AI Agent TAM Atlas
Software vs Labor Market
across 201 Tracks

A living atlas of where AI agents win

A research framework + interactive dashboard mapping the AI Agent opportunity across 201 sub-tracks in 10 categories, comparing the Software TAM (what AI products charge for) against the Labor TAM (the human-labor spend AI agents can actually disrupt).

Why the labor market is the real ceiling

Most AI market sizing focuses on software TAM — "how much SaaS revenue can AI capture?" — currently $497B globally for AI-agent-relevant products. That's the wrong ceiling. AI agents don't compete with SaaS; they compete with human labor.

The total addressable human-labor spend across the same 201 use cases is $9.6T — a 19.4× leverage ratio. This $9T gap is the structural opportunity AI agents are aimed at: turn human-labor spend into software spend (Net Compression), or expand the market by making expensive expert work cheap enough for new users (Net Expansion).

This is why we built the framework: to systematically rank where the leverage is largest, how much of each labor pool AI can actually replace, and which startups are credibly going after each pool.

The framework as an operating tool

Headline takeaway — out of 201 tracks, the framework identifies 22 high-priority deep-dive tracks (Labor/SW ≥ 50×) where the leverage is structurally largest. These are the first places we look when sourcing.

A guided tour of the live atlas

The dashboard below is the same interactive build available at agent-mapping.vercel.app — embedded here so you can read and explore in one place. Hover charts, click bubbles, expand table rows.

Feature-by-feature walkthrough

"What's Changed" banner (top)

Collapsible panel showing the most recent bimonthly update grouped by type — TAM revisions, displacement updates, capability unlocks, and key players & funding. Click any track name → table auto-filters to that track.

KPI cards (6 metrics, recompute live with filters)

• Sub-tracks — count under current filters
• Software TAM Total — sum across visible tracks ($B)
• Labor TAM Total — sum across visible tracks ($B)
• Avg Labor/SW Ratio — leverage across visible tracks
• ✅ Net Expansion — count where AI grows the market
• 🔴 Net Compression — count where AI shrinks total category spend

Filters

Category · AI Impact (Displacement / Expansion / Mixed) · Net Effect · Valuation Tier · Status · full-text Search.

The four charts

1. TAM Totals by Category (log-scale horizontal bar) — gold bars (Labor TAM) tower over purple bars (Software TAM); the gap visualizes the leverage per category.
2. Software × Labor Scatter (log-log bubble) — each bubble is one track; bubble size = displacement rate; click a bubble to drill into that row.
3. AI Impact Donut — track count by Displacement / Expansion / Mixed, hover shows aggregate Labor TAM per slice.
4. Valuation Tier Donut — distribution of lead-company valuation maturity per track.

The table — the core artifact

Every row = one of the 201 sub-tracks. Click any row to expand methodology, representative companies, funding, and investment insight. Columns:

How each number is calculated

4.1 Software TAM

Definition: addressable revenue for the AI software product layer in the track over the next 3-5 years.

Method:

1. Anchor to Tier-1 analyst reports — Gartner, Forrester, IDC, Grand View Research, MarketsandMarkets — for the broader market.
2. Subtract non-AI-replaceable revenue (e.g., billing systems that aren't agent-driven).
3. Cross-check with bottom-up: estimated user base × ARPU of leading product.
4. Triangulate with company-revenue-multiple for category-defining companies.

Worked example — Code Generation Agent SW TAM ($28B)

Largest single TAM revision in 2026-04 (+133% from $12B), showing all three methods converging:

Reading lesson: when one company's ARR exceeds 10-15% of your published category TAM, your TAM is wrong, not the company.

4.2 Labor TAM

Definition: total annual human-labor spend addressable by AI agents in this track, globally. Built bottom-up via the McKinsey/WEF task-fraction methodology.

Worked example — Tax Filing Assistant Agent Labor TAM ($128B)

1. Identify occupations (BLS): Tax Preparers ~90K + Accountants ~1.4M (of whom ~30% do tax) = 420K tax-active accountants in the US.
2. Wages: Tax Preparers ~$50K avg; tax-active Accountants ~$80K avg.
3. US wage bill on tax work: 90K × $50K + 420K × $80K × 30% + seasonal ≈ $16-17B.
4. Task automatability (McKinsey): tax prep scores 71% automatable. → US addressable $11.7B.
5. Global multiplier: US ≈ 9-10% of global tax-prep → multiplier ~10-11×. → $128B globally.

Cross-validation: ILO global accounting workforce × 25% tax × $30K avg comp × 71% ≈ $75B (lower bound). McKinsey GenAI study: $120-180B annual global productivity unlock from AI in tax/audit by 2030. Our $128B sits at the lower end of McKinsey's range — conservative.

Reading lesson: Labor TAM is bigger than people intuit because it counts the full wage bill the AI agent could replace, not just incremental productivity gains.

4.3 Displacement Rate

The most carefully constructed number in the framework. We never take a single analyst's "AI will replace X% of accountants" claim — those ignore that a job is a bundle of dozens of steps. Instead we build it up step-by-step.

6-step methodology (per 50×+ deep-dive track)

1. Workflow breakdown — decompose the track into 15-25 discrete steps grouped by phase. Example: Contract Review Agent breaks into ~22 steps across Intake / Substantive Review / Negotiation / Closing.
2. Per-step capability assessment — classify each step:
   • FULL — AI handles 80%+ autonomously, output production-ready (80-95%)
   • PARTIAL — AI-assisted; human applies judgment for the last mile (30-70%)
   • NO — Human-only today (0-15%)
3. Weight by hours, not step count — Replacement Rate = Σ(Step% × Hour-share). The high-volume boring steps (clause redlining, cross-references) are usually the most automatable, so weighted rate is much higher than naive count.
4. Triangulate with three external sources — Goldman Sachs Occupational Automation Study, McKinsey Task Automation Fractions, WEF Future of Jobs. If our bottom-up rate sits inside the Goldman-McKinsey-WEF range, we have triangulation; outside → re-examine classification.
5. Report as a range, use conservative midpoint — never a point estimate. Lower bound used for Net Effect modeling, because under-investing in over-claimed AI is the more expensive mistake.
6. Update triggers — upward (new benchmark, model release, production data, YC clustering); downward (regulatory ruling, court decision, published failure case).

Worked example — Code Generation displacement (30-50% → 55-75% in 2026-04)

1. JetBrains Jan 2026 survey: Copilot 29% / Cursor 18% / Claude Code 18% production adoption (not experimentation).
2. Claude Opus 4.6 hit OSWorld 72.7%, GPT-5.4 hit GDPval 83% — agent benchmarks crossing thresholds where end-to-end coding tasks become FULL.
3. Cursor's $2B ARR is empirical proof of value capture, validating the productivity gain.
4. Triangulation: Goldman SDE-automation 47-62%; McKinsey high-adoption scenario 50-70%; our 55-75% sits in the upper half of both → revised up.

4.4 Labor/SW Ratio

Simple division. Categorized:

4.5 Net Effect — three canonical case studies

Decided by asking: if AI cuts this service's cost by 10×, what happens to total demand?

🔴 Net Compression — Contract Review Agent (Labor/SW 80×)

Corporate-legal contract review is a saturated B2B service. Companies don't draft more contracts because review got cheaper — they draft the contracts they already need. The buyer (general counsel) cares about output (signed deal), not input (associate hours). When Harvey reviews a contract in 3 minutes vs. an associate's 3 hours, the GC pays Harvey instead of the law firm.

• Mechanism — 1 senior associate billing $400/hr × 3 hrs = $1,200 of human-labor revenue → captured by Harvey at ~$50 marginal cost. Total category dollars shrink because the service is priced by hours, and the hours collapse.
• Demand is inelastic at the existing margin — corporates aren't sitting on a backlog of contracts they wish they could afford to review.
• Investment thesis — pick the picks-and-shovels winner (Harvey, Legora, Ironclad) capturing redistributed dollars. Don't bet on volume growth — bet on margin capture from labor compression. The absolute software TAM is bounded by the labor pool being compressed.

✅ Net Expansion — CBT / Mental Health Agent (Labor/SW 437×)

There is a massive unmet demand pool in mental health. ~75% of people who need therapy don't get it — most because $200/session is unaffordable, not because they don't want it. WHO estimates ~1B people globally have a mental health condition, while only ~150M see a clinician. AI therapy at $20-30/month doesn't compete with the human therapist's existing clients — it serves the 850M who weren't being served at all.

• Mechanism (textbook Jevons effect) — price drops 10× → demand rises >10× (more like 50-100× given the unmet pool). Existing therapists keep their clients (high-acuity cases) at unchanged or higher rates.
• Total category dollars grow — Labor TAM expands as new spend appears.
• Investment thesis — bet on consumer-grade products that win on UX, brand and clinical safety (Woebot, Wysa, Slingshot). The 437× ratio is durable because the labor pool serves only ~15% of the demand.
• The hard part for investors — regulatory ceiling (FDA digital therapeutics path, state-by-state therapy licensing, suicide-risk liability). Moat is regulatory navigation, not the AI itself.

🟡 Mixed — Advertising Agency Super-Agent (Labor/SW 62.5×)

Marketing & growth has both dynamics simultaneously, and which dominates depends on customer segment.

• Compression at the agency layer — mid-market and large brands historically paid $500K-5M/year for creative + planning + buying. AI super-agents handle 60-70% of that workflow. Big-brand spend goes from $2M → $700K, with the same campaign output → labor compression.
• Expansion at the SMB layer — tens of millions of SMBs globally never had agencies because they couldn't afford $50K/month minimums. AI agency tools at $200-500/month bring a previously priced-out audience in. A coffee shop now runs targeted Meta ads with AI-generated creative.
• Net direction depends on which segment grows faster. Currently: brand-side compression ~$100B agency revenue at risk over 5 years; SMB-side expansion ~$80B new spend unlocked. Net: roughly flat in dollars, but huge redistribution.
• Investment thesis — don't bet on the category as a whole. Pick a side: agency-replacement super-agent (compression, finite ceiling) or SMB-democratization tooling (expansion, larger ceiling). Confused positioning is the failure mode.

Pattern recognition rule: if you can't crisply describe the price-elasticity story for a track in two sentences, you don't know its Net Effect yet — you have a Mixed track and should treat it that way.

The watchlist

Tier 1 — 100×+ tracks (structural gaps)

Highest priority for sourcing. Both compression (legal cluster) and expansion (mental health) plays.

Why ratios fell in legal cluster: the AI software TAM caught up fast (Harvey, Legora, super-agent platforms raised aggressively in 2026-Q1), so the leverage compressed. Still 80×+ — investable.

Tier 2 — 50-99× expansion tracks (largest TAM-growth potential)

• Personal Learning Agent (90×) — Khanmigo, Synthesis
• Precision Agriculture (88.9×) — physical-task gap is hardware-limited
• Drug Discovery (35.4× now, was 55× — software caught up) — Insilico Phase IIa, Recursion
• Investment Institution Super-Agent (51.4×) — Hebbia $700M, Rogo, Aaru $1B
• Music Composition (62.5×) — creator-economy expansion play

Tier 3 — Emerging tracks not yet in framework

Likely additions next cycle (2026-06):

• AI Software Engineering Agent — may split from Dev/Tech category given Cursor $60B / $2B ARR scale
• Healthcare Revenue Cycle Management — Overdrive Health, ClaimGlide
• Supply Chain Intelligence — Pollinate
• Physical Security & Surveillance — computer-vision for commercial security

How this was built

Phase 0 — 2025-Q1: Baseline framework

Established the 201-track taxonomy across 10 categories. Built the Labor-vs-Software TAM lens. Cross-validated displacement rates against Goldman / McKinsey / WEF. Hand-curated representative companies and funding data per track.

Phase 1 — 2025-Q1: Deep-dive sheets

Built per-track 4-section workflow analyses (Workflow Breakdown → AI Capability Assessment → Critical Gap Analysis → Strategic Insight) for the 22 highest-priority tracks (Labor/SW ≥ 50×): 9 tracks at 100×+, 13 tracks at 50-99×.

Phase 2 — 2026-Q1: YC W26 mapping

When YC W26 batch was announced (Demo Day 2026-03-24, 199 companies, 74% AI, 64% B2B), we ran every company through the 22-track filter. Output: YC_W26_TAM_Mapping.xlsx — companies ranked CORE / ADJACENT / INFRA per track.

Phase 3 — 2026-04: First bimonthly update

Established the every-2-months update cadence covering 5 update types (new tracks, TAM revisions, displacement updates, capability unlocks, key players). The 2026-04 cycle produced 30 updates — most material being Code Generation Agent SW TAM $12B → $28B (+133%), legal cluster TAM +45-75%, Drug Discovery clinical milestone (Insilico Phase IIa), humanoid robotics breakthrough (Figure AI $39B + Helix 02 VLA), Q1 2026 macro: $300B record VC quarter, AI = 80%.

Phase 4 — 2026-04: Visualization site

Built agent-mapping.vercel.app to make the framework explorable. Self-contained static dashboard (Chart.js + vanilla JS), auto-rebuilds from the master xlsx via build_data.py. Added the "What's Changed" banner so the bimonthly update is the first thing visitors see.

Phase 5 — Ongoing: Bimonthly updates

Next cycle: 2026-06. Tracked: ratio threshold crossings, Net Effect reclassifications, new players in 50×+ tracks, capability unlocks from frontier-model releases, YC S26 batch mapping when announced.

What's not in the framework

• Geographic granularity — numbers are global aggregates; we don't break out US/EU/APAC separately.
• Time-series — only the latest snapshot in the dashboard; the Change Log shows deltas but not full historical curves.
• Hardware/robotics depth — physical-AI tracks noted as gaps but not modeled at component level.
• Regulatory granularity — jurisdiction-specific risks summarized but not enumerated per region.
• Private market liquidity — valuations reflect last-round prices, not exit comparables.

These are deliberate scope choices. Opening any of them up is a 1-2 month follow-on project.

What we cited

Tier 1 — primary, used directly

Tier 2 — industry analysts (triangulated)

• Gartner, Forrester, IDC — Software TAM anchors per category
• Grand View Research, MarketsandMarkets, Mordor Intelligence — sub-segment market sizing
• Roots Analysis, BioMedNexus — Drug Discovery / Healthcare AI TAM
• Research and Markets — Legal AI Software market sizing (2026: $5.59B)

Funding & company data

• PitchBook, Crunchbase, Crunchbase News — deal-level funding amounts and valuations
• TechCrunch, The Information, Bloomberg, Reuters — major fundraises, M&A, IPO announcements
• Tracxn, Contrary Research — company landscape mapping
• Y Combinator official directory + Extruct.ai — YC batch listings

Industry / professional bodies

• AICPA — accounting market structure & licensing
• ABA — legal market workforce, UPL regulation
• CSA, NAR — real-estate workforce data
• AMA — physician workforce

Regulatory & policy

• EU AI Act enforcement timelines — Compliance Check track drivers
• DORA (Digital Operational Resilience Act) — Compliance / FinServ drivers
• US SEC AI disclosure rules — FinServ AI risk
• FDA digital-therapeutics guidance — Mental Health / CBT track
• US IRS guidance on AI-prepared returns — Tax Filing track

Q1 2026 specific (used in 2026-04 update)

• Crunchbase News (Apr 2026) — $300B Q1 2026 VC, AI = $242B (80%)
• Bloomberg — Cursor $60B / $2B ARR; Decagon $4.5B Series D; Sierra $10B / $150M ARR; Harvey $11B
• TechCrunch — Lovable $9B Series D; Figure AI $39B; Hebbia $700M
• Reuters / CNBC — Insilico INS018_055 Phase IIa (Nature Medicine, Feb 2026); Novartis-Recursion; Novo Nordisk × OpenAI partnership
• JetBrains Developer Survey (Jan 2026) — Copilot 29% / Cursor 18% / Claude Code 18% adoption
• Robot Report — Figure 03 White House education pilot; Agibot 10K shipped
• Vendor docs — Claude Opus 4.6 (Feb 5, 2026), GPT-5.4 (Mar 5, 2026), Gemini 3.1 Pro — all 1M context
• OSWorld / GDPval / GPQA benchmark scores — capability evidence

Peer-reviewed

• Nature Medicine — clinical AI validation (Insilico)
• IEEE / arXiv — model capability papers (computer-use, multimodal)