Skip to content

The Record · Part 2 of 5

Shadow AI and the Unanswered Coverage Question

Is pasting owner data into a chatbot a covered event?

By Jared Smith, Founder & CEO, SymphonIQ ·  · 5 min read

The Record, Part 2 of 5: Shadow AI and the Unanswered Coverage Question. Is pasting owner data into a chatbot a covered event?

A project manager has a three-week look-ahead due Friday. On his personal account, he uploads the owner’s drawings and the crew roster to a free chatbot and asks it to draft the schedule.

It saves him an hour. Nobody thinks about the owner’s confidentiality clause. (This scenario is illustrative.)

No firewall was breached. No password was stolen. The data simply left the company.

That’s shadow AI: people using AI tools the company hasn’t approved. It may be the most common AI risk on a jobsite today. And whether insurance covers it is still an open question.

It’s already normal

A 2025 global study by KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries. Almost half “admit to using AI in ways that contravene company policies.” That includes uploading sensitive company information to free public tools.

It shows up in breaches

IBM’s annual Cost of a Data Breach report now tracks shadow AI. The 2025 and 2026 reports measured different things, so read them separately:

  • 2025, the added cost: IBM found that high levels of shadow AI added about $670,000 to the cost of a breach.
  • 2026, how often and the total bill: Shadow AI was involved in 43% of security incidents, more than double the year before. Breaches involving shadow AI averaged $5.39 million.

The gap in written AI rules is getting wider. In 2026, 68% of breached organizations had no AI governance policy or were still writing one, up from 63%. Only 29% ran regular audits for unapproved AI.

IBM doesn’t report construction separately. For scale, the average US breach cost $11.5 million in 2026.

Where uploaded data goes

What happens to an upload depends on the account. One term first: training means the AI company uses your content to improve its future models.

Personal (consumer) accounts can keep it. Anthropic’s 2025 consumer terms extended data retention, meaning how long it keeps your data, to five years. That applies to Free, Pro and Max users who allow training.

OpenAI’s consumer terms let it use content to improve its services unless the user opts out.

Business plans are different, up to a point. Both companies say they don’t train on business customers’ prompts and outputs (what you type in and what comes back) by default.

But “no training” covers your content, not everything around it. Take feedback, such as ratings and comments:

  • OpenAI may “use and exploit Feedback without restriction.”
  • Anthropic “may use that feedback.”

OpenAI also classifies business data “to better understand how our services are used.” It says the resulting metadata, the labels this produces, holds none of the data itself.

On personal accounts, your team’s thinking can become training data. Anthropic says shared chats help future models improve at “coding, analysis, and reasoning.”

OpenAI’s individual plans have a catch. A thumbs up or down means “the entire conversation associated with that feedback may be used to train our models,” even after opting out.

So if an estimator rates a chat where they worked through a bid, their reasoning goes with it.

Deleted doesn’t always mean gone

In May 2025, a court ordered OpenAI to preserve ChatGPT logs, meaning keep the chat records rather than delete them. The order came in The New York Times’ copyright case against OpenAI.

That ongoing duty ended Sept. 26, 2025. But logs already saved, and accounts the Times flagged, stayed preserved.

In February 2026, a federal judge in New York ruled in United States v. Heppner. A criminal defendant’s own research in consumer Claude was neither privileged nor work product.

In plain terms, privilege keeps confidential talks with your lawyer away from the other side. Work product does the same for material prepared for a case.

The court pointed in part to the provider’s privacy policy, which allowed disclosure to third parties, including regulators.

That’s one district court, on narrow facts. The lesson is still plain: claims, disputes and delay analyses don’t belong in consumer AI.

Will your insurance cover it?

It’s not clear yet.

Cyber insurers are adding AI language. Beazley’s new “AI Regulatory Defence & Penalties” endorsement is an add-on to a policy. It supports clients whose “own unintentional misuse of AI” leads to regulatory investigations or fines.

That’s a step. But we found no carrier statement on whether an employee’s voluntary upload is a covered privacy event.

What this means for you

The risk isn’t a hacker. It’s a well-meaning employee on the wrong account. And the insurance answer is still blank.

Do this month

  • Publish a list of approved AI tools. Move approved use to business plans with training turned off.
  • Block consumer AI on company devices and networks. Samsung did this in 2023 after engineers uploaded sensitive code to ChatGPT.
  • Keep claims, disputes, bid strategy and legal analysis out of consumer AI.
  • Tell staff that rating a response counts as feedback. On personal OpenAI accounts, it can send the whole conversation to training.
  • Add “unauthorized AI upload” to your incident procedure.

Ask your broker

Send this in writing:

“Is an employee’s or subcontractor’s voluntary input of confidential, personal or client data into a third-party AI tool a covered privacy event? Under which insuring agreement? Do any rogue-employee, intentional-act, unauthorized-software or contractual-liability exclusions apply?”

The insuring agreement is the part of a policy that grants coverage. Exclusions are the parts that take it away.

Next week

Part 3: The AI drew it. Who stamped it?

Part 3 publishes Tuesday, November 3, 2026, at 7:45 a.m. Mountain. Get it by email · Subscribe on LinkedIn · RSS feed

Free resource: Section 6 of our free AI Use Policy for Contractors covers confidential information, privilege and discoverability. Download it as a PDF or Word file. No form to fill out.

Get The Record by email

Each part in your inbox the morning it publishes.

Then the occasional new guide or template for contractors. Unsubscribe in one click.

We confirm your address by email first. Privacy Policy

General information, not legal or insurance advice. Talk to your broker and counsel.

Sources

  1. KPMG and University of Melbourne: Trust in AI global study, Apr 29, 2025
  2. IBM Cost of a Data Breach 2025, July 30, 2025
  3. IBM Cost of a Data Breach Report 2026 (PDF), July 29, 2026
  4. Anthropic: Updates to our consumer terms, Aug 28, 2025
  5. OpenAI Terms of Use, effective Jan 1, 2026
  6. OpenAI Services Agreement, effective Jan 1, 2026
  7. Anthropic Commercial Terms of Service, effective June 17, 2025
  8. OpenAI: Enterprise privacy at OpenAI, updated Jan 8, 2026
  9. OpenAI Help Center: How your data is used to improve model performance, checked October 2026
  10. Engadget: OpenAI preservation order modified, Oct 11, 2025
  11. O’Melveny: United States v. Heppner, Feb 2026
  12. Proskauer: Heppner privilege and work product, Feb 18, 2026
  13. Insurance Journal: Beazley expands AI cyber cover, Sept 25, 2026
  14. Carrier Management/Bloomberg: Samsung bans generative AI, May 3, 2023

About The Record

The Record is written by Jared Smith, founder and CEO of SymphonIQ, which builds governed, single-tenant AI for engineering and construction.

Subscribe on LinkedIn Book a working session
Book a working session