LEADSUREAI trade operations

Top 10 verticals being bought in AI data licensing deals

Prepared Oct 13, 2026. Companion to buyer-landscape-and-routes-oct13.md and data-licensing-leg-oct13.md — buyer mechanics, referral terms, and route triangulation established there are not repeated except where a vertical's ranking turns on them. This document answers one question: which industries should the finder pipeline hunt in, ranked by evidence that buyers actually pay for that industry's data?

Evidence rules used throughout: company marketing statements are labeled (company claim); figures from named publications are labeled as reporting; anything I could not source is flagged as thin rather than filled in.

Ranking methodology

Verticals are ranked on three things, in this order: 1. Documented buyer demand for that vertical specifically — a buyer naming the vertical in its program, thresholds, hiring, or deal behavior beats generic "all industries" language. 2. Price signals — published payout bands, completed deals, or payout totals attributable to the vertical. 3. A capability story — a plausible account of what the data trains (usually agent/RL-environment performance on that industry's judgment work).

Generic market size (TAM studies, "AI in X" vendor content) was explicitly not used — it measures software spend, not data buying, and it floods search results for exactly the verticals with the weakest licensing evidence (insurance and logistics especially). Where two verticals are close, the text says so.

One structural note before the list: the single most-documented "vertical" in this market is not an industry at all — it is the full linked workflow archive (Slack/Teams + email + tickets + docs) of any documentation-mature white-collar company, priced at $100K–$2M+ for live companies (micro1 materials; Fast Company reporting at https://www.fastcompany.com/91622918/ai-companies-are-offering-businesses-up-to-2-million-for-something-their-employees-create-every-day). Mode's CEO states data is "priced by the terabyte" and white-collar headcount drives value (same source). The verticals below are where that generic demand concentrates, specializes, or gets a named specialist buyer.

The Top 10

1. Legal services

(a) What's bought: Matter workflows minus privileged substance — internal email/Slack about matters, drafting and redlining histories, research memos, appellate and motion briefs in progress, billing and matter-management records (iManage/NetDocuments-type systems), intake and conflict-check processes, and the disagreements/judgment calls around strategy. Seller-side legal guidance already treats privileged material as an exclude-entirely category, which defines the sellable residue: process and work product patterns, not client confidences (https://www.mondaq.com/unitedstates/privacy-protection/1851302/selling-company-data-for-ai-training-what-to-know-before-you-say-yes). (b) Documented buyers: Mode/BizBuyData states outright that "AI labs are prioritizing legal and accounting data" and cut its law-firm eligibility threshold to 6 employees — the lowest bar any program publishes for any vertical (Mode launch release, Oct 13, 2026: https://www.businesswire.com/news/home/20261013463203/en/Mode.Inc-Launches-$5-Billion-Enterprise-Data-Monetization-Initiative-to-Help-Businesses-Nonprofits-License-Data-to-Leading-AI-Companies). micro1 lists "legal and compliance organizations" among its target company examples (https://www.eu-startups.com/job/enterprise-data-partnership-founders/). Mercor's stated verticals lead with legal (see buyer-landscape-and-routes-oct13.md §1B), and its APEX evaluation suite is built on professional tasks in law, medicine, finance, and consulting (https://finance.biggo.com/news/afb2bf61-ddd5-4bf0-92f8-7eeaf19eaefd). Surge AI's environments lead gives legal work as his canonical example of archive value: RL gyms where agents practice "drafting appellate briefs" (MIT Technology Review, Oct 13, 2026: https://www.technologyreview.com/2026/10/13/1145755/ai-labs-buying-data-from-dead-startups-to-train-models/). (c) Price/fee signals: Mode referral: greater of $50,000 or 6% of the deal — on the live-company band ($100K–$2M) a legal deal at the top of the band is worth up to ~$120K to the referrer (Mode release, URL above; Fast Company, URL above). No named law-firm deal price is public — the signal is threshold design, not a closed comp. (d) Why buyers want it: Legal reasoning is high-value, text-native, and verifiable — briefs and motions either persuade or they don't, which makes legal archives ideal RL material for long-horizon professional agents. Law is also one of the four domains labs chose to benchmark first (Mercor APEX: law, medicine, finance/banking, consulting). (e) Supply: Excellent hunting terrain. Legal services are fragmented into thousands of independent firms; the 30–200-employee band is densely populated (regional firms, boutique litigation, insurance-defense, PI mills). Owners can sign. The constraint is not supply, it's the privilege screen — a meaningful share of any firm's archive is client-confidential and excluded. (f) Chad-fit: Strong. His seller-screening value (what's privileged vs. sellable process data) is exactly the diligence buyers want pre-done, and law firms surface through the same accountant/lawyer referral channels he already plans to work. BizBuyData's 6-employee floor also means small firms qualify — the firms a Seattle network actually reaches.

2. Finance & accounting

(a) What's bought: Two distinct products. (i) Accounting/tax firm workflow: engagement letters, workpapers, review notes, client-email threads, tax-season process artifacts in practice-management systems. (ii) Financial-work artifacts: M&A data rooms, diligence files, financial models, decks, filings, underwriting and credit memos — the raw material of finance RL environments. (b) Documented buyers: Mode cut its accounting-firm threshold to 10 employees, paired with legal in the same "labs are prioritizing" sentence (Mode release, URL in #1). Halluminate exists solely to license real financial artifacts (data rooms, filings, models, decks) for finance RL environments, sells to four of the top five closed-source US labs (company claim), raised a $30M Series A on Oct 1, 2026, and is hiring a first Data Platform Lead whose job is to "source, license, and purchase real-world financial data" (https://www.ycombinator.com/companies/halluminate/jobs/aoRoZRG-data-platform-lead, https://www.techtimes.com/articles/328407/20261001/finance-ai-tops-out-51-due-diligence-startup-raises-30m-fix-it.htm). micro1 lists "financial services firms" as a target example (URL in #1). Mercor's APEX includes investment banking (same coverage as #1). (c) Price/fee signals: Same Mode referral math as legal (greater of $50K or 6%). Halluminate's $30M raise to buy data is a capital-commitment signal, not a price. No per-deal finance price is public. Effectively tied with legal at the top — legal ranks first only because its threshold (6) is lower than accounting's (10) and Surge's named example is legal. (d) Why buyers want it: Finance tasks are the canonical long-horizon agent test: multi-document, numerically checkable, high-dollar judgment. Models visibly fail at due-diligence work (the "51%" reference in Halluminate's coverage), and real deal artifacts are the scarcest fix. (e) Supply: Accounting is superb — thousands of independent CPA firms in the 10–200 band, owner-signed, documentation-heavy by professional obligation. Investment-banking artifacts are the opposite: concentrated in few firms and locked down; the findable finance supply is accounting firms, regional banks' operations teams, and boutique advisory shops. (f) Chad-fit: The best fit on this list. Former CPA is not a nice-to-have here — it is the credibility and the screening skill (he can read workpapers and know what a buyer will pay for). His CPA peer network is a direct seller channel into the exact firm size Mode wants.

3. Software & engineering (development workflow)

(a) What's bought: Codebases with their surrounding workflow — commit histories, pull-request reviews, issue trackers, incident postmortems, architecture debates in Slack, CI failures and fixes. Buyers consistently treat code-plus-discussion as worth more than code alone. (b) Documented buyers: Mercor has publicly recruited startup CTOs/co-founders who "can authorize access to a substantial, production codebase" for evals/training, and says some have signed (TechCrunch, Oct 29, 2025: https://techcrunch.com/2025/10/29/how-ai-labs-use-mercor-to-get-the-data-companies-wont-share/). SimpleClosure sells code and data as separate line items and reports 47% of shutdown deals now include code or data (MIT Technology Review, URL in #1). Surge's named RL-gym example #1 is "debugging a complex IT architecture" (same source). micro1 lists "technology companies" first among its target examples (URL in #1). (c) Price/fee signals: The clearest small-deal comp in the market: a 2.5-year-old crypto startup netted ~$25K for its codebase alone, priced separately from ~$15K for its workspace data (MIT Technology Review, URL in #1). At the top end, Spirit Airlines' 516 repos / ~30M lines were a named component of Google's $10M winning bid (Reuters/PYMNTS coverage in buyer-landscape-and-routes-oct13.md §1A; Inc adds that micro1 also bid on Spirit: https://www.inc.com/moses-jeanfrancois/micro1-1-billion-dollar-bet-on-your-companys-data/91417578). (d) Why buyers want it: Coding agents are the labs' most commercialized agent category, and real repos with real review/incident history teach the parts synthetic code can't: legacy constraints, trade-off arguments, and how bugs actually get found. (e) Supply: Deep but contested. Every dead startup has a repo; SimpleClosure/Sunset already work this supply, and GitHub-native companies are the most-approached sellers in the market. Differentiation is in the linked workflow (issues + chat), not the code. (f) Chad-fit: Good via his tech/agency network and Seattle's startup density; weaker than finance/legal because he has no special diligence edge on code quality — buyers' engineers screen this, not accountants.

4. Healthcare — two markets, one compliance wall

(a) What's bought: Two different products that are often conflated. (i) Clinical data: de-identified patient records, medical imaging, audio — sold through governed marketplaces, not as raw company archives. (ii) Healthcare operations workflow: billing/RCM processes, prior-auth correspondence, scheduling and intake operations, provider-office administration. (b) Documented buyers: Protege is the specialist: it licenses de-identified health records, imaging, and audio from "hundreds of providers" (company claim) with revenue share, works with "the majority of the Magnificent Seven" (company claim), and raised a $30M Series A extension in Jan 2026 ($65M total, a16z-led) on "rapid adoption across healthcare, media, audio, motion capture" (https://siliconangle.com/2026/01/08/protege-raises-30m-grow-governed-marketplace-ai-training-data/, https://markets.financialcontent.com/kelownadailycourier/article/bizwire-2026-1-8-protege-raises-30-million-led-by-a16z-to-unlock-access-to-data-for-ai-development). micro1 names "healthcare administration teams" as a target example (URL in #1) — note: administration, i.e., product (ii). Mercor's APEX includes primary medical care, and its vertical list includes healthcare (buyer-landscape-and-routes-oct13.md §1B/§2). (c) Price/fee signals: No public per-deal price for either product. Protege's raise and provider count are scale signals only. Historical context (pre-AI-licensing-wave reporting, treat cautiously): de-identified record sets at population scale have traded at very large values (a ~50M-record acquisition reported around $1B in 2021, via http://dev.to/velocityai/medical-data-as-commodity-why-your-health-records-are-the-most-valuable-training-data-16bh — a secondary blog source; flagged as weakly sourced). Evidence for demand is strong; evidence for price is thin. (d) Why buyers want it: Clinical reasoning is a flagship benchmark domain (APEX medicine) and real longitudinal records are genuinely scarce — they can't be scraped or synthesized credibly. (e) Supply: Structurally hard. Clinical data sits in health systems and EHR-governed entities with HIPAA duties; the sellable small-company supply is mostly product (ii) — RCM firms, billing services, home-health and clinic operators' back offices — plus de-id-ready datasets held by data companies, which Protege already aggregates. (f) Chad-fit: Moderate. His lane is the administration slice (billing/RCM firms look like accounting firms operationally, and his screening skills transfer). Clinical data itself is a specialist game he should refer into Protege rather than work directly — and seller-side guidance already flags health/benefits data as exclude-or-de-id-first (Mondaq, URL in #1).

5. Professional services, consulting & BPO operations (the general white-collar archive)

(a) What's bought: The canonical archive itself: SOPs, playbooks, knowledge bases, project-management histories, QA processes, decision records, CRM and workflow metadata, task-execution histories (micro1's published data list: https://theoutpost.ai/news-story/micro1-commits-1bn-to-buy-company-data-for-ai-agents-training-31971/). (b) Documented buyers: This is micro1's core program design — target examples are "consulting businesses," "professional services businesses," "BPOs & recruiting firms" (URL in #1) — backed by a $1 billion, 12-month commitment to license enterprise operational data, financed by Citi and Hercules Capital (Business Wire, Oct 8, 2026, via https://pulse2.com/micro1-commits-1-billion-to-license-enterprise-ai-training-data/). Mode's program ($5B initiative headline — company claim) covers the same ground at 20+ employees (Mode release, URL in #1). Handshake ($10K flat referral) and Scale (40+ employees) buy the same product without vertical specialization (terms in referral-program-inventory-oct13.md). (c) Price/fee signals: The market's best-published bands live here: micro1 partnerships pay >$100K qualified, >$500K for large datasets, >$1M for highly unique data (company claims reported at https://theoutpost.ai/news-story/micro1-commits-1bn-to-buy-company-data-for-ai-agents-training-31971/); Mode advertises $100K–$5M per company (company claim); Chad's own Ambaum deal ($175K agreed, ~20–40-person services firm) is a real datapoint inside the band (data-licensing-leg-oct13.md). (d) Why buyers want it: This is the RL-environment feedstock: how offices actually route approvals, escalate exceptions, and exercise judgment under incomplete information — micro1 says exactly this in its release. (e) Supply: The deepest of any vertical here — every mid-size services firm, agency, MSP, and BPO qualifies structurally. The screen is documentation maturity, not industry. (f) Chad-fit: Strong — it is literally his own seller profile (Ambaum), and his agency/e-commerce network is dense with 30–200-person services firms. Caution from his own research: agency archives skew client-owned; his screening is the value-add.

6. Media & content (video, audio, text rights)

(a) What's bought: Licensed content corpora — video footage and broadcast archives, audio (podcasts, call-free recordings, music), publisher text, gameplay data — cleared for training, usually exclusive or semi-exclusive. (b) Documented buyers: Troveo, the specialist marketplace: 7,000+ rights holders, 95% signed exclusively (company claim), 8M+ hours of video, and an Apr 2026 expansion into audio (4M hours), text, gameplay, and "agentic trajectories" (enterprise workflow) (Business Wire, Apr 28, 2026: https://www.businesswire.com/news/home/20260428319383/en/Troveo-Accelerates-AI-Model-Development-Expands-AI-Training-Data-Platform-to-Five-New-Categories-Announces-$20-Million-in-Payouts). Protege also aggregates media/audio (URL in #4). (c) Price/fee signals: The hardest payout total in the entire licensing market: Troveo has paid out >$20 million to content owners (company-reported, same release). At the top of the market, publisher deals are large but concentrated: OpenAI–News Corp reported at >$250M over five years and Reddit's disclosed licensing at $203M aggregate, with an industry observer noting mid-size publishers realistically see "low hundreds of thousands a year at best, and often nothing" (industry commentary, Sep 6, 2026: https://www.linkedin.com/pulse/retrieval-licensing-training-ai-deal-structure-nobodys-menon-zpcoc — commentary, not reporting; treat the deal figures as reported-elsewhere claims). (d) Why buyers want it: Multimodal training (video understanding, ASR, world models) needs rights-cleared real footage that cannot be scraped legally at scale. (e) Supply: Rights holders, not operating companies — studios, production companies, creators, publishers. Aggregation is the game and Troveo/Protege already play it at scale; a finder competes on relationships with unrepresented rights holders (regional broadcasters, production libraries, podcast networks), not on company-hunting. (f) Chad-fit: Weakest of the top six for the finder model — with one exception: his content-business history (ContentRunner) and agency network touch exactly the mid-size content producers Troveo signs. It's a side door, not the main hunt. Ranked this high on payout evidence alone; on Chad-fit alone it would be ~#9. Flagged as the list's most methodology-sensitive placement.

7. Customer support operations

(a) What's bought: Resolved ticket histories (Zendesk/Intercom/Help Scout), chat and email threads with outcomes, QA scorecards, escalation paths, help-center corpora — the intent→resolution pairs agents learn from. (b) Documented buyers: Support data is a named component of the archive product: Scale's data-partnership materials name Salesforce/Zendesk sources (buyer-landscape-and-routes-oct13.md §2 ranked call/support archives #6 with "no standalone price evidence"). Demand for support automation is proven at scale — Zendesk says its AI agents were trained on 19 billion tickets (vendor claim about its own corpus: https://www.techtarget.com/enterprise-software/news/366613094/Zendesk-upends-pricing-for-AI-in-customer-service) — but that is a platform training on its own hosted data, not labs buying third-party support archives. (c) Price/fee signals: None standalone, as of this research. Support archives monetize as part of a full workflow archive (bands in #5), not as a priced category. This is the thinnest price evidence in the top 10. (d) Why buyers want it: Support is the most measurable agent task (resolved/not, CSAT, reopen) — ideal RL reward structure — and every services company generates it as exhaust. (e) Supply: Enormous and usually bundled: any BPO, SaaS company, or e-commerce operator has years of tickets. Standalone sellers are rare because the tickets live inside a broader archive. (f) Chad-fit: Good as a component pitch — his e-commerce network's support histories ride along with their ops archives — but there is no support-only deal to hunt. Rank reflects component value inside #5-type deals, not a standalone market.

8. Insurance (claims & underwriting workflow)

(a) What's bought: Claims files and adjuster notes, FNOL intake records, coverage-interpretation correspondence, underwriting submissions and decline/accept rationales, fraud-flag histories — judgment-dense, document-heavy workflow. (b) Documented buyers: Mercor's stated verticals include "finance & insurance" (buyer-landscape-and-routes-oct13.md §1B). Beyond that: no insurance-specific program, threshold, specialist buyer, or deal surfaced in this research. Targeted searching returned only AI-for-insurance vendor content (claims-automation startups, implementation consultancies), which evidences insurer software spend, not data buying. (c) Price/fee signals: None found. Zero. (d) Why buyers would want it: Claims adjusting is the textbook "tricky judgment call" archive Surge's environments lead describes wanting (MIT Technology Review, URL in #1) — coverage calls are documented, consequential, and reviewable. The capability logic is strong; the market evidence has not caught up. (e) Supply: Structurally awkward: claims data concentrates in carriers and large TPAs (not 30–200-employee signable firms); the findable supply is independent adjusting firms, MGAs, and insurance BPOs — real but thin. (f) Chad-fit: Moderate via professional-services adjacency (adjusting/BPO firms resemble his services-firm targets). Verdict: a watch-list vertical, ranked on one buyer's vertical list plus capability logic — the weakest evidence in the top 10, and effectively tied with #9–#10.

9. Logistics & transportation operations

(a) What's bought: Shipment and exception histories, dispatch decisions, rate/pricing negotiations, carrier communications, TMS records — and at the extreme, an entire airline's operational archive. (b) Documented buyers: micro1 lists "logistics and operations businesses" among its target examples (URL in #1); Mercor lists logistics among its verticals (buyer-landscape-and-routes-oct13.md §1B). The Spirit Airlines dataset — the market's biggest completed deal — is substantially a logistics/transportation operations archive: flight and aircraft operations logs and pricing data on 7.2 billion competitor flights were headline components (Inc, URL in #3). (c) Price/fee signals: Spirit's $10M is a bankruptcy mega-deal, not a vertical rate; no logistics-specific payout band exists. General archive bands (#5) are the only pricing guide. (d) Why buyers want it: Operations under disruption (weather, capacity, exceptions) is judgment work with measurable outcomes — prime RL-environment material, and micro1's release language about "exceptions that require judgment" reads like a logistics brief. (e) Supply: Freight brokerages and 3PLs are fragmented and owner-operated at exactly the target size — good hunting structure — but documentation maturity varies wildly and much workflow lives in phone calls and TMS notes. (f) Chad-fit: Moderate. PNW has real freight/3PL density (Seattle/Portland), but Chad has no special channel here; it competes for his time with better-fit verticals.

10. Real estate

(a) What's bought: Two different things: (i) structured listing/transaction data (MLS data) licensed for AI training/retrieval, and (ii) brokerage operations archives (transaction coordination files, agent communications, CRM histories). (b) Documented buyers: For (i), real evidence: Bright MLS (100,000+ subscribers) reports it is "already in talks with vendors that want to train AI models on its listing data" and has built a licensing structure treating AI training as a separate addendum to its data agreements (RISMedia, Sep 3, 2026: https://www.rismedia.com/2026/09/03/how-ai-tools-are-exposing-mls-data-and-your-brokerage-to-risk/). For (ii), Mercor lists real estate among its verticals (buyer-landscape-and-routes-oct13.md §1B) — and nothing else surfaced. (c) Price/fee signals: None published for either product. MLS licensing terms are negotiated privately. (d) Why buyers want it: Listing/transaction data grounds real-estate AI (valuation, search, agent copilots); brokerage workflow is ordinary white-collar archive material. (e) Supply: Product (i) is controlled by ~500 MLSs — a handful of gatekeepers, not a finder market. Product (ii) supply is plentiful (independent brokerages and transaction-coordination firms at target size) but undifferentiated from #5's general archive. (f) Chad-fit: Low-moderate. His e-commerce/agency network doesn't map to brokerages; this is a "take it if it walks in" vertical, not a hunting ground. Effectively tied with #8 and #9 — three different weak-evidence shapes: named-on-a-list (insurance), named-plus-one-big-deal (logistics), and a real-but-gatekept market (real estate).

Just missed (11–15)

  1. Manufacturing / industrial — Named on Mercor's vertical list and plausible (maintenance logs, QA records, engineering-change workflows), but no program, deal, or price evidence found; much plant-floor workflow is still paper/vendor-system-bound, which fails the documentation test.
  2. Sales & marketing operations — Commonly assumed hot (see below); in practice the sellable asset (CRM/call history) is buried inside #5 archives, and no buyer names sales data as a sought category.
  3. Education — Content licensing is real (scholarly publishers report meaningful AI licensing revenue — Wiley disclosed $49M for FY2026, via the commentary at https://www.linkedin.com/pulse/retrieval-licensing-training-ai-deal-structure-nobodys-menon-zpcoc), but that is publisher content, not school/tutoring-company workflow archives; no broker program targets education companies' operations data.
  4. Construction — Zero licensing-demand evidence found. Project data is fragmented across owners/GCs/subs with contested ownership — the exact client-data problem, squared.
  5. Hospitality & energy (tie) — Zero evidence in either. Hospitality workflow is high-turnover and systems-light; energy data is concentrated in utilities/majors with regulatory overlays. Neither appears on any buyer list found.

Commonly assumed hot — but the evidence does not support it

What this means for the hunt (one paragraph)

Hunt in this order: accounting firms and law firms first (named prioritization, lowered thresholds, Chad's CPA credibility, fragmented signable supply), services/consulting/BPO archives second (deepest supply, best-published price bands, his own seller profile), dev-workflow startups third (real code comps, contested supply), and treat healthcare-admin, logistics, real estate, and insurance as opportunistic — take them when the network surfaces them, don't spend sourcing time there yet. Media/content is a real market with the biggest documented payouts, but it runs on rights-holder relationships, not company hunting; it is a separate muscle (possibly via his content-business past) rather than part of this pipeline.

‹ Daily Deals  ·  Data Licensing Pipeline ›