Where the agent lives: notes on Harvey and Docusign

The Harvey–Docusign partnership announced on 8 May is, on the surface, an unremarkable piece of enterprise plumbing. Harvey, the legal-reasoning startup that has spent two years convincing law firms it can read a contract, plugs into Docusign's Intelligent Agreement Management platform. Legal teams will reach Harvey Knowledge through Docusign Iris, Docusign's contract assistant, and get risk summaries, drafting help and approvals inside the workflow they already use. Sales, procurement, HR and finance get the same lens on whatever paper crosses their desk.¹

I want to argue that this is more interesting than it looks, and interesting in a direction the standard "AI eats legal" commentary keeps missing.

The obvious read is the wrong read. The instinct is to file this under legaltech consolidation: Harvey gets distribution into Docusign's installed base, Docusign gets a credible reasoning layer it didn't have to build, and somewhere a managing partner reads the press release and worries about billable hours. All of that is true and none of it is the point. The point is where the agent is being placed.

Those are different places, and the distinction matters more than almost anything else about how legal AI plays out from here.

The first inversion. For most of the last two years, the dominant model for legal AI has been the assistant-at-the-lawyer's-elbow. Harvey itself was built this way: a workspace a lawyer logs into, uploads documents to, and queries. Thomson Reuters' CoCounsel works the same way. So does most of the long tail. The mental model is the associate's screen, augmented. The unit of value is the lawyer-hour made more productive.

This is a perfectly defensible product. It is also, I'd argue, a strategic dead end for anyone who wants to be the layer rather than the tool. A workspace lawyers visit is a workspace lawyers can stop visiting. The switching cost is a bookmark.

Docusign, whatever you think of it as a company, sits somewhere structurally different. It sits on the contract itself, at the moment the contract becomes real. Every agreement that flows through IAM is a piece of structured data with a known counterparty, a known signing date, a known set of obligations, and increasingly a known relationship to every other agreement the company has signed. That is not a workspace. That is a substrate.

Putting Harvey's reasoning into Iris means the reasoning happens where the document lives, not where the lawyer happens to be sitting. Procurement drafts an NDA and the risk summary surfaces in the same screen where they're about to send it. Sales gets a redline before the deal desk has to be paged. HR catches a non-compete that won't survive in California without anyone billing 0.4 of an hour to notice.

The second inversion. This is also where the "AI replaces juniors" framing gets the causation backwards. The work being absorbed here is not what juniors do well. It is what juniors do badly, slowly, and at a cost the business has always resented paying: first-pass review of standard paper that nobody senior wants to look at and nobody junior is good enough to clear without supervision. The bottleneck this dissolves is not associate talent. It is the queue in front of the legal team.

The work being absorbed here is not what juniors do well. It is what juniors do badly, slowly, and at a cost the business has always resented paying.

What this does to legal headcount is a real question, but it is downstream of a more immediate one: what does it do to the legal team's relationship with the rest of the business? Today legal is a gate. The contract goes out, comes back marked up, sits in a queue, comes back again. Every business person I have ever read complaining about their legal team is complaining about latency, not quality. Iris-with-Harvey-inside is, structurally, a latency play. The gate becomes a filter, and the filter runs in the document.

The commercial model question. Here is where I get less sure and more interested. Harvey's pricing to date has been seat-based, sold to firms. Docusign's pricing is platform-based, sold to enterprises. The Harvey-inside-Iris configuration is going to expose something the legal AI market has been studiously avoiding: what is the unit of value when the agent is doing the review and a human is approving the output?

It is not a seat. The procurement manager who gets a risk summary is not a Harvey user in any meaningful sense; they are a beneficiary of one Harvey call. It is not quite a per-document fee either, because the value of the call varies wildly between a £500 NDA and a £50m supply agreement. And it is emphatically not an hour, because no hour was spent.

I think this is the first large legaltech deployment where time-based pricing simply has no surface to attach to. Neither side of the partnership bills hours. The work being displaced (associate first-pass review) was billed in hours, but the work replacing it is being delivered through a SaaS platform and a reasoning API. Somebody is going to have to invent the pricing here, and whatever they invent will become the template the rest of the market argues about for the next three years.

What I am watching. Three things, in order of how much they would change my read.

First, whether Harvey's law-firm customers treat this as competitive or complementary. If firms see Iris-with-Harvey as the tool their corporate clients use to do work the firm used to do, the partnership has a political problem on the firm side that the press release does not acknowledge. Harvey has spent two years selling to firms. Docusign sells to the clients of those firms. These two distribution motions point at the same revenue from opposite sides.

Second, whether Docusign actually ships the integration depth the announcement implies, or whether Iris-with-Harvey turns out to be a thin wrapper that throws documents over a wall to Harvey's API and surfaces the response. The difference between those two architectures is the difference between a substrate play and a referral deal.

Third, the pricing. Whatever Docusign and Harvey settle on for how this is sold will tell us more about where legal AI is actually heading than any number of conference panels on the future of the profession.

The gate becomes a filter, and the filter runs in the document.

The piece I would not write is the one about whether this is good or bad for lawyers. It is going to be both, in the usual unevenly-distributed way, and that argument has been had. The more useful question is the structural one: agents move to where the work is, not where the workers are, and a contract platform is closer to the work than a law firm has ever been.

That is the bit worth watching.

Footnotes

¶