Seven clauses. That is the entire show tonight. This is AI Law, on Terms.Law Radio, one oh one point three, and before your business signs up for any AI tool, there are seven clauses that decide nearly all of the legal risk. For each one, I will give you what it is, what good looks like, and a single test question a non-lawyer can send to the vendor and understand the answer to. Have a pen ready.

Clause one: confidentiality. Every vendor agreement protects somebody's secrets. The question is whose. Plenty of agreements define the vendor's pricing, technology, and documentation as confidential in loving detail, then handle your data in a different section with softer verbs. What good looks like: your inputs and outputs are expressly defined as your confidential information, protected with the same standard of care the vendor demands for its own. The test question: is my data defined as confidential information in this agreement, in the same words you use to protect yours? If the answer requires a tour of three documents, that is an answer too.

Clause two: training rights. I did a full episode on this one, so tonight, just the checklist version. The promise not to train usually covers a defined term, and the definition has edges. Feedback, telemetry, and safety review windows often sit outside it. What good looks like: a written commitment covering content, outputs, feedback, and fine-tuning, at your tier, not just at some premium tier you did not buy. The test question: do you train on my content, my outputs, or my feedback at the tier I am buying, and where exactly does the agreement say so? Make them point to the sentence.

Clause three: security. Not the marketing page, the commitment. What good looks like: named certifications the vendor actually holds, encryption in transit and at rest, and a breach notification promise with a number in it, hours or days, not language like without undue delay doing unsupervised work. The test question: which security certifications do you hold today, and how many hours after discovering a breach involving my data will you notify me? A vendor that will not put a number on notification is telling you where you sit in their queue.

Clause four: output ownership. Most commercial AI vendors now say you own the output, and then the surrounding sentences do quiet work. Outputs may not be unique to you. The vendor may keep rights to use outputs to improve services. Ownership may be conditioned on your account being paid up and compliant. What good looks like: assignment or ownership of outputs to you, conditions you can actually live with, and no license back to the vendor broader than operating the service. The test question: do I own the output, and which sentence in this agreement takes any part of that ownership back?

That is four of seven. This is AI Law, on Terms.Law Radio.

Clause five: indemnity. If AI output gets you sued, the question is whether the vendor stands in front of you or behind you. Several major vendors now offer copyright indemnities for business customers, and they are real, but they are conditional, and the conditions are the whole game. Commonly you must be on the right tier, using the safety features, unaware of the problem, and inside the usage rules. What good looks like: an indemnity that covers third party infringement claims arising from output, with conditions your team can actually operate within, written down and monitored. The test question: if your output triggers a claim against my company, do you defend and pay, and exactly which behaviors on my side void that protection? Ask for the disqualifier list. It exists.

Clause six: the liability cap. Scroll to the section in capital letters. Somewhere in there, the vendor's total liability is capped, commonly at the fees you paid over the previous twelve months, sometimes less, and consequential damages, meaning your lost profits and your customer's losses, are excluded entirely. Read that cap next to your actual exposure. If a bad output could cost your business six figures and the cap is a few thousand in subscription fees, you now understand the deal you are signing: the risk is priced onto you. What good looks like: a cap you have at least consciously accepted, carve-outs for confidentiality breaches and indemnity obligations, and no cap so low it makes every promise in the agreement ornamental. The test question: if this tool causes my business a serious loss, what is the most your company will ever actually pay me under this agreement? The answer is a number, and the number is usually small.

Clause seven: exit and portability. Every subscription ends someday, and the exit clause decides whether it ends gracefully. What good looks like: you can export your data and your outputs in a usable format, you have a stated window after termination to retrieve them, deletion afterward is confirmed on a stated timeline, and nothing important is held hostage to a final invoice dispute. Fine-tuned models and configurations deserve a special look: if you spent a year customizing on their platform, ask what walks out the door with you. The test question: when I cancel, what do I get back, in what format, within how many days, and when is the rest deleted?

So what do you do with the seven? Something very simple. Put the seven test questions in one email to the vendor before you buy. Not a negotiation, just questions. A serious vendor answers most of them in a day, because serious vendors have answered them a hundred times. Evasion, boilerplate, or a link to the same document you were asking about, each of those is information, and it is information you collected for the price of one email. Then weigh the answers against what you are actually feeding the tool. A brainstorming toy needs two of these clauses. A tool touching client data needs all seven.

Before your team adopts the tool, upload the agreement to the free Terms.Law analyst and check the data, ownership, liability, and exit provisions. It is at terms dot law.

That is AI Law for tonight. The fine print: this is general commentary and education, not legal advice about your company, and listening does not create an attorney client relationship. AI law and vendor terms both move fast, so verify the current documents before you rely on them. The analysis belongs to Sergei Tokmakov, California attorney. Seven clauses, one email. Good night.