AI and Jobs: What Data Shows for Sales Tech
By Stanislav Chirk11 min read
Headlines say AI will take your job. The data says something else.
Anthropic's March 2026 labor market study introduces observed exposure: not what AI could do, but what it actually does in professional settings. The difference matters for anyone thinking about sales automation.
Executive Summary
94% → 33%
Computer & Math: theory vs observed
75%
Computer programmers — highest AI coverage of any occupation
−14%
Job entry rate for workers 22–25 in exposed roles (post-ChatGPT)
47%
Higher average pay in high-exposure vs zero-exposure occupations
~30%
Workers with zero AI exposure in any task
−0.6pp
BLS employment growth drop per +10pp observed AI coverage
The Situation
Anthropic's study measures observed AI exposure — what Claude actually does in professional settings — not theoretical capability. The gap between the two is the most important number in the report.
Across every occupational category, real usage trails theoretical feasibility by 30–65 percentage points. AI is being used, but far more narrowly than models suggest it could be.
How It's Measured
Three data sources combined into one metric:
- O*NET — task-level data for ~800 US occupations
- Anthropic Economic Index — actual Claude usage in professional settings (Aug + Nov 2025)
- Eloundou et al. (2023) — theoretical LLM feasibility scores per task
Automated use counts fully; augmentative use counts at half weight. Work-related context weighted over casual use.
Key Data Points
- Computer & Math: 94% theoretical → 33% observed (−61pp)
- Office & Admin: 90% theoretical → ~25% observed (−65pp)
- Sales & Related: ~60% theoretical → ~15% observed (−45pp)
- Computer Programmers: 75% coverage — highest of any occupation
- High-exposure workers earn 47% more on average than zero-exposure workers
- ~30% of all workers show zero AI exposure in any task
- No systematic unemployment increase since late 2022
Important Caveat
The study uses Anthropic Economic Index data — actual Claude usage patterns. This measures what Claude users do, not what all AI tools do. The observed exposure figures are a floor, not a ceiling.
The gap between theory and observed usage is where sales automation lives today. Tools that automate structured workflows — config, quote, qualification — are capturing real value. The human remains essential for judgment, relationship, and exceptions.
1. The gap: theory vs reality
AI covers a fraction of what it theoretically could. In Computer & Math occupations, theory suggests 94% of tasks are LLM-feasible — real observed coverage is 33%. The gap is 61 percentage points.
The study combines three data sources to build observed exposure — a measure of what AI actually does in professional settings, not what it could theoretically do:
- O*NET task data — enumerates tasks for ~800 US occupations
- Anthropic Economic Index — actual Claude usage, weighted toward automated and work-related contexts
- Eloundou et al. (2023) — theoretical LLM feasibility scores (β) per task
Result: a measure that weights automated use more than augmentative, and work-related context over casual use. Observed exposure is lower than you'd expect from capability hype.
The Gap in Numbers — Anthropic Labor Market Study, 2026Figure 2
| Occupation | Theory → Observed | Gap |
|---|---|---|
| Computer & Math | 94% → 33% | −61pp |
| Office & Admin | 90% → ~25% | −65pp |
| Sales & Related | ~60% → ~15% | −45pp |
Source: Anthropic, Labor market impacts of AI, Figure 2. Reproduced under fair use.
Top Exposed Occupations
01Computer ProgrammersCoding, debugging, code review — extensively automated via API75% coverage
02Customer Service RepsFirst-party API traffic — high share of automated, work-related usageHigh coverage
03Data Entry KeyersReading source documents and entering data — primary task sees significant automation67% coverage
~30% of all workers have zero coverage — cooks, mechanics, bartenders, lifeguards. Physical and situational tasks remain outside current AI reach.
Who are the highly exposed workers? The demographics are counterintuitive. Compared to zero-exposure workers, the most exposed group is:
- 16pp more likely to be female
- 47% higher average earnings
- ~4× more likely to hold a graduate degree (17.4% vs 4.5%)
- More likely to be white or Asian
AI is not primarily threatening low-wage, low-skill work. It's concentrated in educated, well-paid, knowledge-intensive roles.
2. Who's exposed, who isn't
Participant
High Exposure
Role
Structured, language-heavy, knowledge-intensive tasks
// Gains
- Routine structured tasks automatable — frees capacity for higher-value work
- AI augmentation available now, not theoretical — 75% coverage for programmers
- Qualification, config, proposal drafting: clear AI wins in sales-adjacent roles
- Higher-paid workers (47% premium) — AI tools accessible and already in use
// Risks
- Hiring of workers aged 22–25 into exposed roles down ~14% post-ChatGPT
- Displacement shows as fewer new hires, not visible layoffs — entry-point erosion
- BLS projects −0.6pp employment growth per +10pp observed coverage
- Computer programmers, customer service, data entry most at risk of scope reduction
Participant
Low Exposure
Role
Physical, situational, contextual work
// Gains
- No systematic unemployment increase — data confirms stability through late 2025
- Physical and manual tasks remain outside current AI reach
- Relationship selling, negotiation, complex exceptions: durable human advantage
- Contextual judgment — the moat that current LLMs cannot replicate
// Risks
- Indirect pressure as AI-augmented competitors operate faster and cheaper
- Adjacent structured tasks may shift to AI as capabilities advance
- Cooks, mechanics, bartenders: low exposure today — not a permanent guarantee
- Complacency risk — "low exposure" is a 2026 snapshot, not a long-term forecast
3. No unemployment spike — but watch the hiring data
The headline finding: no systematic increase in unemployment for highly exposed workers since late 2022. But the study flags a more subtle signal that matters more for sales teams.
Hiring of workers aged 22–25 into AI-exposed occupations has slowed by ~14% post-ChatGPT (Figure 7). Job finding rates at less-exposed occupations remain stable at 2% per month; entry into the most exposed jobs has dropped by about half a percentage point. The effect is only observed for workers under 25 — not for older workers.
What the data confirms, suggests, and doesn't show
- Confirmed: No spike in unemployment rates for high-exposure occupations since late 2022
- Confirmed: Observed AI coverage is far below theoretical capability across all categories
- Confirmed: BLS employment growth projections are weaker for higher-coverage occupations (−0.6pp per +10pp coverage)
- Suggestive: ~14% slower hiring of 22–25 year olds into exposed roles — consistent with AI reducing entry-level demand, but just barely statistically significant
- Suggestive: Young workers not hired may be remaining at existing jobs, taking different jobs, or returning to school — not necessarily unemployed
- Not confirmed: Whether the hiring slowdown is AI-driven or cyclical — the study explicitly flags this as an open question
- Not measured: Impact of non-Claude AI tools — the data is Anthropic-specific and likely understates total AI usage
4. What this means for sales tech
Presales and qualification are classic "structured but contextual" tasks — perfect for AI augmentation, not full replacement. The structured part (config, validation, proposal generation) is where the gap lives and where tools are already delivering.
AI Sales ToolCo-Sales.AI
Handles structured presales; human owns relationship and sign-offInterviews the prospect, validates requirements against the product catalog, and delivers a complete quote. The agent handles the structured part; the human handles edge cases and final approval.
01Prospect interview — agent asks qualification questions, captures requirements in structured format
02Catalog validation — requirements matched against product rules, constraints, and pricing logic
03Quote generation — complete, accurate proposal produced without rep involvement in the structured steps
→Rep reviews, handles exceptions, owns the relationship and closing conversation
AI Sales ToolTalkulate
Turns a short conversation into a personalized proposalTurns a short conversation into a personalized proposal. The AI collects context and runs your logic; the rep owns the relationship and closing.
01Context collection — conversational interface captures prospect needs, budget signals, and decision criteria
02Logic execution — your qualification and pricing rules run automatically against the collected context
03Personalized proposal — formatted output ready for rep review, not a generic template
→Rep delivers the proposal, handles objections, and closes — AI handled the structured preparation
The gap between theory and observed usage is where most sales automation lives today. Tools that augment clear workflows (config, quote, qualification) thrive. The human remains in the loop for judgment and trust.
5. Bottom line
01Theory vs Reality
AI's observed impact lags far behind its theoretical capability — by 30–65 percentage points depending on occupation. This gap is the most important number in the Anthropic study and the most underreported finding in AI coverage.
02No Spike — Watch Hiring
No unemployment increase for exposed workers. But hiring of 22–25 year olds into exposed roles is down ~14%. Displacement is happening at the entry point, not through layoffs — and it's only visible in the hiring data.
03Sales: Capture the Structured Gap
Automate config, proposal, and qualification — the structured tasks where AI coverage is real and growing. Keep humans for relationship, negotiation, and exceptions. That's the durable division of labor the data supports.
Bottom Line
That's what the data shows. The hype is ahead of reality — which leaves room for tools that get the balance right. The window to capture the structured gap is open. It won't stay that way.
References and Sources
Primary Source
[1]Massenkoff, Maxim and Peter McCrory. "Labor market impacts of AI: A new measure and early evidence." Anthropic, March 5, 2026.
[2]Handa, Kunal et al. "Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations." Anthropic Economic Index, 2025.
Task Exposure Methodology
[3]Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models." arXiv:2303.10130, 2023.
[4]O*NET Online — Occupational Information Network. US Department of Labor task database used for occupational exposure scoring.
Labor Market Evidence
[5]Brynjolfsson, Erik, Bharat Chandar, and Ruyu Chen. "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence." Digital Economy, 2025.
[6]Gimbel, Martha et al. "Evaluating the Impact of AI on the Labor Market: Current State of Affairs." The Budget Lab at Yale, October 2025.
[7]Hampole, Menaka, Dimitris Papanikolaou, Lawrence DW Schmidt, and Bryan Seegmiller. "Artificial intelligence and the labor market." NBER Working Paper, 2025.
[8]Johnston, Andrew and Christos Makridis. "The labor market effects of generative AI: A difference-in-differences analysis of AI exposure." SSRN 5375017, 2025.
BLS Employment Projections
[9]US Bureau of Labor Statistics. Occupational Outlook Handbook — Employment Projections 2024–2034. Published 2025.