Working PI
by Isaac Peck, Publisher
The courts are already starting to ask expert witnesses a question most PIs have not yet asked themselves: what exactly are you typing into ChatGPT?
On May 18, 2026, a federal magistrate judge in Connecticut ordered an expert witness to hand over her AI prompts to opposing counsel in a major environmental case. The ruling does not directly govern most PI work, but it signals where courts are heading, and it puts a new layer on every conversation about how investigators use AI in their work.
The PI profession is well past the point of debating whether to use AI. When Working PI hosted a webinar on artificial intelligence for private investigators earlier this year, the live audience answered a poll question before the presentation began: are you using AI tools today? 73 percent said yes. That is up dramatically from just 21 percent who answered affirmatively in the Working PI Nationwide Private Investigator Survey run a year prior in early 2025.
The real questions now are practical. How can investigators use AI to be more productive and efficient in their businesses, all without creating confidentiality exposure, getting blindsided in discovery, or losing the trust of the attorneys and clients who hired them?
Here’s what PIs need to know right now.
How PIs Are Actually Using AI
In May 2026, Working PI hosted a webinar, AI for Private Investigators: A Practical Starting Point, featuring Jay Marin and Chris Reeves, two PIs who are actively integrating AI into their PI work (To view the full webinar, visit WorkingPImag.com/AIwebinar). Over 700 PIs registered for the webinar and nearly 350 showed up live to the event.
Jay Marin is the President and CEO of Specialized Investigations Consultants, a Florida firm with offices in Miami, Lakeland, and Naples. And Marin’s director, Chris Reeves, spent a decade in the U.S. Air Force and another decade in cybersecurity before transitioning into investigations.
Marin and Reeves offered up a variety of concrete use cases for attendees. Report writing is the first one most investigators recognize. Marin described the familiar problem of sitting down at eight or nine at night to write up work done at two in the afternoon, then finishing at one or two in the morning. With the right setup, he says, that same report can be drafted in twenty or thirty minutes.
“AI is not replacing the investigation. They need us for that,” Marin says. “We need a body, a person, a being there. But everything that surrounds the job and the investigation, that’s where AI helps.”
Marin pointed out that for most PI businesses, the time burden of administrative work scales with the size of the operation. The smaller the shop, the more the owner is also the report writer, the email responder, the invoice generator, and the case manager. “When we all started and it was just one or two of us, that was very cumbersome,” Marin says. “We would always be busy trying to get out the work product and then actually conducting the field work as well.”
Reeves walked attendees through their actual workflow. The firm uses Claude for much of their work. “Claude is a much more robust system,” Reeves explains. “To run our business and to use it in our field of investigation, Claude has a lot more features than [other mainstream models].” He builds workflows inside Claude projects, where templates, reporting standards, and case-specific instructions all live in one place. The system knows what kind of report he is writing based on the project, and it produces what he describes as a 90 percent draft in two or three minutes.
Here’s the catch: Reeves isn’t saying AI is doing 90 percent of the work. He’s using AI to help write up investigations he’s already done the work on. By the time Reeves asks AI to help, he’s already done the relationship mapping, fact verification, contradiction analysis, and more. He plugs his structured notes into the tool and it produces a draft in the firm’s format and tone. “It’s not like Claude is doing any research and then taking that and throwing that into a report,” Reeves says. “It’s everything that I’ve done. I’m just using Claude to write it out and summarize it.”
For research and OSINT work, the use case shifts. Reeves uploads documents (public records, business filings, corporate registries) and asks the system to summarize, cross-reference, or flag inconsistencies. He gives the model role-specific instructions on the front end. “I want you to operate as an expert OSINT analyst” is a typical instruction. He tells the model to cite its sources, identify which document and page each finding comes from, and refuse to fabricate or infer beyond what the source material supports.
Marin added that the instructions can go further, specifying the legal and regulatory framework the investigator operates under. “You can tell it, hey, I want you to operate as an investigator compliant with Florida Chapter 493 standards,” Marin says. “That way, anything that is getting pulled or researched is complying with the rules and regulations of what governs us here in the state of Florida for our profession.”
For client communication, the use case is more straightforward. A long surveillance day ends with a frustrated client asking for an update. The investigator knows what to say but does not want to spend twenty minutes finding the right tone. The model drafts an email in two versions, one warm and one firm, and the investigator picks and edits.
Document review is another area where AI can be particularly useful. Reeves describes uploading a forty-page investigative report and asking the model for a structured summary that returned in under two minutes. “These days in 2026, you really don’t need to be sitting there reading that and spending all that time reading that,” Reeves says. The summary becomes a starting point, not a substitute for review, but the time savings on long, dense documents is substantial.
Of course, many investigators are already using AI to draft reports and summarize documents. However, Marin and Reeves have a unique infrastructure and subscription plan that ensures the confidentiality of their data is protected. Marin and Reeves run Claude on an enterprise plan, and they also operate a server in a private data center with over 700 gigabytes of RAM and dedicated GPU cards, which lets them run AI models entirely on hardware they own.
“If you have your own [local model], it’ll help you make sure that all your personal information is kept,” Reeves says. “Nobody’s information is getting uploaded and sent out there.”
This setup points to an exposure most PIs aren’t thinking about but should be: confidentiality.
(story continues)
Confidentiality is Complicated
Confidentiality has emerged as a key concern not only for PIs, but for professionals everywhere whose work carries contractual or professional duties of client privacy. Most public AI tools, especially the free chatbots most people use, handle data very differently from enterprise systems.
When a user types information into a free ChatGPT account, that input is potentially used to train future versions of the model and is retained on the provider’s servers under terms most users never read. The Claude Pro tier at $20 per month allows users to toggle off data sharing, but that protection is a setting, not a contractual guarantee. Real contractual protections (no training on inputs, no retention beyond defined windows, audit rights) kick in at the team and enterprise tiers, where pricing starts at $100 per month minimum.
For a PI working a routine background check, those distinctions may not matter much. For a PI working under attorney direction on active litigation, they matter enormously.
The legal exposure is direct. In U.S. v. Heppner (S.D.N.Y. 2026), a federal court held that communications with a public AI chatbot are not communications with an attorney. The defendant in that case used a consumer (not commercial) chatbot account to ask legal questions. Law enforcement later seized the defendant’s electronics and sought that information. The defendant claimed both attorney-client privilege and work-product protection. The court rejected both arguments, pointing to the platform’s privacy policy, which made clear that input data could be used for training and shared with third parties. By using the public service, the defendant had voluntarily disclosed the information to a nonconfidential party. No privilege attached.
Imagine an investigator working a defense case feeds case details into a free (or contractually unprotected, lower tier) ChatGPT account to brainstorm investigative angles. Without the right subscription and contractual protections, that confidential information now lives on someone else’s servers. Even if the PI strips names and tells himself he has been careful, a stripped prompt can still give the case away.
The PI now has two problems. The first is contractual. Most PI engagement agreements with law firms include confidentiality clauses prohibiting disclosure of case information to third parties without authorization. Feeding case details into a public AI tool is a third-party disclosure. The PI may be in breach of contract before opposing counsel ever finds out.
The second is work-product doctrine. For PIs working under attorney direction, work-product doctrine is the main legal shield protecting investigative materials from discovery. That protection can be waived by voluntary disclosure to a third party outside the attorney’s circle of confidentiality. A free chatbot is exactly such a third party. In other words, a PI who uses an AI tool without proper contractual protections may be waiving work-product protection on all of the work the tool touched. Once the protection is waived, opposing counsel has a path to the prompts and outputs; they can subpoena the provider, compel the PI to produce them in discovery, or use them at deposition.
Marin has thought about this directly. “Who knows what’s really going to happen down the road?” Marin asks. “That’s why for us, we’ve kind of pushed a local large language model and have our own local model so we can keep our stuff internal.” His concern is not just about today’s terms of service, but what might happen to his firm’s data in the future. For example, what happens if ChatGPT or Claude is acquired, changes its policies, or eventually monetizes its archive of user inputs in ways the original users never anticipated.
Those concerns are one of the main reasons that Marin and Reeves are investing in enterprise tools and a local server in a private data center. They want primary control over their data to avoid any confidentiality breaches or work-product challenges.
Expert Witness Ruling
The most significant recent development in this area is Conservation Law Foundation v. Shell Oil Co., decided May 18, 2026, by Magistrate Judge Thomas O. Farrish in the District of Connecticut.
The plaintiff’s expert witness, Dr. Naomi Oreskes, a Harvard historian of science, used AI prompts to help narrow Shell’s massive document production into the subset she relied on for her expert report. The defendants moved to compel production of those prompts as part of expert discovery. The plaintiff resisted, arguing first that AI prompts were not part of an expert’s methodology, and second that a Rule 29 agreement between the parties protected the prompts as “notes.”
Judge Farrish rejected both arguments. On the methodology question, he wrote that “the process by which Dr. Oreskes culled down the defendants’ document production into a subset to be worked with is an aspect of that methodology.” On the Rule 29 question, he held that a generic agreement to withhold “expert notes” was not “quite clear” enough to cover AI prompts specifically. The plaintiff was ordered to produce the prompts.
To be clear, most PIs do not appear in courtrooms to testify as expert witnesses. Thankfully, when PIs work as defense investigators or as consulting, non-testifying experts under attorney direction, they fall under a different framework. The Supreme Court’s 1975 decision in U.S. v. Nobles established that investigative work performed by a defense investigator can qualify as attorney work product, with strong protection against disclosure. The current Federal Rules of Civil Procedure preserve that strong protection for consulting experts under Rule 26(b)(4)(D), which generally shields their work absent “exceptional circumstances.”
This makes it all the more important that PIs operating under the work-product doctrine use contractually tight and confidential versions of AI, so as to avoid voluntary disclosure to third parties—which can void work-product confidentiality protections.
However, for PIs who do testify as experts, or who work closely with the experts who do, this is a clear signal that your AI use is going to be examined the same way your other investigative tools are. The era of treating AI as a back-of-house productivity tool, invisible to the case file, is closing. Document your AI workflow as carefully as you would any other part of your investigation, talk with retaining counsel about how the work was done, and make disciplined choices about what tools to use for what kind of case.
This last ruling is one of the first federal decisions on AI in expert discovery, and of course it won’t be the last. So far, it seems likely that courts will treat AI use like any other piece of investigative work: protected as work product when it’s handled correctly, but exposed in discovery when those protections are waived or when the investigator testifies as an expert.
(story continues below)
(story continues)
The Insurance Angle
As PIs adopt AI tools, liability is an important part of the conversation. As a national provider of liability insurance for more than 14,000 professionals, OREP sees firsthand how new technologies can translate into real-world claims. Bryan Crosco, Senior Underwriter at OREP Insurance, says verification is non-negotiable when it comes to AI. “AI can help a PI work faster, but it can also make it easier to miss errors if investigators assume the technology is always right. If you’re using AI to draft sections of your reports or summarize case documents, double-checking it is mandatory,” Crosco says. “From a liability standpoint, the PI is still on the hook, regardless of how the information was produced.”
The confidentiality piece is just as important. PIs handle some of the most sensitive information in the marketplace, including private financial records, medical information, surveillance footage, witness identities, and case theories developed under attorney direction. A confidentiality breach, even an accidental one, can expose a PI to a lawsuit from their client, a complaint to their state licensing board, and a discovery fight in the underlying case the PI was hired to support. “Confidentiality isn’t just an ethical obligation, it’s a contractual one,” Crosco says. “When a PI types case details into an AI tool that doesn’t have the proper protections, they may be in breach of their engagement agreement before the work is even finished.”
For PIs carrying combined General Liability and Errors & Omissions coverage, the professional liability piece is what responds when an AI-related error in a report becomes the subject of a claim. However, Cyber insurance is becoming more important than ever given how much PIs are relying on technology. Cyber coverage is what responds if a PI faces a data breach, is hacked, or has a system outage that materially affects their business. OREP recommends $1 million in Cyber Liability coverage for PIs handling sensitive client data, though for PIs on a tighter budget, a $100,000 Cyber Liability policy is available for $125 per year.
Building Guardrails
For PIs who want to use AI productively without creating new risk, the practical guidance is straightforward. Start with the right tool. Free, public AI tools should be out of bounds for any work tied to a client matter. The convenience is not worth the disclosure. If the use is genuinely confidential (case facts, witness names, investigative theories, attorney strategy), the tool needs to provide contractual protection, not just a settings toggle. That means a Teams or Enterprise account with documented no-training and no-retention contract terms, or a locally deployed model that runs entirely on hardware your firm controls.
From there, protocols. Marin and Reeves moved AI from an experiment to a formal part of their operations. “The same way that we have standard intake processes and a standard surveillance log, we have a standard AI workflow,” Marin says. Marin’s firm has built repeatable workflows in the four areas where AI saves them the most time, and they specify context, format, and role in every prompt rather than asking the AI for open-ended help.
Before you trust a tool with real client work, test it on fake cases. Both Marin and Reeves run new tools through fictionalized files before putting any real case in front of them. Build a fake case file with fake names, dates, and details. Run it through the AI tool the way you would a real case. See where the tool helps and where it misfires. This lets investigators understand the capabilities and limits of a particular system without ever putting privileged information at risk.
Lastly, verification and review are key. “You still have to verify and review,” Marin says. “It’s a lot easier to review it and make sure it’s right than to do everything from beginning to end.”
Where This Goes From Here
Given how widely PIs are now using AI, the question is no longer whether AI can be useful in a PI’s business, but how PIs can use it effectively and safely, given the confidentiality and discovery concerns emerging in the courts. The May 18 ruling is the first of what will likely be many federal decisions addressing AI in litigation work, and the contours of work product protection in an AI-enabled practice will be worked out in courts over the next several years.
Yes, AI can make PIs more efficient, and the use cases for it are likely to increase. But as Marin and Reeves both pointed out, AI does not replace the investigation itself. It’s up to the PI to verify, check sources, and to bear professional responsibility for what ends up in the final report.
These are exciting times to be an investigator and the tools available to investigators are rapidly changing. The profession has been through technological shifts before, from the typewriter to the database, from the camera to GPS, and PIs who adapt thoughtfully always come out ahead.
This is a developing story. Stay up-to-date on the latest news in your profession by subscribing to Working PI‘s email edition at WorkingPImag.com. Stay safe out there!
About the Author
Isaac Peck is the Publisher of Working PI magazine and the President and Senior Broker of OREP.org, a leading provider of E&O insurance for the PI profession. Working PI is the most widely read print magazine for investigators nationwide, reaching over 25,000 PIs. PIs who become OREP Members enjoy two CE courses (15 hours) at no charge (Visit OREP.org/PI-Members for details). Isaac received his master’s degree in accounting at San Diego State University. Email Isaac at isaac@orep.org or call toll-free (888) 347-5273. CA License #4116465.
We’re always listening. Send your story submission/idea to the Editor: kendra@orep.org.
