What Data Do AI Companies Buy? Slack, Email, Jira, ERP and CRM Explained

AI companies and the brokers who supply them buy records that show how work gets done: email, Slack or Teams chat, documents, Jira tickets, CRM and support histories, call transcripts, and billing or ERP records. Labs also buy computer-use trajectories and reinforcement-learning environments. What buyers say they don't want is your customers' identities, which are removed or transformed first.

Why do AI companies want internal workplace data?

Because they are building agents that do office work, and the public internet doesn't show how office work is done. Business Insider reported that OpenAI, Anthropic, Meta and others are building AI agents for businesses, and that Gartner estimates annual spending on data for AI will rise from less than $1 billion last year to almost $6.5 billion in 2027.

William Almeida, chief revenue officer of AI training company micro1, told Business Insider that AI developers increasingly want "long-horizon" data that shows how work gets done end to end. If a company won't sell its Slack, CRM or email records, he said, the rest of the corpus is less valuable.

Mode CEO Dan Novaes made the same point on The LFG Show: buyers "don't care about who your customers are." They want "the operational decision-making of companies."

Which systems do buyers say they want?

Here is what five programs publish on their own sites. A check means the system or category appears on that buyer's page; a blank means it isn't listed, not that it's refused.

Sources, viewed October 29, 2026: Mode, micro1 (lists data by business function rather than by tool; chat and email per its revenue chief's comments to Business Insider), Troveo, Mercor (says it supports 100+ tools), Polyshares.

What does Mode's BizBuyData site list?

Mode runs the BizBuyData program; bizbuydata.com redirects to its program page. That page groups what Mode buys into six categories:

1. Email & calendar: "Business conversations and coordination"
2. Documents & files: "Working materials and company knowledge"
3. Chat & collaboration: "How teams solve problems together"
4. Calls & transcripts: "Recorded discussions and decisions"
5. CRM & sales: "Customer journeys and commercial context"
6. Billing & operations: "Invoices, payments, and work systems"

Under "The sources we buy," it names 20 systems: Microsoft, Gmail, Outlook, Google Drive, Slack, Jira, Confluence, Salesforce, HubSpot, Zendesk, Notion, Monday.com, Dropbox, Zoom, AutoCAD, Figma, NetSuite, Paychex, QuickBooks and ServiceTitan, plus "+1K more" (data.mode.inc). AutoCAD, Paychex and ServiceTitan signal that Mode wants trades, payroll and field-service records too, not just tech-company data.

Two process points from the same page: "We agree on the systems and date ranges before any export," and "no passwords are handed over"; your team signs in to the approved systems.

What makes each data type useful?

Slack and Teams. Chat is where problems get escalated, argued over and resolved. It's also where people overshare. Private channels and DMs may hold health disclosures, HR complaints and pasted credentials. Business Insider's headline on this market, "AI labs are fueling a wild market for your Slack DMs," is a reminder that buyers know it. Decide deliberately whether DMs are in scope, and check your workspace agreement and employee policies first.

Email. Email captures commitments to outsiders: quotes, approvals, negotiations, vendor coordination. It's also full of customer names and attachments, which is why de-identification matters most here.

Jira and project tools. Tickets show how a request becomes work: triage, priority, assignment, resolution. They're structured, which makes them easy to learn from. In software companies they can also contain API keys and customer data pasted into bug reports.

CRM and support. Salesforce, HubSpot and Zendesk record how deals progress and how complaints get handled. They're mostly about customers, so ask exactly which fields survive de-identification. When Google won the auction for Spirit Airlines' internal data, a spokesperson said: "We are buying the company's internal data and custom software, but we are not buying their customer or credit card information" (Business Insider). That $10 million winning bid still awaits court approval; the hearing was moved to November 6, 2026 (Travel Weekly).

ERP, accounting and billing. QuickBooks and NetSuite show how money moves: invoices, approvals, reconciliations. micro1 lists "reconciliations, approvals, reporting workflows" among the finance data it wants. These records hold customer financial details and your margins. For accounting and tax firms, client ledgers and tax return information are restricted: 26 U.S.C. §7216 makes it a crime for a tax-return preparer to knowingly or recklessly disclose or use return information. Firm counsel should review before anything is scoped. See accounting firms.

Code. Mode's list doesn't name code repositories. Other buyers want them: DataVendor sells codebases as marketplace listings, and SimpleClosure's AssetHub sells source code from startups that are shutting down. Treat code as a separate IP decision.

What else are AI labs buying besides company records?

The fastest-growing category isn't records at all. It's practice material for agents.

Computer-use demonstrations and trajectories. A trajectory is the step-by-step record of how a task was completed on a computer: what was clicked, typed and checked. Business Insider reported that OpenAI, when it announced a 2025 version of its Computer Use tool, said it had used datasets of people demonstrating how to complete tasks. A Columbia researcher told BI that this kind of data is expensive because it's hard to obtain and clean. micro1 now offers to license "qualifying trajectories" produced by companies' own AI agents "for $1–2M annually, with your approval" (micro1).

Reinforcement-learning environments. These are simulated workplaces, such as clones of business software with realistic data, where an agent attempts a task and is rewarded when the outcome is correct. Scale AI says "nearly half of all of our new data training projects involve reinforcement learning environments" (Scale). Business Insider reported that Google was in talks to invest more than $1.5 billion in Mechanize, a startup that builds virtual work environments to train agents (Business Insider). DataVendor lists RL environments and "tasksets" as products vendors can sell, priced per environment or per task.

Employee activity data. Some tech companies tried collecting it directly. Meta announced a program to capture employees' keystrokes, mouse movements and screen context for AI training, then suspended it after the supposedly anonymized data turned out to be viewable across the company (Business Insider). That's a useful warning for any business licensing its own people's work.

Why does this matter to you? Because your records are inputs to those environments. Real tickets, emails and approvals show builders what realistic work looks like, which is part of why historical company data commands a price.

What should stay out of a license?

• Credentials and secrets: passwords, API keys and tokens pasted into tickets or chat.
• Customer data you hold for customers under contracts or data processing agreements.
• Client-confidential and privileged material at law firms. Consult firm counsel or an ethics adviser.
• Tax return information at tax-preparation firms (IRC §7216).
• Protected health information. HIPAA de-identification requires either the Expert Determination or the Safe Harbor method (HHS); don't assume a broker's general process meets it.
• Employee HR and medical records, and, for California staff, anything you haven't given CPRA-compliant notice about.

Mode says sensitive material is removed and names and contact details are removed or transformed before onward delivery. That reduces risk; it doesn't eliminate it, and no buyer's process should be treated as making records fully anonymous. Exclusions you set at scoping are more reliable than scrubbing afterward.

How should you inventory your data before talking to a buyer?

An hour of preparation makes any conversation with a buyer shorter and your negotiating position stronger.

1. List your systems. Email, chat, documents, tickets, CRM, support, accounting and any industry tools. Note when each was adopted; years of history are a value driver.
2. Count the people. Full-time desk employees who use those systems, by department. This is the first eligibility filter.
3. Flag client-owned and regulated data. Which systems hold customer content under contract, privileged or client-confidential files, tax return information or PHI? Those are scoping decisions for counsel, not for a buyer.
4. Check your vendor terms. Your Slack, Microsoft or Google workspace agreements and any industry software licenses may limit bulk exports or third-party use.
5. Check for legal holds. Records under litigation hold or regulatory retention need special handling.
6. Decide on DMs and private channels before anyone asks, and decide how you'll tell employees.

With that list, you can ask each buyer for a scoped, written offer instead of a range.

What makes company data more valuable?

• Headcount. Novaes said value is driven by "the amount of employees" doing white-collar work. Mode's program page currently lists 6+ employees for legal; 10+ for accounting, finance and medical; and 30+ for others. Its October 13 press release said 20 for general businesses, so confirm during intake.
• Years of history and volume. Mode reviews "years of history, record volume and quality."
• Connected systems. Troveo says "buyers pay a premium for work they can follow across communication channels, knowledge and code repositories, and operational tools."
• Exclusivity. Business Insider reported lower demand for a codebase a rival lab already has.
• Domain. Law and accounting, fields labs are trying hardest to automate, see especially high demand (Business Insider).

What it adds up to in dollars is covered in how much company data is worth. Mode's page headline says a company's data "could be worth $100K–$5M," and Mode calls its calculator illustrative, not a guaranteed offer.

DataLicensingAudit.com is a referral partner of Mode's BizBuyData program and earns a fee if a company we introduce licenses data through Mode (disclosure). For a side-by-side of the buyers above, see AI data brokers compared.