10 legal risks AI startups underestimate

Early-stage AI companies tend to spend their legal attention on the fundraise and the cap table. The problems that slow down sales, or turn into claims later, are usually somewhere else: in how the product collects data, and in the contracts it's sold on.

These are the ten we see founders underestimate most often.

1. Recording consent, including the people you didn't mean to record

If your product records or transcribes conversations, consent rules apply. In California, Washington and several other states, recording a private or confidential conversation generally requires everyone's consent, whether it happens on a call or in person. Always-on microphones also pick up bystanders who never agreed to anything.

Fix: an audible or visible notice at the start, a way to opt out, recording triggered by an action rather than always on where possible, and signage for in-person capture. Recording calls with AI.

2. Voiceprints and other biometric data

If your product identifies people by voice or face, you may be collecting biometric data. Illinois' Biometric Information Privacy Act requires written notice, written consent and a published retention and destruction policy, and lets individuals sue for up to $1,000 per negligent violation and $5,000 per intentional or reckless one. Since a 2024 amendment, repeated scans of the same person by the same method count as one violation. Texas and Washington have their own biometric laws.

Fix: know whether your features create biometric identifiers, get consent before collecting them, and publish a retention schedule you actually follow.

3. Training on customer data

Enterprise customers will ask whether you use their data to train or improve your models. If your terms are silent or vague, the answer becomes a negotiation on every deal, and using recordings or customer content for your own purposes is exactly the theory plaintiffs are testing against AI vendors.

Fix: decide your position, write it into your terms and DPA, and make training on customer data off by default or limited to what customers opt into.

4. Keeping everything, forever, where anyone can see it

Raw recordings and transcripts kept indefinitely, broad internal access, and third-party model APIs that send data to other providers or regions all show up in security reviews.

Fix: a retention schedule, role-based access to raw data, and a current list of subprocessors that your contracts commit you to maintain.

5. Selling on an order form with no real contract behind it

A one-page order form gets the first customers live quickly. It usually leaves out a limitation of liability, data protection terms, IP ownership, and who's responsible for getting end-user consent.

Fix: a short master agreement and a data processing addendum that every order form points to, with the customer responsible for consent when they deploy your product to their own users.

6. AI calling and the TCPA

Under a February 2024 FCC ruling, outbound calls using AI-generated voices count as calls with an "artificial" voice under the Telephone Consumer Protection Act. That generally means prior consent (written consent for marketing calls), caller identification and an opt-out, with statutory damages per call.

Fix: consent capture and calling rules built into the product before you scale outbound. AI voice agents and the TCPA.

7. Promising SLAs your suppliers don't give you

If your product depends on a cloud provider, a model API or compute you resell, your customer SLAs can easily be stricter than what your suppliers promise you. When they go down, you owe credits they don't.

Fix: match your customer SLAs, credits and liability caps to your supplier contracts, with a buffer, and carve out upstream outages where you can.

8. Liability when your AI is part of the customer's workflow

Once your output feeds a customer's decisions, such as a missed deadline, a wrong number or a bad recommendation, someone will ask who's responsible. Without contract terms, that question gets answered after the fact.

Fix: clear statements of what the product does and doesn't do, human review where it matters, warranty disclaimers that fit your actual promises, and a liability cap with sensible exceptions.

9. IP that isn't owned by the company

Code and models written by founders before incorporation, by contractors without assignment clauses, or by advisors on a handshake may not belong to the company. Investors and acquirers will check.

Fix: written IP assignments from every founder, employee and contractor, including work done before the company existed. Who owns the code?

10. Investor consent rights over commercial deals

Preferred stock terms often give investors a veto over certain actions, such as selling or exclusively licensing key assets, taking on debt above a threshold, or changing the business. A large customer or partner deal can trigger one.

Fix: know your protective provisions before you negotiate exclusivity, IP licenses or financing terms with a customer or partner, and get any consent you need early.

Start with the contracts you sell on

Most of these are fixed by a handful of documents: your customer terms, a data processing addendum, a privacy policy that matches what the product does, and IP assignments. Getting them right before enterprise sales start costs far less than renegotiating them deal by deal.