Is Selling Company Data to AI Safe? Privacy, Employee & Contract Risks (Checklist)

It can be done with acceptable risk, but it isn't safe by default. Most of the risk isn't in the AI lab. It sits in your own paperwork: client contracts and NDAs, employee notices, California privacy law, professional confidentiality rules, SaaS terms, and how well identifiers are actually removed. Clear each item below with your counsel before you sign anything.

What exactly are you selling?

Under Mode's BizBuyData program, the product is an "agreed copy of company-owned operating records" (email, documents, chat, CRM and billing data) that Mode buys and then resells "for AI research and training," according to Mode's program page. Mode says it "exports and de-identifies the agreed copy" in a secure workspace and removes or transforms names and contact details "before onward delivery." Notice the order: Mode receives the export first and de-identifies it second. Many of the questions below depend on that order.

The best-known precedent shows how quickly this turns contentious. In August 2026, Google won a bankruptcy auction for Spirit Airlines' internal operational data with a $10 million bid. Google's spokesperson told Business Insider it was buying internal data and custom software, not customer or credit card information. The sale still needs court approval, and the hearing was moved to November 6, 2026 after objections from unions and 121 members of Congress, who warned that conventional de-identification may not hold up against AI re-identification (Travel Weekly). micro1 has filed a competing $12.5 million bid. Spirit excluded passenger databases and hired Tonic.ai for de-identification, yet the objections still center on the privacy of employees whose messages are in the data.

One disclosure before the checklist: we're a paid referral partner of Mode, and we earn a fee if your company licenses data through Mode (details). That gives you a reason to read us skeptically. It's also why the risks come first on this page.

Which risks actually matter?

Do your contracts allow it?

Start here, because it's the cheapest risk to find and the most expensive to miss. Mode's page says it reviews "ownership and access terms" before making an offer, and the word that matters in its pitch is company-owned. A great deal of what sits in your inbox belongs, contractually, to someone else.

• Customer MSAs and NDAs. Confidentiality clauses usually cover anything the customer discloses, in any form. Licensing a de-identified copy may still count as a use or disclosure outside the permitted purpose. Search for "confidential information," "permitted use," "aggregate," "de-identified," "machine learning" and "AI."
• Data processing agreements. If you process personal data on a client's behalf, as SaaS companies, agencies and outsourcers do, you are usually limited to the client's instructions. GDPR Article 28(3)(a) requires processors to act "only on documented instructions from the controller" (GDPR). Licensing that data to a third party is not one of those instructions.
• Shared channels and threads. Email threads quote outside senders. Slack Connect channels contain messages from other companies. Supplier portals and shared drives carry other firms' documents.
• Your own commitments. Read your privacy notice, your employee handbook and any customer-facing security promises. Selling data your notice says you don't share is how deception claims start.

If a system can't be cleaned of third-party confidential material, take it out of scope. A smaller deal is better than a breach-of-contract claim.

Are you giving away trade secrets?

Possibly, and this is a risk to the business itself rather than a compliance question. The value buyers describe is "the operational decision-making of companies," as Dan Novaes put it. That's also a fair description of your pricing logic, sales playbooks, escalation scripts and vendor terms. Renaming the people in a thread doesn't hide how you quote a job or win a renewal.

Under the federal Defend Trade Secrets Act, information qualifies as a trade secret only if its owner has taken "reasonable measures to keep such information secret". Handing it to a reseller without tight use restrictions works against that. Decide which systems hold your competitive edge (pricing sheets, proposal libraries, source code, M&A folders) and keep them out, or price them separately. Mode's page says the agreement sets out the rights and restrictions on the filtered copy. Read those terms as closely as the price.

Do you need to tell employees, or get their consent?

No statute squarely answers this for licensed workplace records, and it's the issue that has stalled the Spirit sale. Emory law professor Ifeoma Ajunwa told Business Insider that whether employees' consent extends to resale is "an ethically gray area." The same reporting describes the backlash at Meta, which suspended a keystroke and mouse-tracking program after an internal leak.

Here is what does exist:

• Monitoring-notice laws. New York Civil Rights Law §52-c and Connecticut §31-48d require employers to give notice before monitoring email or internet use. Those notices describe monitoring, not resale, so your existing acknowledgment probably doesn't cover licensing.
• Slack's own gate. Workspace owners can export public channels. Exporting private channels and DMs requires an application, which Slack frames as a tool "for legal and regulatory compliance" (Slack Help). Commercial licensing doesn't fit that description neatly.
• California law (next section) treats employees as consumers with rights over their data.

Our view: give written notice before export, describe the systems and date ranges, and exclude DMs, HR, medical-leave and legal channels unless counsel advises otherwise. Whether you need consent rather than notice is a question for your employment counsel.

What does California's CPRA require?

Two facts change the analysis for any company with California staff. First, the CCPA's employee and B2B exemptions expired on January 1, 2023, so employees' personal information is fully covered (CPPA FAQ). Second, the law applies if your business has more than $26,625,000 in annual gross revenue, or meets one of the data-volume tests (CPPA).

"Deidentified" information falls outside the definition of personal information, but only if three conditions in Civil Code §1798.140(m) are all met:

1. You take reasonable measures to ensure the information can't be associated with a consumer or household.
2. You publicly commit to keep it in deidentified form and not to try to re-identify it.
3. You contractually require every recipient to comply as well.

The harder question is timing. If a raw export goes to the buyer for money and de-identification happens afterwards, ask counsel whether that transfer is a "sale" of personal information, with the notice and opt-out obligations that follow. Outside California, most state comprehensive privacy laws exclude people acting in an employment or commercial context; see, for example, Virginia's definition of "consumer". Mode's referral page lists the UK among the countries it covers, and UK or EU staff bring GDPR into play. See de-identified vs anonymized data.

What if you're a law, accounting or medical practice?

The general rules still apply, and on top of them come professional duties that de-identification after export doesn't satisfy.

• Law firms. "A lawyer shall not reveal information relating to the representation of a client unless the client gives informed consent" (ABA Model Rule 1.6). Matter files, privileged email and work product are off-limits. See our law-firm guide.
• Accounting and tax firms. 26 U.S.C. §7216 makes it a crime for a tax-return preparer to knowingly or recklessly disclose or use tax-return information without authorization. QuickBooks files, client email and billing narratives can all contain it. See accounting firms.
• Medical practices. Selling protected health information generally requires patient authorization (45 CFR §164.502). Data counts as de-identified under HIPAA only through Safe Harbor or Expert Determination (HHS).

Any of these firms should get a written review from firm counsel or an ethics adviser before any export. We don't describe any arrangement as compliant.

Do Slack, Google and Microsoft terms let you export the data?

Generally, the vendors agree the data is yours. Whether a given export method is allowed is a separate question.

The Slack row matters most. Since May 2025, Slack has also throttled message-history calls from non-Marketplace apps to one request a minute (Slack changelog). If a buyer's tool connects to Slack as an outside app, its terms may forbid exactly this use. Mode says your team "signs in to the approved systems" and that Mode handles the export. Ask which method it uses for each system.

How well does de-identification hold up?

It works better on structured fields than on conversation, and no honest vendor will promise 100%. The research is old and consistent:

• Latanya Sweeney estimated that 87% of the US population is likely unique on ZIP code, gender and date of birth alone (Sweeney).
• A 2019 Nature Communications study estimated that 99.98% of Americans could be correctly re-identified in any dataset using 15 demographic attributes.
• The FTC says hashing identifiers does not make data anonymous.

Workplace text is worse than tables. "The only controller in our Boise office who just came back from leave" identifies a person without containing a single name. Mode says it removes sensitive material and transforms identifiers ("Company 01 / Person 01" in its demo), and that it built its own tools after other vendors wouldn't let it audit their PII removal. Mode names no third-party certification or audit. For the mechanics, see data anonymization for AI training.

Call recordings carry their own risk. In Ambriz v. Google (Feb. 2025), a federal judge let California wiretap claims proceed against Google's contact-center AI, partly because Google allegedly had the capability to use call data to train its models (ZwillGen). The case has developed since, but if "calls and transcripts" are in scope, have counsel review your recording-consent language.

What should you ask Mode before signing?

Put these in writing. Mode's Terms of Service say information you submit through the site isn't covered by any confidentiality obligation beyond its privacy policy until a written confidentiality agreement is signed, so sign the NDA first.

1. Raw data custody. Where does the raw export sit during de-identification, who can access it, and when is it deleted? Will you get a deletion certificate?
2. Method. Get a written description of the de-identification method and sample output from your own records, not a demo.
3. Review right. Can you inspect the filtered copy and veto items before onward delivery?
4. Recipients. Which categories of buyers receive the data? Does each contract ban re-identification and onward sale? (CPRA prong three depends on this.)
5. Public commitment. Will Mode publicly commit not to re-identify?
6. Extraction. How is each system exported, and does that path comply with Slack's and other platforms' terms?
7. Your warranties. What do you represent about rights and consents, what do you indemnify, and is your liability capped?
8. Breach. Who notifies whom, how fast, and who pays if the filtered copy leaks downstream?
9. Insurance. Dan Novaes said Mode is "working on" an insurance policy for potential data leaks (LFG Show). Ask for a certificate before you rely on it.
10. Exclusivity and refresh. Mode says the agreement defines scope "and any exclusivity." Find out what it blocks.
11. Eligibility. Mode's page currently lists 6+ employees for legal; 10+ for accounting, finance and medical; and 30+ for others. The Oct 13 release said 20 for general businesses, and Mode's referral page lists different targets. Confirm during intake.
12. Timeline and price. The release says payment is expected within about 60 days. Mode's FAQ now says about three months from first call to payment. The headline on Mode's program page (as of Oct 29, 2026) says company data "could be worth $100K–$5M"; Mode calls its calculator "illustrative, not a guaranteed offer." Nothing about price is guaranteed until a signed agreement says so.

One red flag applies to every buyer, not just Mode: Business Insider reported six-figure "exploding offers" that expire in 48 hours. If anyone rushes your counsel, slow down.

The short version

Selling company data to AI is a contract-and-consent project with a privacy-engineering component, not a quick export. Scope it narrowly, clear your contracts and professional duties, tell your people, test the scrubbed output yourself, and negotiate custody and recipient restrictions in writing. Then compare the number against what's left. For what that number might be, see how much company data is worth. Nothing here is legal advice. Talk to your counsel.