It is 4:51 p.m. on a Friday. The VP of Sales from the companion piece is in her last twenty minutes. The laptop goes back at 5:00. The badge already went back at noon. HR has run her through the standard exit checklist: laptop, monitors, charger, badge, MFA token, desk key, return of confidential information, signed certification.
She has not signed anything about her skill. Nobody has asked. [S: as my toddler would say: oopsie doopsie, BIG ONE.]
At 4:53 she opens her personal Claude account on her phone. The same one she has been using on weekends for two years to refine the deal-memo skill. The skill is already there. The system prompt. The custom instructions. The library of 3,000 Gong call transcripts. Four years of deal memos. The price book. The redlines.
At 4:54 she creates a new project. She copies everything over.
At 4:58 she logs out.
At 5:00 she returns the laptop to the receptionist.
On Monday she starts at a competitor. On Tuesday morning, on her new firm-issued laptop, she logs back into her personal Claude account and queries the skill. It is calibrated to her old company’s customer base, priced against her old company’s price book, and tuned to her old company’s objection handling.
You did not put any of that on the exit checklist.
This is the story of every senior departure in 2026 where the employee had thirty minutes and a phone. It is happening right now at yours.1
What the skill actually contains
The system prompt names your enterprise pricing tiers, your top three competitors, and the canonical objection-handling response your CRO drafted in 2024.
The custom instructions reference your sales methodology, your discount approval matrix, and your contract redline preferences.
The library is 3,000 call recordings, four years of deal memos, and 14 contract redlines with markup attributable to your specific positions.
The retrieval embeddings encode the same information statistically. Even if she “deletes” the source files later, the index still works.
The skill is a complete extract of your sales operation, packaged for query. Not a memory aid. A data product.
Four exposures she just created
The DTSA. 18 U.S.C. § 1836 gives you a federal civil cause of action for misappropriation. Your customer lists, pricing, and contract positions are trade secrets. Moving them from a system you control to one you do not is textbook misappropriation. Statutory damages, exemplary damages up to two times actual, attorneys’ fees, civil seizure remedy. The DTSA does not care that she has it on a personal account. It cares that she has it. [M: misappropriation doesn’t care which account it’s on.]
The Economic Espionage Act, criminal version. 18 U.S.C. § 1832. Federal felony to knowingly steal a trade secret with intent to convert it to the economic benefit of someone other than the owner. The U.S. Attorney’s office will not bring most of these. They will bring the ones with a clean factual record. Yours may be cleaner than you think.
The Computer Fraud and Abuse Act. 18 U.S.C. § 1030. After Van Buren, “exceeds authorized access” turns on whether the employee accessed gated resources. The case law for ongoing personal-account access containing company data is thin. Your facts may make it.
Breach of contract. The PIIA. The confidentiality agreement. The acceptable-use policy. The exit certification. Each is a separate promise. Each is breached when the library crosses the line. Stack them with the DTSA claim.
Four theories. Same fact pattern.2
“I just took my brain”
The defense fails for the skill.
The skill is not unaided memory. It is aided memory. The aid is the model. The library is the prompt material. The system prompt is the calibration. None of it is in her head. All of it is in a system she trained against company-owned data.
A useful test. Could she reproduce the skill, the same outputs, the same calibration, by sitting in front of a blank prompt window and typing from memory? No. The skill exists because the library exists. The library is the company’s. [S: there’s an interesting hypo in here for someone with a photographic memory.]
That is not a residuals argument. That is a misappropriation argument with a UI on top.
The detection problem
You will not find it on her laptop. Her laptop is back. The forensic image will show normal end-of-employment activity. Nothing actionable.
You will not find it in your enterprise admin console. The skill never lived there.
You will not find it in your DLP. The skill moved by being authored slowly over two years on a personal account that was always outside the perimeter. [M: most DLP hasn’t caught up with AI.]
You will find it, if at all, by watching her at her new employer. By the cadence of how quickly she gets up to speed. By the pricing patterns in deals where you both bid.
By the time you find it, the cause of action is two years old, the remedy is depreciated, and the library is on a third platform.
If your company has had a key technical employee depart in the last year, the skill transfer to the next employer is the asset that left, not the documents.
Talk to a Talairis attorney →What her new employer just inherited
The receiving company is in worse shape than it knows. It has a new VP whose first deals will be informed by a tool trained on your data. The model does not emit attribution. The new employer will not know which outputs are clean and which are trade-secret-derivative.
When you sue under the DTSA, you will name the receiving employer. They will say they did not know. They are now subject to a discovery process that examines every output, every chat thread, every deal she touched. The cost of failing to ask a candidate “did you bring a skill” is six figures of outside counsel and an injunction.
What to do
The cemented answer: update the AI use policy and the PIIA before the next departure. The drafting is small. The avoided exposure is not. This is the only lever that scales.
Beyond that, options.
Don’t allow personal AI accounts for work in the first place. Add the skill to the exit checklist by name. Pull the access log. Issue a preservation letter to the employee, the new employer, and the vendor. Hold the separation agreement until the audit closes. Sue early if the facts warrant the DTSA seizure remedy.
For the receiving employer: ask every senior hire whether she has built or trained an AI artifact in her last role, in writing, and quarantine personal AI accounts for the first thirty days. The candidate who refuses to answer is telling you something.
Get counsel before the next exit interview
Before the next senior departure, before the next executive transition, before the next layoff that removes anyone with admin access to a library, get counsel in the room. Bring the AI use policy. Bring the PIIA. Bring the exit checklist. Bring the access log. Update them once, correctly, before the next person is in her last twenty minutes on a Friday afternoon with a phone.
The artifact you are not auditing for is the artifact that walks out the door.
A closing thought
Why does the skill walk out the door?
Because nobody put it on the checklist.
Because the artifact has no physical footprint and the perimeter has no sensor for it.
Because the receiving employer did not ask.
Because the model does not emit attribution.
The skill leaves on Monday. It will be on every deal she touches for the next nine months. The first time you find out is when a customer mentions a price quote that maps too cleanly to your matrix.
By then the library is on the third platform.
Welcome to the exfiltration nobody told you to look for.
- Most exit interviews still run on a 2018 checklist. Laptop, badge, MFA token, return of confidential information, signed certification. Nobody asks about the personal Claude account. Nobody asks about the project. Nobody asks about the system prompt. The artifact leaves with her. ↩
- DTSA, EEA, CFAA, and breach of contract run on the same fact pattern. Different elements, different remedies, different statutes of limitation. None of these statutes was drafted for an artifact built slowly on a personal account, but each of them reaches it. ↩
