Harvey Builds AI Product Films for Individual Sales Prospects

Harvey’s October 21 account of its marketing team shows how a B2B marketer can turn prospect-specific product videos into repeatable work. The team built a library of product screens and animations that AI assembles into films, while people choose the story and review it.

For a launch film, the work began with a storyboard in Figma, the design tool. Human feedback shaped individual scenes and the ending. In their joint account, creative director Shawn Farsai and president and co-founder Gabe Pereyra write:

“We review on the canvas, sometimes 50 comments at a time.”

Those comments become production decisions. The software revises the board and asks which notes should become rules for future work. A good correction can therefore influence the next film as well as the one being reviewed.

Give each prospect the relevant product story

The account-specific films address a sales problem: showing a prospective customer the part of the product that matters to them. The team first got usable demonstrations through 15 to 20 minutes of back-and-forth per film. Repeating that exchange for every target account would still require substantial attention.

The authors put the initial success in proportion:

“Anyone can get one flashy demo out of an agent.”

Here, an agent is software that acts inside design tools from instructions. The repeatable process needs decisions and material it can use again without reconstructing the product each time.

Harvey built two shared libraries. Product Map holds the product screens. Motion Primitives holds reusable animations, including how changes appear in a document. When a component is corrected, films using that component inherit the fix, according to the authors.

The team then defined 44 demonstration flows, each focused on a product moment a buyer asks to see. Information from sales calls and account records helps shape a script, with a fictional company shown on screen. That gives the buyer a relevant example without requiring a completely new visual language for every account.

For a small B2B marketing team, the transferable choice is what stays consistent across sales conversations. A repeated product action can become an approved scene. The reason a particular buyer needs it still requires a decision about their business.

That follows the question in our coverage of shared AI responsibilities in marketing jobs. Someone has to maintain the material colleagues rely on. Here, that responsibility includes deciding whether a product screen still represents what the company sells.

Creative direction survives the batch

The production account describes one batch of 36 films taking 69 minutes. That happened after the team built the shared libraries and production system. It isn’t the time required to start from nothing, and it doesn’t establish a total cost saving.

Getting there involved practical failures. Agents trying to work in the same file queued and timed out. The team separated production into three files, each with its own building agent, with another coordinating exports and delivery.

A common tracker records each film’s account, story, status, and links. Automated checks flag obvious visual problems, such as clipped text or small type. Neither the tracker nor those checks proves every film passed human review or helped win a deal.

The creative team’s earlier November 2025 design essay placed content verification with people while automation handled repetitive work. That was a statement about product design, not a measurement of this marketing system. The current account shows human direction in a different setting: deciding what a film communicates before production and correcting it afterward.

Our reporting on product teams changing handoffs raised a related question about where decisions happen. In this case, reviewing the storyboard gives the person responsible for the message a place to change it before watching finished exports.

Judge the work buyers receive

There are clear limits to the evidence. Harvey sells legal AI and benefits from showing its own team doing sophisticated work with it. The case study doesn’t publish a comparison of full production costs or a measured effect on sales. The dates on which the individual films were produced aren’t specified.

A founder-marketer shouldn’t treat the reported batch time as a staffing plan. Building and maintaining the library is work. Reviewing a relevant product claim is work, too. The operating example is useful because it shows where that effort went.

For a trial, choose one product moment sales repeatedly explains. Approve its screen sequence, then adapt it for a different buyer need. Have the person who owns the product claim review that version before it reaches a prospect. Record which changes required a new creative decision and which reused an approved component.

That gives AI content production a test tied to the buyer’s experience. Expand the library when the adapted film shows the right product behavior and a relevant use for it. A batch that finishes quickly still needs a reason for each prospect to watch.