Automation That Eats Its Own.

Key Takeaways

• Anthropic launched enterprise Cowork plugins in February 2026 including a dedicated HR plugin that automates offer letters, onboarding, performance reviews, and compensation analysis, while Deloitte is deploying Claude to 470,000+ employees across 150+ countries with role-specific AI personas and a certified practitioner base of 15,000.

• Anthropic’s own March 2026 labor market research found that 97% of observed AI tasks are theoretically feasible for automation, that job-finding rates for workers aged 22-25 in exposed occupations dropped 14%, and that the highest-exposure workers are disproportionately female by 16 percentage points, mapping directly onto HR’s demographic composition.

• The jagged frontier problem is acute in HR: Claude Cowork handles structured, template-driven tasks like offer letters well, but performance reviews require contextual judgment where AI models have demonstrated measurable gender bias, and compensation equity analysis carries political and organizational nuance that no model can navigate.

• Multiple Claude service outages in March 2026 exposed a new operational risk: when AI plugins become synchronous dependencies in hiring workflows connected to DocuSign, Gmail, and Google Drive, a platform disruption doesn’t slow HR operations, it stops them entirely.

The Automation That Eats Its Own.

The company building HR’s AI future is also documenting its displacement, and the evidence is in the same research library

A senior HR business partner at a mid-size tech company builds a Claude Cowork workflow that connects Google Drive templates, compensation band data, and DocuSign routing. What used to take her team three days per offer letter now takes twenty minutes. She becomes the internal evangelist. Presents the time savings at a leadership offsite. Gets a shoutout in the CEO’s all-hands.

Six months later she takes a role at another company. The workflow breaks the following Monday. IT finds three orphaned API connections, two outdated compensation band files, and no documentation. For the next quarter, her former team processes offer letters manually. Not just slower than the AI-assisted process. Slower than the old manual process, because that process was abandoned months ago and no one remembers the steps. The tool that was supposed to reduce HR’s administrative burden has created a new kind of institutional fragility, one that doesn’t show up on any dashboard until someone leaves.

This is not a hypothetical. It is the kind of scenario practitioners are already describing as Claude Cowork’s HR plugin rolls into production across enterprises. And it is the least alarming thing happening in this space right now.

The Promise: AI Across the Employee Lifecycle

In February 2026, Anthropic launched a suite of enterprise Cowork plugins spanning HR, design, engineering, operations, and finance. The HR plugin is purpose-built to automate core employee lifecycle processes: drafting offer letters, generating onboarding plans, writing performance review frameworks, and running compensation analysis. It connects to Google Drive, Gmail, and DocuSign, creating end-to-end workflows that move documents from draft to signature without a human touching a routing slip.

The product addresses three friction points enterprises had consistently reported: difficulty deploying AI to non-technical staff, insufficient control over who accesses what, and a lack of tools tailored to specific job functions. The HR plugin was designed with practitioners, not just for them, and it targets the administrative work that consumes the majority of an HR generalist’s week.

Deloitte’s deployment makes the scale concrete. Announced in October 2025 and now in active rollout, the partnership puts Claude in the hands of 470,000+ employees across 150+ countries. Deloitte has built role-specific “personas” so that an accountant gets financial analysis AI and a developer gets coding AI, established a Claude Center of Excellence, and certified 15,000 practitioners in generative AI on Anthropic’s models. This is the largest enterprise AI deployment Anthropic has disclosed, and Deloitte’s Trustworthy AI framework is meant to provide the governance rails.

On the surface, this is the HR technology story the industry has been waiting for. A capable AI tool, designed for the function, backed by enterprise governance, deployed at scale. The structured, template-driven tasks that eat HR’s time, the offer letters, the onboarding checklists, the compliance documentation, are exactly the kind of work that AI handles well. Clear inputs. Defined formats. Low ambiguity.

The problem is everything adjacent to those tasks. And the problem is what the company building this tool already knows about what happens next.

The Jagged Frontier of HR AI

Ethan Mollick’s concept of the jagged frontier, drawn from the landmark BCG study where consultants using AI completed 12.2% more tasks 25.1% faster at 40%+ higher quality but were 19% less likely to solve complex problems outside AI’s capability zone, maps precisely onto Claude Cowork’s HR applications.

Inside the frontier, the plugin performs. Offer letters follow templates. Onboarding checklists have defined regulatory requirements. Benefits enrollment documentation has clear formatting rules. These are tasks with right answers, and AI finds them quickly. An HR coordinator who used to spend Tuesday afternoon assembling an offer packet can now have it ready before lunch.

Outside the frontier, the failures are quieter and more dangerous.

Performance reviews sit in the most treacherous part of the jagged landscape. They require contextual judgment about individual contribution, team dynamics, organizational politics, and development trajectory. Research from Textio found that when ChatGPT generated performance reviews, it produced gendered language patterns that tracked occupational stereotypes. Reviews for roles stereotyped as female-dominated used different descriptive language than those for male-stereotyped roles, reproducing exactly the bias patterns that HR has spent two decades trying to train out of human managers.

Compensation equity analysis is worse. The numbers are easy. The interpretation is organizational dynamite. Whether a 12% pay gap between two similar roles reflects market adjustment, tenure difference, or systemic inequity is a question that requires political judgment, institutional memory, and the ability to read a room. An AI model that produces a technically accurate analysis with a tone-deaf recommendation can do more damage than no analysis at all. The CHRO who presents AI-generated comp findings to the board without interrogating the framing isn’t saving time. She’s outsourcing judgment to a system that doesn’t understand what’s at stake.

HR teams deploying Cowork across the employee lifecycle without mapping where the tool excels and where it fails are walking the jagged frontier blind. And the data on what’s behind the fog is coming from an unexpected source: the vendor itself.

The Exposure Paradox

In March 2026, Anthropic published research by economists Maxim Massenkoff and Robert McCrory titled “Labor Market Impacts of AI.” The findings are rigorous, peer-reviewed-quality work based on Anthropic’s Economic Index data, and they paint a specific picture of who AI automation affects most.

Ninety-seven percent of tasks observed across previous Economic Index reports fall into categories rated as theoretically feasible for AI automation. Actual adoption remains substantially lower, with Computer and Math occupations showing 33% observed exposure against 94% theoretical capability, but the gap between what AI can do and what organizations are currently doing is the adoption lag, not the capability limit. That gap will close.

The demographic data is where HR leaders should feel a chill. Workers in high-exposure occupations are 16 percentage points more likely to be female than those in unexposed roles. They are 47% higher-earning. They are more than twice as likely to hold graduate degrees. HR is a female-dominated profession. HR practitioners tend to be educated, mid-to-upper income knowledge workers. The demographic profile of the most AI-exposed workforce segment is, almost perfectly, the demographic profile of an HR team.

The research also found suggestive evidence of a 14% drop in job-finding rates for workers aged 22 to 25 entering AI-exposed occupations. BLS projections show 0.6 percentage points lower employment growth for every 10-point increase in AI task coverage. Entry-level HR coordinator and recruiter roles, the pipeline for every CHRO, fit this exposure profile precisely. If the entry-level roles that train future HR leaders are the first to be automated, the profession doesn’t just shrink. It loses its apprenticeship model.

The paradox is this: the same company shipping the HR automation tool and marketing it as a productivity multiplier also published the research documenting that the function it’s automating is among the most exposed to displacement. This is not a contradiction that Anthropic is hiding. The research is on their website. But the tension between the product team’s narrative (“free HR from administrative burden”) and the research team’s findings (“administrative roles face 90%+ theoretical AI task coverage”) is one that every HR leader evaluating Cowork plugins should sit with.

The Skill Leveler Trap

The BCG study that established the jagged frontier also revealed something that complicates the HR automation story further. Below-average performers gained the most from AI assistance, improving by 43% while top performers gained less. AI acts as a “skill leveler,” compressing the gap between the best and the rest.

In a vacuum, this sounds like good news for HR teams. The generalist who struggles with compensation analysis or onboarding documentation gets a dramatic quality boost. The junior recruiter who can’t write a compelling offer letter suddenly produces polished output. Mollick has noted this effect repeatedly: AI disproportionately helps the people who need the most help.

The trap is what happens to critical evaluation. When an HR generalist who has been producing C-grade compensation analyses suddenly produces B+ work with AI assistance, the improvement feels enormous. It feels like the tool is reliable. The generalist, experiencing the biggest quality jump of anyone on the team, has the strongest experiential reason to trust the AI’s output and the weakest analytical foundation for catching its errors.

Microsoft Research documented the cognitive mechanism in a 2025 study: higher confidence in AI tools was associated with reduced critical thinking effort. Workers who trusted generative AI the most were the least likely to question its output. The shift was behavioral. Workers moved from generating original analysis to reviewing AI analysis, and review is a cognitively thinner activity than generation. You catch less because you’re doing less cognitive work.

In HR, this plays out with specific consequences. The generalist who lets Cowork draft a performance review and makes minor edits is not performing the same cognitive work as the specialist who builds a review framework from scratch. The generalist may not notice that the AI’s language patterns differ by gender. May not catch that the development recommendations for a high-performer are generic. May not realize that the review’s tone is calibrated to a different organizational culture because the model was trained on a different distribution of examples. The quality improvement feels real because it is real. But the error-detection capacity is degrading at the same time, and no one is measuring that.

The Governance Gap

The legal landscape around AI-assisted employment decisions is tightening faster than most HR teams realize. A February 2026 analysis from K&L Gates, written by eleven labor and employment law specialists, maps the regulatory terrain: Colorado’s SB 24-205 takes effect in 2026. California, Illinois, New Jersey, and Texas have all passed or proposed legislation governing AI in employment decisions. The EU AI Act classifies AI used in hiring and performance management as high-risk.

The core legal principle emerging across jurisdictions is straightforward: employers remain liable for outcomes produced by third-party AI tools. If Claude Cowork generates a performance review that contains biased language, the employer owns that outcome, not Anthropic. If an AI-assisted compensation analysis produces a discriminatory recommendation, the company that acted on it bears the regulatory risk. HR technology vendors can now be held liable under agency principles in some jurisdictions, but this supplements rather than replaces employer accountability.

Cross-functional AI governance committees, spanning HR, Legal, IT, and Compliance, are becoming a regulatory expectation rather than a best practice. Organizations that treat AI governance as an IT problem or a legal problem or an HR problem will find that regulators expect it to be all three simultaneously.

And then there’s the institutional knowledge problem. When the senior HRBP who built the Cowork workflow leaves, she takes with her the understanding of why certain templates were chosen, which compensation bands were included, what approval chains were configured, and which edge cases the automation doesn’t handle. IT doesn’t know what she built because it was designed to be accessible to non-technical users. That’s a feature, not a bug, right up until the builder walks out the door and the system becomes an undocumented black box connected to DocuSign with access to salary data.

The governance gap isn’t just about compliance. It’s about organizational memory. Citizen-developed AI workflows are the new shadow IT, except they’re connected to systems that make employment decisions.

The Dependency Risk

On March 2, 2026, Claude experienced a significant service outage. Consumer-facing services went down. Users across the platform reported disruptions, with Downdetector logging thousands of reports. Further disruptions followed later in the month, with response delays and intermittent availability.

For individual users, this was an inconvenience. For HR teams that had wired Claude Cowork into their offer letter pipelines, it was an operational crisis.

When a Cowork plugin connects to Google Drive for templates, pulls compensation data from a connected spreadsheet, generates an offer letter, and routes it through DocuSign for signature, every step in that chain depends on the AI platform being available. An outage doesn’t degrade the process. It halts it. The offer letter that was supposed to go out Monday morning doesn’t go out. The candidate who was expecting a response gets silence. The recruiter who abandoned the manual process three months ago can’t just switch back, because the manual process no longer exists in a form anyone can execute quickly.

This is the operational risk of single-vendor AI dependency for critical HR processes. It’s the same risk that organizations learned about with cloud infrastructure a decade ago, except the blast radius now includes hiring velocity, candidate experience, and the ability to execute on compensation decisions that have legal deadlines.

The March outages are early warnings, not catastrophes. But they expose an architectural question that most HR teams haven’t asked: what is our fallback when the AI agent is unavailable? If the answer is “we’ll figure it out,” the answer is wrong.

What Actually Works

None of this means HR teams should avoid Claude Cowork or similar tools. The productivity evidence is real. The administrative burden is real. The opportunity to redirect HR time from document assembly to strategic work is genuine.

But the path matters more than the destination, and most organizations are taking shortcuts that will cost them.

Start by mapping the frontier before deploying across it. Not every HR task is the same. Offer letters, onboarding checklists, and benefits documentation are inside the frontier: structured, template-driven, low-ambiguity. Performance reviews, compensation equity analysis, and talent calibration are outside it: high-context, politically sensitive, bias-prone. Deploy AI aggressively for the first category. Deploy it cautiously, with mandatory human review by someone qualified to catch what the model misses, for the second.

Build governance before you need it. Cross-functional AI governance committees are becoming a regulatory floor, not a ceiling. Staff them with HR, Legal, IT, and Compliance. Document every Cowork workflow: what it connects to, what data it accesses, who built it, and what happens when that person leaves. Treat citizen-developed AI workflows like you’d treat citizen-developed financial models. The cost of documentation is trivial compared to the cost of an undocumented system making employment decisions.

Invest in the skills the tool doesn’t build. If AI automates administrative HR work, the function’s value shifts to critical evaluation of AI outputs, organizational design, change management, and the political judgment that no model can replicate. Those skills don’t develop through AI use. They develop through deliberate training, mentoring, and practice. Mollick’s framing is right: HR is R&D now. But R&D requires researchers, and researchers require development programs that most HR teams haven’t built.

Maintain operational independence. Never let an AI platform become a synchronous dependency for a process with legal deadlines or candidate-facing commitments. Build and maintain manual fallback processes. Test them quarterly. The March outages were not the last outages.

And protect the pipeline. If entry-level HR roles are the first to be automated, redesign them rather than eliminating them. Junior practitioners need exposure to the full employee lifecycle to develop the judgment that makes senior practitioners valuable. Automating away the apprenticeship creates a profession that can hire experienced people but can’t develop new ones.

The Real Question

Anthropic is not a villain in this story. The company published its labor market research openly and the data is some of the most rigorous analysis of AI workforce impacts available. The Cowork HR plugin was designed with practitioners and addresses real pain points. The tension is structural, not conspiratorial: a company that builds automation tools and a company that studies automation’s displacement effects are, in this case, the same entity looking at the same phenomenon from two sides of its own organization.

The question for HR leaders is whether they will engage with both sides.

The easy path is to adopt the plugin, capture the productivity gains, present the time savings to leadership, and hope the displacement data describes someone else’s function. The harder path is to read the research alongside the product documentation, map the jagged frontier for your own team, build the governance and skills infrastructure that the tool doesn’t provide, and make deliberate choices about which parts of the function you automate and which parts you protect.

The company that built the tool has already told you what the evidence says. The question is whether you’ll read it before or after it describes what happened to your team.