This is the Counsel's Desk, on Terms.Law Radio. Ninety two point five. This is the Washington to Contract report. Tonight's question sounds abstract and is anything but: when Washington and the states both claim the right to regulate artificial intelligence, whose rule ends up in your contract? One rulebook or fifty. The answer decides your compliance budget, your vendor negotiations, and a clause most businesses have never heard of that I will spend real time on tonight, the compliance allocation clause.

Start with where things stand, per public reporting. On June second, the President signed an executive order on artificial intelligence built around a voluntary framework: developers may share pre release access with the government for up to thirty days, agencies get new coordination jobs, and the framework itself is due to be finalized by August first. Alongside it came a companion directive that got less attention and deserves more: it instructs the federal government to challenge state AI laws. California, Colorado, and Texas have each enacted their own AI statutes, and other states are drafting. The administration's position is that a patchwork of state rules should yield to a single federal approach. The states' position, predictably, is no. That disagreement is called preemption, and it will be worked out slowly, in courts and in Congress, on a timeline nobody controls.

I want to be careful about what I am not saying. I am not predicting who wins. Preemption fights are genuinely uncertain; they turn on the wording of statutes that do not exist yet and on doctrines that judges apply unevenly. What I can tell you is what the uncertainty itself does to a business, because uncertainty is not neutral. It lands somewhere. It lands in contracts.

Consider a company that sells an AI screening tool to employers in a dozen states. Colorado has obligations for high risk AI systems. California has its own set. Texas has another. Under a single federal rule, that company builds one compliance program. Under fifty rules, it builds a matrix. And under today's condition, which is neither, it has to decide who bears the cost of not knowing. That decision gets made in the contract whether or not anyone writes it down. Silence is also an allocation. It usually allocates the risk to whoever can least afford to litigate it later.

So here is the clause I promised, compliance allocation, in three parts.

Part one, who complies with what. The vendor warrants that the system, as delivered, complies with the AI specific laws of a defined list of states as of a defined date. Not compliance with all applicable law, a phrase that means nothing until a judge tells you what it meant, but named statutes, named states, named date. The customer, in turn, takes responsibility for how it uses the tool: its own deployment decisions, its own notices to employees or consumers. Delivery compliance versus deployment compliance. Most of the AI statutes themselves distinguish developers from deployers, so the contract should too.

Part two, change in law. The unstable part. If a new state statute takes effect, or a federal law preempts the state ones, or a court enjoins a statute the contract assumed, what happens? The mature answer is a change in law mechanism: the vendor updates the system within a stated period, the parties share defined categories of cost, and if compliance becomes commercially impracticable, either side can exit on notice without penalty. If the preemption fight resolves in either direction, this is the clause that decides whether your agreement adapts or breaks.

Part three, the representations you give downstream. If your customers ask you to promise that your AI features comply with every state's law, present and future, decline politely and offer the structure above instead. An open ended future law warranty is not diligence. It is a lottery ticket your customer is holding against you.

You're listening to the Washington to Contract report, on the Counsel's Desk, Terms.Law Radio.

Now the planning question, and the phrase I keep using: the strictest state floor. While one rule or fifty stays unresolved, a business has three options. Comply with nothing and hope. Comply state by state with a matrix that changes quarterly. Or find the strictest requirements among the states where it actually operates and build to that level everywhere. I keep coming back to the third, for unglamorous reasons.

First, it converts a legal question into an engineering one. Instead of asking which statute applies to this particular user, a question that depends on domicile and place of harm and choice of law, you ask what the most demanding statute requires and do that for everyone. Second, it front loads work you will likely need anyway. The recurring themes across the state statutes, notice that AI is being used, some form of impact or risk assessment, a human appeal path for consequential decisions, documentation of training and testing, appear in enough places that building them once is rarely wasted, whichever rulebook survives. Third, it reads well in hindsight. If a regulator or a plaintiff ever asks what your compliance posture was during the uncertain period, the strictest standard in any state where I operated is an answer. Waiting for the courts is also an answer, just not one you want read back to you later.

The honest costs, because there are costs. A strictest state floor means you may be running assessments a lighter touch state never required, and it can slow a product launch. For an early stage company the burden is real. So here is the scaled version: apply the floor to the highest risk uses, decisions touching employment, credit, housing, health, insurance, and take the lighter path on low stakes features. Risk tiering is the approach most of the statutes themselves take, so you are borrowing their logic rather than inventing your own.

What would change my analysis. If the federal preemption effort produces an actual statute passed by Congress with an express preemption clause, the fifty rulebook problem shrinks and the floor can relax toward the federal standard. If instead the effort proceeds mainly through litigation, expect years of partial answers, one statute enjoined here, another upheld there, and the floor keeps earning its keep. Watch the voluntary framework due August first, per public reporting, as an early signal of what the federal side thinks the substantive baseline should look like. Voluntary today has a way of becoming the reference point for reasonable tomorrow.

Three steps before the quarter ends. List the states where you have users, employees, or deployments, and pull the AI statutes for those states, the actual texts, not summaries. Identify your highest risk AI uses under those texts, anything touching employment, credit, housing, or health. And put a compliance allocation clause with a change in law mechanism into your next AI vendor renewal, because the cheapest time to allocate this risk is before anyone knows how it comes out.

The practical question is whether your contract gives you an exit if the policy environment changes. The Terms.Law analyst and the related contract checklists are at terms dot law.

The fine print. This broadcast is commentary and general information, based on public reporting and government documents as of July tenth. It is not legal advice and not investment advice, and listening does not create an attorney client relationship. Policy moves fast, and preemption fights move in lurches, so verify the current state of any statute or order before you act. I'm the AI voice of Terms.Law Radio. The analysis belongs to Sergei Tokmakov, California attorney. Stay tuned, stay skeptical, and build to the floor that lets you sleep. Good night.