AI PM Employment Contract Guide: IP Clauses, Non-Competes, and What to Watch Before You Sign
TL;DR
AI job offers include contract clauses that were not standard in traditional PM roles. Broad IP assignment clauses can claim ownership of side projects you built before joining. Non-compete scope in AI can effectively block you from the entire industry for 12 to 24 months. Confidentiality clauses in AI companies sometimes extend to model behavior you observed while building products. And training data provisions can create personal liability if a product you shipped used data improperly. This guide covers the five clause categories that AI PMs should scrutinize, what is negotiable, and the five questions to ask a lawyer before signing.
The AI PM Minute
One tactic to make you a sharper AI PM, twice a week. 60 seconds to read. Free.
No fluff. Unsubscribe anytime.
IP Assignment Clauses: Who Owns Your AI Work
Intellectual property assignment clauses assign ownership of your work product to your employer. In most PM roles, this is unremarkable: the product you build belongs to the company. In AI roles, the scope of these clauses has expanded significantly and can reach well beyond your work hours.
Standard IP assignment
LowWork product created during employment using company resources belongs to the employer. This is industry standard and expected. If you build an AI feature at work using company time and compute, the company owns it.
Broad invention assignment
HighWork product that is related to the company's business or reasonably anticipated business, even if created on personal time. If you are building a personal AI writing tool at night and your employer is an AI writing company, this clause may claim ownership of your side project.
Pre-employment invention schedules
High if uncheckedMost good contracts include an exhibit where you disclose prior inventions you want to exclude from assignment. If you have AI projects, open-source contributions, or side products that predate the job, list them explicitly in the exhibit. Omitting them can retroactively assign them to the new employer.
Model outputs and prompts as inventions
MediumEmerging clause in frontier model companies: prompts, prompt engineering techniques, or workflow automation you develop while using company AI tools may be classified as inventions. Read carefully if the clause says 'methods, processes, or techniques' without further limitation.
Training data contributions
Low to MediumIf you contribute data, annotations, or fine-tuning examples to a company model, some contracts treat this as a contribution to a collective work with joint ownership implications. Usually company-favorable but worth understanding.
What to negotiate: Ask for an explicit carve-out for personal projects that do not use company resources and do not relate to the company's core business. In California, Labor Code 2870 already limits employer IP claims to work done on company time using company resources, but other states have no such protection. Fill out the prior inventions exhibit completely.
Non-Compete Clauses in AI: Why Scope Matters More Than Duration
Non-compete agreements restrict where you can work after leaving a company. In AI, scope is the more dangerous dimension than duration. A 12-month non-compete that covers "artificial intelligence product management" as a category can effectively shut you out of the entire industry.
Garden variety non-compete
Cannot work for a direct named competitor for 12 months. Enforceable in most states. Manageable if the named competitors are few and specific.
Watch for: List of named competitors that is so broad it covers every meaningful AI employer in your market.
Category non-compete
Cannot work in a broadly defined category such as AI-powered productivity software or language model applications for 12 to 24 months. These are aggressive because AI touches almost every software category.
Watch for: Undefined or overbroad category descriptions. Push for specific product categories, not technology categories.
Non-solicitation of customers
Cannot solicit former customers for 12 to 24 months. Standard and usually fine for PMs who are not in sales roles.
Watch for: Provisions that extend to customers you worked with tangentially or that cover a customer's entire corporate family, not just the contacts you had.
Non-solicitation of employees
Cannot recruit former colleagues for 12 to 24 months. Becomes problematic if you later want to hire a former direct report at a new company.
Watch for: Provisions that cover all employees of the company, not just those you directly managed or recruited.
State enforcement varies dramatically
California, North Dakota, Oklahoma, and Minnesota effectively do not enforce non-competes. If you are in California, a non-compete is largely unenforceable after employment ends regardless of what the contract says. If you are in Massachusetts, Florida, or New York, enforcement is real. The FTC issued a rule banning most non-competes in 2024, but its enforcement has been legally contested. Check your state's current status before assuming a non-compete is or is not enforceable.
Confidentiality Clauses and AI-Specific Gotchas
Confidentiality clauses in AI companies have expanded to cover categories that did not exist five years ago. The broad scope is driven by competitive sensitivity around model architecture, training data, evaluation results, and product roadmaps.
Model capability information
What the model can and cannot do, its failure modes, benchmark performance, and evaluation scores may be classified as confidential. This can affect what you say publicly after leaving, including in interviews, blog posts, or talks about your work.
Training data composition
The datasets used to train the company's models, including data sources, annotation methods, and data quality processes. Sharing this after departure can create trade secret liability even if the data itself is public.
Prompt and system design
The system prompts, few-shot examples, and prompt architectures used in production. Some companies classify these as trade secrets. If you reference them in a case study after leaving, you may be in breach.
Safety and alignment findings
Internal red team results, alignment testing outcomes, and safety evaluations are usually highly confidential. Companies are sensitive about public disclosure of failure modes before fixes are shipped.
Customer data and AI outputs
What AI outputs you observed during development may be restricted if they implicate customer data. This affects what you can discuss in your portfolio or in technical blog posts about your work.
Duration of confidentiality
Many AI contracts have indefinite confidentiality obligations, not the standard 2 to 5 year limit seen in other industries. Push for a defined sunset on confidential information that has since become public knowledge.
PM-specific risk: As a PM, you document product decisions in PRDs, strategy memos, and retrospectives. These documents often contain exactly the information classified as confidential: model performance data, competitive positioning, roadmap rationale. Before using any work product samples in your portfolio after leaving, check whether they fall under the confidentiality scope of your former contract.
Negotiate From Strength in Your AI PM Career
The AI PM Masterclass covers compensation, equity, career strategy, and how to position yourself for the highest-leverage AI PM roles. Taught live by a Salesforce Sr. Director PM.
Five Red Flags in AI Employment Contracts
Most employment contracts are standard. But these five provisions appear more frequently in AI company contracts and warrant immediate pushback or legal review.
Moonlighting ban that covers all AI-adjacent work
HighSome AI companies ban any outside work that relates to artificial intelligence, machine learning, or data science during employment. For an AI PM, this can mean you cannot contribute to open-source projects, write educational content, or advise a friend's startup in any AI domain. Push for a carve-out for non-competing, non-conflicting outside work done on personal time.
Clawback provisions tied to non-compete compliance
HighIf the company discovers you violated the non-compete, they can claw back signing bonuses, accelerated vesting, or equity received at separation. These provisions create significant financial exposure and may incentivize aggressive enforcement even in states where the underlying non-compete is unenforceable.
AI development activity monitoring
MediumProvisions allowing the company to audit your personal devices or online accounts for evidence of unauthorized AI development. Rare but appearing in contracts at companies handling highly sensitive model weights. Understand the scope before accepting.
Broad arbitration clauses covering IP disputes
MediumForcing IP ownership disputes into private arbitration (rather than court) can disadvantage the employee, who typically has fewer resources than the company. Some states limit mandatory arbitration for employment disputes. Know your rights.
Assignment of future inventions without sunset
HighA clause that assigns inventions created up to 12 months after employment ends to the former employer. Justified as protecting against employees who pre-develop exit inventions on company time. But 12 months is long in a fast-moving field. Push for 6 months or a narrow scope tied to work performed during employment.
Your Pre-Signing Checklist
You have 3 to 10 days from offer to signature in most cases. Use this checklist before you sign any AI PM offer, regardless of company size or prestige.
Engage an employment attorney for a contract review ($300 to $600 flat fee)
Not optional for any offer above $200K total comp. An attorney sees language you will miss and knows which provisions are standard vs aggressive in your state. Worth every dollar.
Complete the prior inventions exhibit in full
List every personal project, open-source contribution, or side business you want to protect. Vague descriptions are better than nothing, but specific is best. This is the exhibit that protects your side projects.
Request specific rather than categorical non-compete scope
Push back on 'AI-powered products' as a category and ask for a named-competitor list or a specific product category definition. A shorter, narrower scope is more negotiable than you think, especially at early-stage companies.
Ask whether the confidentiality provision has a sunset
Request that confidentiality obligations for general know-how expire after 2 to 3 years, or that they carve out information that has become publicly available through independent disclosure.
Understand the moonlighting policy explicitly
Ask HR directly: can I write a technical blog post about AI while employed? Can I advise an early-stage company that is not a competitor? Can I teach a workshop? Get the answer in writing.
Check the arbitration clause for class action waivers
Class action waivers in arbitration clauses limit your ability to join other employees in collective actions. Some states restrict these. Know what you are agreeing to.
Negotiation is expected and professional
Asking for contract modifications does not rescind offers. Employers expect senior candidates to negotiate legal terms, just as they expect salary negotiation. Frame requests professionally: "I want to make sure I understand the non-compete scope before I sign. Can we narrow it to direct named competitors?" A company that withdraws an offer because you asked a reasonable contract question is a company that would have been difficult to work with.
Accelerate Your AI PM Career
The AI PM Masterclass covers career strategy, compensation negotiation, and how to position yourself for senior AI PM roles at the companies shaping the industry.
Related Articles
Before you go: get the AI PM Minute
One tactic to make you a sharper AI PM, twice a week. 60 seconds to read. Free.
No fluff. Unsubscribe anytime.