works-council co-determination in Germany, Spain, Italy, the Netherlands and France adds another layer. Gartner finds 27% of HR leaders believe GenAI has increased their organization's risk exposure.

10. The CHRO leadership gap is the single biggest predictor of disappointing AI ROI. Gartner (December 2025, n=2,986 employees): "AI deployment decisions are often made without any involvement of HR." Only 27% of executives have a comprehensive AI strategy; only 20% believe their workforce is AI-ready. SHRM: only 1 in 4 HR professionals played a leading role in AI implementation, yet two-thirds believe HR should lead change management. Deloitte's 2026 State of AI in the Enterprise (3,235 leaders): worker AI access rose 50% in 2025, yet only 33% redesigned career paths and 30% reimagined organizations based on AI-usage patterns.

The Three Redesigns

The argument behind the data above is that getting AI right in HR requires three simultaneous redesigns, and most organizations are doing one or two and wondering why the ROI is thin.

Redesign 1: HR Workflows. Map every HR workflow against the automation / augmentation / reinvention framework. For each, classify tasks as AI-only, human+AI, or human-only — then rewrite role purposes and skills around the classification. Pick two or three high-volume, low-risk workflows (case management, scheduling, job-description generation) for first agentic deployment. Aim for Bain's 15–20% average HR labor-time savings, with up to 35% on operations teams; benchmark against IBM AskHR's 75% ticket reduction.

Redesign 2: The HR Operating Model. The traditional Ulrich three-pillar model is being restructured along three new axes. An AI-agent layer absorbs the bulk of Tier-0/Tier-1 shared-services volume. HR product teams replace siloed COEs — product owners pair with engineering, data and design, owning value streams (talent acquisition, growth & development, employee service, total rewards). HRBPs evolve into AI-era business advisors coaching managers on human-AI team design, ratio-setting, and change leadership. Gartner's 2024 Productivity Impact of AI Survey: adapting the operating model itself has the highest predicted impact on AI productivity gains (29%) — outpacing AI knowledge sharing, acceptance, or even AI skills. The gains accrue to firms that change how HR is structured, not to those that simply plug in better tools.

Redesign 3: The Enterprise Organization. Two organizational archetypes are emerging. PwC frames them as the Diamond (strong leadership + AI-managed middle + narrow base as AI absorbs entry-level work — JPMorgan's stable-headcount-with-internal-shifts model fits here, as does Amazon's signaled corporate reductions) and the Hourglass (expanded entry-level of AI-literate generalists + expanded specialist tier + thinner middle, designed to preserve apprenticeship and avoid starving the future leadership pipeline). Microsoft's 2025 Work Trend Index names the resulting entity the Frontier Firm — hybrid human-plus-AI-agent teams operating with shifting "work charts" rather than rigid org charts. The rising role is the "agent boss" — an individual who builds, delegates to, and manages multiple AI agents.

Governance and Trust — The Binding Constraint

The compliance perimeter for HR AI is widening fast. EU AI Act high-risk obligations apply from August 2, 2026 (subject to ongoing Digital Omnibus simplification discussions). NYC Local Law 144 enforcement is active and intensifying (the December 2025 NY State Comptroller audit criticized DCWP enforcement as ineffective — a signal that regulatory pressure will increase, not relax). UK Data (Use & Access) Act 2024 introduces automated-decision rules. Colorado AI Act, Illinois AI Video Interview Act, California AB-2930 add state-level requirements. Works-council co-determination in Germany, France, Spain, Italy, Austria, and the Netherlands means HR cannot simply "deploy" AI in Europe without negotiated rollouts.

Per SHL's November 2025 "AI at Work in 2025" release (n=1,009 U.S. adults): 74% say being interviewed by AI would change their perception of a company; only 27% fully trust employers to use AI responsibly; 59% believe AI is making bias worse, not better. Trust is the binding constraint. The CHRO who pairs AI deployment with explicit bias auditing, human-in-the-loop registers, candidate-facing transparency, and works-council engagement is buying the organization permission to scale; the CHRO who does not is buying a regulatory enforcement action or a public-trust crisis, with timing as the only open variable.

Seven-Stage CHRO Roadmap

The full seven-stage roadmap is in the dashboard above (Roadmap tab). The compressed version, for those who want it in one place:

  1. Establish a CHRO-led enterprise AI position (next 90 days). Demand a seat on the enterprise AI governance committee alongside the CIO, CFO and CRO. If you are in the 73% of HR leaders Gartner identifies as not in the AI strategy conversation, treat this as a board-level governance failure.
  2. Redesign HR work before redesigning HR roles (next 6 months). Inventory workflows against automation/augmentation/reinvention. Pick 2–3 high-volume, low-risk workflows for first agentic deployment. Benchmark to 15–20% labor-time savings.
  3. Productize the HR operating model (6–18 months). Restructure COEs as product teams owning value streams. Re-skill the existing HR workforce. Collapse the shared-services structure once AI absorbs >50% of Tier-1 volume.
  4. Lead the enterprise skills transformation (6–24 months). Stand up an internal talent marketplace (Mastercard / Schneider / Unilever pattern); target 25–35% reduction in voluntary turnover from proactive mobility. Pair with explicit redeployment economics (Standard Chartered's $49K/role) so finance sees AI-era reskilling as P&L, not cost.
  5. Build the governance and trust layer (continuous, start now). Map every HR AI system against EU AI Act Annex III. Commission an independent bias audit. Establish a human-in-the-loop register. Halt any high-risk HR AI without current bias audit, no human override, or no impact-assessment documentation.
  6. Own the "what to do with time saved" narrative (next 6 months). Only 7% of organizations provide this guidance (Gartner). Define, by role family, the value-creating activities that should absorb freed capacity. The single most actionable, lowest-cost intervention available.
  7. Measure differently (12–24 months). Replace "AI adoption rate" and "hours saved" with a True ROI Index covering depth and diversity of AI use, time reallocated to value-creating activity, employee AI confidence, voluntary attrition of critical talent, and bias-audit pass rates.

Caveats

Several honest caveats apply to the data and recommendations above. Vendor-sourced metrics dominate — several headline figures (Gloat's talent-marketplace ROI, IBM's productivity savings, Vodafone's SuperTOBI uplift, Microsoft's Work Trend Index) come from companies that sell the underlying tools. They are directionally consistent across multiple independent sources but should be read as indicative ceilings, not guaranteed outcomes. Selection bias in case studies — Standard Chartered, Mastercard, Unilever, Schneider Electric, IBM and PwC publish their results because they are succeeding; median enterprise outcomes are closer to Gartner's 88% "no significant value" figure. Self-reported survey data from SHRM, Mercer, Deloitte and Gartner rely on HR-leader self-assessment, particularly for "highly successful" deployment claims. Forward-looking statements from Microsoft Work Trend Index, Mercer, and WEF are expectations, not realized outcomes. Regulatory timelines may shift — the EU AI Act's August 2, 2026 enforcement date is subject to the Digital Omnibus simplification process; plan to comply on the announced timeline but watch for adjustments. Healthcare and public-sector HR-specific AI outcomes are still thin in primary sources — Cleveland Clinic and Kaiser Permanente deployments are predominantly clinical-administrative; HR-specific quantified outcomes in these sectors lag finance, tech, and consumer goods by 12–18 months in the published record.

The technology has arrived. The supply of CHROs who can lead the three-redesign work — workflows, operating model, enterprise organization — at the same time, with governance and trust as the binding constraint, is the actual bottleneck. The dashboard above is the evidence. The roadmap is the action set. The Katies from last week's essay are the people who will execute it.

Sources

Get the newsletter in your inbox

A weekly, curated digest on AI and organizational design.