The agreement layer is where legal AI actually lives

The Docusign–Harvey partnership announced this morning is, on the press release, exactly the sort of thing you would expect: a legal AI vendor wires into a signature workflow, both parties get a logo on each other's slides, customers get a demo. I would normally let it pass without comment. I'm writing about it because I think the framing the market is reaching for, "Harvey is now an agreement-management feature", gets the direction of the arrow backwards, and the inversion matters for how anyone serious about legal AI should be thinking about distribution.

Start with what each side actually has. Harvey has the model layer and a defensible book of large-firm and in-house customers paying serious money for drafting, review, and research. Docusign has something Harvey does not, and cannot build: the agreement itself, in its executed, dated, counter-signed, stored, cross-referenced state. Every NDA, every MSA, every order form, every employment contract, every procurement agreement that an enterprise has signed in the last decade is sitting in a Docusign vault or an equivalent, with metadata, with signing parties, with version history. That corpus is the substrate. Drafting and reviewing documents in the abstract is a parlour trick next to drafting and reviewing them against the actual book of agreements the company is already bound by.

This is not a controversial claim if you look at where the value accrues in any other vertical AI deployment. Bloomberg's terminal is not valuable because of its analytics; it is valuable because of the data feed those analytics run on. Epic's AI features are not interesting on their own; they are interesting because they sit on the patient record. Legal has been late to this realisation because the legal-tech vendor stack was historically so fragmented that no one held the corpus. Docusign quietly built that position while the conversation was elsewhere.

The second inversion is on the workflow side. A great deal of legal AI discussion treats "the contract" as the unit of work, draft it, review it, redline it, sign it, done. That is what a junior associate sees. It is not what a general counsel sees. The GC sees a contract as one event in a multi-year relationship that touches sales (who negotiated it), procurement (who sourced the counterparty), finance (who books the revenue or cost), HR (if it is an employment matter), and compliance (always). The drafting moment is maybe ten per cent of the lifecycle. The other ninety per cent is the agreement living in the business, being referenced, being amended, being breached, being renewed, being audited.

Harvey-as-a-product is built for the ten per cent. Harvey-wired-into-Docusign is built for the ninety. That is a materially different proposition, and it is the one in-house teams will pay for, because the in-house lawyer's actual problem is not "draft me an NDA", it is "tell me which of our 4,000 active vendor contracts have an auto-renewal clause triggering in the next ninety days and which of those have an MFN provision we are now in breach of because of the deal we signed last Tuesday." You cannot answer that question without the corpus. The model is necessary but very far from sufficient.

The drafting moment is maybe ten per cent of the lifecycle. The other ninety per cent is the agreement living in the business.

Which is why I think the natural reading, "Docusign is distribution for Harvey", has it the wrong way round. Distribution implies the model is the product and the channel is the means. In this configuration the corpus is the product and the model is the means. Harvey is doing the reasoning, but the thing the customer is buying is enriched access to their own agreement estate. The model is interchangeable in a way the corpus is not. If Anthropic ships a better legal reasoner next quarter, Docusign swaps it in. If Harvey wanted to leave, it would walk away from the very thing that makes its reasoning useful in this context.

I want to be careful here, because there is a respectable counter-case and I don't want to flatten it. The counter is that Harvey's specific post-training on legal corpora, its workflow surface, and its existing relationships at AmLaw 50 firms constitute a real moat that a generic frontier model wired to Docusign couldn't replicate. There is something to this. Firm-side legal work, the bet-the-company litigation, the novel M&A structure, the regulatory grey zone, does benefit from domain-specific tuning and from a product surface built by people who understand how lawyers actually work. Harvey's defensibility is real on that axis. But it is defensibility against other legal AI vendors, not against the platform underneath. The firm-side market and the in-house market are different beasts, and the in-house market is where the volume sits.

The third thing worth saying is about professional services more broadly. I have written before about time-based pricing being structurally doomed once agents do the leverage, and this announcement is a small but clear data point in that direction. The work that used to fill an associate's afternoon, pulling up the precedent agreement, checking the standard clauses against the client's preferred positions, flagging the deviations, is now a query against the corpus that returns in seconds. That work was billable. It is now infrastructure. The firms that adapt are the ones that figure out what to charge for once the leverage is gone, and "access to the corpus, intelligently queried" is one plausible answer, but it is not an answer that bills by the hour.

The cross-departmental piece named in the announcement, sales, HR, procurement, is the part that most reads as filler but is actually the most consequential. Agreements are the connective tissue between functions. A sales contract is also a finance event, a delivery commitment, and a compliance posture. Once an AI layer can reason across all of that simultaneously, the boundary between "the legal function" and "the operating system of the company" starts to blur. That is a much bigger claim than today's partnership announcement supports, but it is the direction the line points, and it is why I think pieces like this one are worth writing about something that on the surface is just another integration.

What I am watching for next: pricing structure (does Docusign price the Harvey-enabled tier per seat, per agreement, or per query, because each implies a different theory of value), the first big in-house customer announcement (firm-side is the easier sell; in-house is the volume), and whether any of the other agreement-management incumbents respond with their own model partnership. The shape of those moves will tell us whether the corpus-as-moat reading is right, or whether I have over-rotated on a press release.

I think it is right. But the next ninety days will say.

¶