Buried inside the Eastern District of Michigan's pro se page is a short, almost clerical paragraph about using AI tools to prepare court filings. The court is calm about it. Chatbots and writing assistants are now common. You can use them to understand legal concepts, draft language, organize your thoughts. Then it makes the part that matters explicit: the filer remains responsible for the truthfulness, accuracy, and legal sufficiency of every document submitted. The tool helps. The duty does not move.

It is easy to read that as a small administrative notice. My read is that it is something more useful. It is a published wiring diagram for how AI is going to be evaluated inside any institution that has a signature at the end of the workflow.

The Ninth Circuit gave the same signal from a higher altitude. In LNU v. Blanche, the court treats AI-assisted drafting as normal enough to address directly, while reiterating that verification and judgment cannot be delegated. The tool is permitted. The accountability is not transferable. Two different tiers of the federal judiciary, looking at the same drafting reality, arrived at the same shape: speed in, duty unchanged at the gate.

This is not a story about courts being slow, or AI being scary, or lawyers behaving badly. It is a story about a product design pattern that every builder working in a regulated or consequential domain is about to live inside.

The pattern the courts just drew

Strip the legal vocabulary out and you are left with a clean operating shape. AI compresses the drafting loop from days to minutes. The verification loop, the part where a named human takes responsibility for the artifact, runs on a different and largely unchangeable clock. You can ten-x the first part. You cannot meaningfully compress the second part without breaking the legitimacy that makes the institution useful in the first place.

The interesting move is not to complain about the tempo mismatch. It is to design for it.

The design problem is this: between the moment an agent finishes generating something and the moment a human signs their name to it, where does the state live? Who holds custody of the draft, the sources, the prompts, the revisions, the model version, the tool calls, the rejected branches? What is handed to the signatory, and in what shape, so that they can actually exercise the judgment the institution requires of them?

If the answer is "a Word document and a vibe," you have built nothing durable. You have built a faster way to produce work the institution will eventually distrust.

If the answer is a workflow that preserves state across the drafting loop, packages it for the human at the verification gate, and keeps the signature as the load-bearing moment, you have built something the institution can actually accept at scale. That is the moat.

State-preserving routing, not faster generation

I keep coming back to a small but consequential distinction. Most AI product roadmaps in regulated domains are still organized around generation speed. Faster drafts. More variants. Bigger context windows. Better tool use. All real, all useful, all easy to demo.

None of it is the hard part anymore.

The hard part is the routing layer between generation and sign-off. State-preserving routing means the workflow knows what the human is being asked to assent to and surfaces exactly what they need to do that honestly. It carries the draft and the chain of decisions that produced it. It flags the parts that came from the model with low confidence. It keeps cited authorities pinned to verifiable sources rather than letting them drift into plausible-sounding fabrications. It makes the act of signing legible, not ceremonial.

That is a product, not a policy. It is closer to how I'd think about a decision-layer integration problem than a compliance feature. The model worked. The drafting worked. The question is whether the system can hold context, provenance, and custody all the way to the moment a human takes ownership of what the agent produced.

Builders who solve that get a quiet, durable advantage. The institution accepts their output faster because the verification surface is cleaner. The signatory can actually sign with their eyes open. The speed of the drafting loop becomes usable instead of suspicious.

Why this generalizes well beyond courts

The legal example is sharp because the duty is written down. But the same pattern is showing up wherever AI accelerates creation against a human accountability gate that is not going to move.

Marketing teams subject to regulatory review for financial services, pharma, or political advertising are running into it. Contracts and disclosures where a named officer signs the document are running into it. Engineering change approvals where a licensed professional stamps the work are running into it. Research outputs where a principal investigator is responsible for the integrity of the data are running into it. Anywhere a human signature, license, or named accountability sits at the end of a workflow, AI speed runs ahead of the verification clock and lands at the same design problem.

This is the same shape I tried to name in The Evidence Contract when the EU's border system went live: probabilistic execution meeting deterministic institutional gates. The court guidance just adds the personal version. Not only must the system produce auditable evidence. A specific human must be able to read it, judge it, and put their name on it.

The surface area is enormous. The product opportunity is to build the routing primitives that make this gate usable across domains: state capture, provenance pinning, source verification, draft packaging, sign-off telemetry, rollback. None of it is glamorous. All of it is leverage.

What this makes possible

The builders who internalize the duty-persistence pattern get to do three things their competitors cannot.

They can sell into regulated workflows without spending the first six months of every deal arguing about whether AI is allowed. The court has already answered that question. AI is allowed. The duty did not move. The conversation becomes about how the workflow respects the gate, which is a conversation builders with state-preserving routing can win quickly.

They can compress the drafting loop hard, because the verification layer is cleanly designed to absorb that speed. Generation can iterate dozens of times against the same accountability surface without producing chaos. This is the opposite of the tempo mismatch I wrote about when creative loops outrun the record they leave behind.

They can charge for the verification surface itself. In regulated domains, the willingness to pay is concentrated at the gate, not at the generator. The model is increasingly commoditized. The routing layer that makes a named human comfortable putting their signature on AI-assisted work is not.

The judgment underneath

There is a quiet discipline in reading the court guidance and not arguing with it. The instinct in a lot of AI conversations is to treat the duty as friction, something to be lobbied away or routed around once the technology is good enough. That instinct misreads what the duty is for.

The signature is not a bottleneck. It is the thing that makes the output mean something inside the institution. Erode it, and you erode the value of the speed you just bought. The accountability surface is the asset. The Stoic move, if it needs naming once, is the dichotomy of control: the model's capability is one thing, the institution's duty point is another, and the builder's leverage lives in the workflow they actually own between the two.

That is also where the operator texture matters. I've watched teams ship genuinely impressive AI outputs into enterprise environments and then stall because the signatory at the gate could not, in good conscience, sign. Not because the work was wrong. Because the workflow gave them no way to verify it without redoing it. The fix was never a better model. It was a better hand-off.

A closing turn

If you are building with frontier models inside any workflow that ends with a named human taking responsibility, the useful question is not how much faster you can draft. It is where duty attaches when the draft becomes real, and whether your system makes that moment honest.

The courts have drawn the gate clearly. The signature has not moved. Build the routing layer that meets it, and AI speed becomes something an institution can sign.

Sources