If you've been watching the AI agent announcements over the past six months, you've probably noticed a pattern. Every major AI company is racing to build agents that don't just answer questions — they do the work. They connect to your email, your calendar, your spreadsheets, your project management tools. They plan tasks, execute multi-step workflows, and hand back finished deliverables.
The question every worker should be asking isn't "Is this impressive?" It's "Which parts of my job do these agents actually touch?"
I've spent the past few weeks mapping the agent capabilities announced by OpenAI, Google, Anthropic, and Meta against the task-level data on occupational exposure. Here's what the evidence shows — and where the real risk sits.
What the Four Major Players Are Actually Building
Let me start with the product layer, because the capabilities define the exposure.
OpenAI has shipped two workplace-focused agents. ChatGPT Work, launched in July 2026, is built on GPT-5.6 and Codex technology. It connects to Slack, Google Drive, email, calendars, CRMs, and project management tools. It creates documents, spreadsheets, presentations, reports, and web apps from connected workplace information. It supports "Scheduled Tasks" that monitor updates and prepare recurring reports. OpenAI says nearly 100% of its internal teams now use it — its sales team turns customer discovery notes into a proof of concept in 24 hours, and its finance team cut month-end close from days to hours.
OpenAI Presence, launched the same month, goes further: it's designed for production deployment of agents that "answer questions, resolve incidents, use company systems, execute approved actions, and hand off to humans when needed." Each deployment starts with a specific task — billing issues, insurance claims, IT service requests. The agent gets only the knowledge and system access needed for that task. OpenAI's own English-language phone support channel now resolves 75% of calls without human assistance.
Google is building what it calls the "Agentic Enterprise." At Cloud Next '26, it announced the Gemini Enterprise Agent Platform — a unified system for building, scaling, governing, and optimizing agents. It also launched Gemini Spark, a 24/7 personal AI agent for Gemini Enterprise and Google Workspace customers that can act on your behalf. Google's agent strategy is infrastructure-heavy: it wants to be the platform other companies use to deploy agents across their workforce.
Anthropic has taken a different approach with Claude Cowork, which extends agentic capabilities to the same chat interface people already use. The company analyzed 1.2 million anonymized Cowork sessions from May 2026 and found that the single largest category of use — 33.4% — was "business process and operations": pulling scattered updates into a single report, building onboarding checklists, reconciling spreadsheets. Content creation and copywriting came second at 16.4%. Anthropic noted that roughly half of all usage was "the work around the work" — tasks that span jobs but aren't anyone's core responsibility.
Meta is embedding agents internally. The company is tracking employee mouse movements and keystrokes to train its AI models, with CTO Andrew Bosworth telling staff: "The vision we are building towards is one where our agents primarily do the work and our role is to direct, review and help them improve". Meta's "Hatch" agent is being pushed to employees, though the company has recently softened performance review criteria tied to AI usage after internal pushback.
The Task-Level Exposure Map: Where the Risk Actually Sits
The critical insight from the research is that AI agents don't replace jobs. They automate tasks within jobs. And the tasks they automate best are specific, identifiable, and measurable.
Anthropic's study, based on millions of real workplace AI interactions, developed a metric called "Observed Exposure" to measure how frequently AI performs tasks associated with specific jobs. The findings show that the highest exposure roles are those involving repetitive digital tasks.
Here's the task-level breakdown I've built from the research:
Task type | Automation exposure | What agents do | Who's most exposed |
|---|---|---|---|
Information gathering and routing | Very high | Pull data from multiple systems, summarize, route to the right person | Coordinators, junior analysts, reporting roles, executive assistants |
Document generation from templates | Very high | Produce drafts, slide decks, reports, spreadsheets from connected data | Admin staff, marketing ops, sales ops, junior consultants |
Data entry and reconciliation | Very high | Extract, deduplicate, format, and reconcile structured data | Data entry clerks, medical records specialists, bookkeeping staff |
First-pass research and summarization | High | Synthesize sources, produce briefs, compile competitive intel | Market research analysts, legal assistants, news editors |
Customer support tier-1 resolution | High | Answer routine queries, process returns, update account records | Customer service representatives, call center agents |
Status reporting and updates | High | Monitor projects, compile status reports, flag exceptions | Project coordinators, PMO staff, operations analysts |
Code generation and testing | Medium-high | Write boilerplate, run tests, suggest fixes | Junior programmers, QA testers, technical documentation writers |
Meeting notes and follow-ups | Medium-high | Transcribe, summarize, assign action items | Anyone in meeting-heavy roles |
Judgment-heavy decisions | Low | Provide input only; human decides | Managers, clinicians, lawyers, senior strategists |
Relationship and trust work | Low | No reliable capability | Sales, therapy, teaching, caregiving, community management |
Anthropic's study found that in some high-exposure roles — computer programmers, data entry clerks, customer service representatives, medical record specialists, and market research analysts — up to 67% of tasks could already be automated using existing AI tools. Notably, these aren't "low-skill" jobs. They're cognitive, information-intensive roles.
A Coface study analyzing 923 professions found that approximately one in eight occupations crosses the threshold of 30% automatable tasks — what the study calls "deep transformation." The most exposed sectors: engineering, IT, administrative functions, finance, and law.
The "Middle Lane" Problem: Why the Most Exposed Jobs Aren't the Lowest-Paid
Palantir CEO Alex Karp made a point that's worth taking seriously. The job market is splitting into two lanes that do well — people who build AI systems, and people who do the human work AI can't — and the middle is at risk. Workers in the middle "mainly move, summarize, or aggregate information".
This matches what the data shows. The roles with the highest exposure aren't manual labor. They're the connective tissue of white-collar work: the coordinator who pulls updates into a report, the analyst who compiles competitive research, the support agent who follows a script, the junior marketer who writes first drafts.
Karp's framing is useful because it names the risk clearly. If your day is a pipeline of emails, tickets, meeting notes, spreadsheet clean-up, and turning messy inputs into neat outputs, AI agents can copy the format quickly. The visible output — the summary, the deck, the handoff — is exactly what agents are built to produce.
Gartner's data reinforces this. The share of enterprise applications with embedded AI agents went from under 5% in 2025 to 40% in 2026 — an 8x increase in one year. And the job titles with the highest replacement rates in 2026 include legal document assistants (92%), junior programmers (87%), basic market analysts (85%), and news editors (81%).
These are not jobs that require no skill. They're jobs where the visible output is information processing, and the valuable part — the judgment, the context, the relationship — is invisible to the systems that measure productivity.
The Adoption Reality Check: Agents Are Everywhere, but Trust Is Lagging
Before we over-index on the threat, it's worth looking at what's actually happening inside organizations.
Adoption is near-universal among large employers. Writer's 2026 enterprise survey found that 97% of executives say their company deployed AI agents in the past year, with 52% of employees already using them. But only 29% see significant ROI, and 79% report challenges adopting AI — a double-digit increase from 2025.
Governance is lagging. KPMG's global survey of 2,500 tech executives across 27 countries found that 88% of companies have integrated agentic AI into their systems, but governance and oversight are still catching up.
The midmarket is moving faster than enterprises. IDC's 2026 survey found that 78% of midmarket organizations have moved beyond pilots, 46% have deployed agentic AI across multiple departments (ahead of enterprise at 40%), and 55% report measurable positive ROI.
Meta's internal experiment is instructive. The company pushed agents aggressively, tracked employee usage, and planned 10% global layoffs. But WIRED reported that Meta has now ended its "tokenmaxxing" incentive program — the practice of evaluating employees based on how much they used AI tools. Performance evaluations will no longer depend on AI usage. Some employees said the change freed them from feeling pressured to use AI in situations where it didn't make sense. This is a rare admission that adoption pressure can backfire.
And there's a hard lesson from Meta's own experience. A 36Kr analysis of Meta's year-long agent experiment found that after deploying agents to take over routine work, incidents increased by 40% and engineers spent 70% more time firefighting. The project originally envisioned cutting team sizes by 60%. The reality was messier.
The Five-Level Exposure Framework: Where Does Your Job Sit?
Based on the task-level data, I've built a simple framework for assessing where a job sits on the automation exposure curve.
Level 1: Fully automatable. The task has a single correct output, structured inputs, and clear success criteria. Examples: password resets, order status lookups, data entry, template-based document generation. Agents handle these end-to-end. The human role shifts to exception handling.
Level 2: Mostly automatable with review. The task produces a first draft that requires human judgment to finalize. Examples: first-pass research summaries, competitive briefs, support responses for common issues, standard reports. Agents do the work; humans verify and adjust.
Level 3: Agent-assisted, human-led. The agent handles components but can't own the outcome. Examples: campaign strategy, complex customer escalations, multi-stakeholder project coordination. The human decides what to do with the agent's output and owns the result.
Level 4: Human-led, agent-supported. The core of the task requires judgment, relationships, or accountability that agents can't replicate. Examples: performance management, client relationship building, clinical decision-making, teaching. Agents provide inputs; humans make decisions.
Level 5: Human-only. The task depends entirely on trust, empathy, physical presence, or ethical accountability. Examples: therapy, caregiving, conflict mediation, high-stakes negotiation, anything involving vulnerable people. Agents have no meaningful role.
The risk is concentrated at Levels 1 and 2. If your job is primarily composed of tasks at these levels, the automation pressure is real and immediate. If your job is mostly at Levels 3–5, the pressure is about skill redefinition — learning to work with agents — rather than task elimination.
The problem is that many jobs are mixed. A customer support agent might spend 60% of their time on Level 1–2 tasks (routine queries, account updates) and 40% on Level 3–5 tasks (escalations, emotional de-escalation, judgment calls). When agents absorb the Level 1–2 work, the job doesn't disappear — it becomes a higher-intensity version of the Level 3–5 work. The hours stay the same. The difficulty goes up.
What Workers Can Do With This Information
I'm not going to tell you to "learn to code" or "become an AI engineer." That's not actionable for most people. Here's what is:
Map your own tasks against the five levels. Write down every recurring task in your week. Classify each one. If more than half your time is spent on Level 1–2 tasks, you have a real exposure problem — and you need a plan.
Identify the "invisible" parts of your job. The tasks that don't show up in your job description but matter — the relationship building, the context you hold in your head, the judgment calls you make in ambiguity. Those are your moat. Protect them and make them visible.
Learn to supervise agents, not compete with them. The workers who will do well in the next 18 months aren't the ones who refuse to use agents. They're the ones who learn to brief them well, verify their output, catch their errors, and explain their limitations. That's a skill, and it's learnable.
Track the correction work. Every time you fix an agent's output, log it. Date, task, error, time spent. This is evidence that the efficiency gains come at a human cost — and it's the data you need when performance conversations happen.
Watch the governance gap. If your employer is deploying agents without clear policies on data access, error handling, and human oversight, that's a risk to you. Ask the questions. Document the answers.
What I'm Still Trying to Figure Out
I don't have this solved. Here's what I'm wrestling with:
1. How do we measure "exposure" when tasks are mixed? A job might be 60% automatable and 40% not. Does that make it safe? Or does it mean the remaining 40% becomes unbearable?
2. What happens to the people who can't move up the levels? Some workers are excellent at structured, repetitive work and struggle with judgment-heavy, ambiguous work. What's the path for them?
3. Is the "agent as coworker" framing accurate? Meta says agents will "primarily do the work" and humans will "direct, review, and help them improve." That sounds like a promotion to manager. But managing an unreliable agent that produces "confident garbage" isn't management — it's babysitting.
4. How long is the transition? The Denmark study found zero wage or employment effects from AI so far. But the agent products shipping in 2026 are more capable than anything that existed when that study was conducted. When does task-level automation become job-level displacement?
5. What am I missing? This is the big one. What are you seeing in your workplace that doesn't show up in these reports?
No feedback yet — submit the first.