Trajectory.ai cut model onboarding from three months to one week, and the gap tells you something about where legal AI is heading

Ronak Malde, of Trajectory.ai, offers a precise and uncomfortable benchmark for anyone still treating legal AI deployments as inherently slow, expensive, and bespoke. The first time Trajectory.ai worked with Harvey, it took three months to stand up the full model and get it operational. That figure reflects not incompetence but the actual state of the infrastructure at the time: nothing was settled, every piece had to be reasoned through, and the people doing the work were, as Malde puts it, building the airplane as it was flying.

That first engagement is now the baseline, and the baseline has been overtaken quickly. Malde says Trajectory.ai can now train a model with Harvey in under a month. For a subsequent customer, the onboarding came in at roughly one week. Three months to one month to one week, across a handful of engagements. The direction is not ambiguous.

What Malde describes is the difference between pioneering and operating. The first Harvey engagement required Trajectory.ai to make foundational decisions without precedent: how to structure training data, how to set up the pipeline, how to validate that the resulting model was actually doing what a legal professional would need it to do. Every one of those decisions, once made and tested, became a reusable part of the playbook. The second engagement was faster because the first one existed. The third was faster still.

The first engagement, it took like three months to like set up the entire thing. We're kind of building the airplane as it was flying and now we're able to train a model with Harvey in under a month. and then onboarded a new customer and within a week we were able to train a model and really get the flywheel going. Ronak Malde

This kind of compounding is easy to describe and easy to underestimate. A reduction from three months to one week is not a 75 percent improvement in efficiency. It is a qualitative shift in what the product can offer to a customer. At three months, model onboarding is a major institutional commitment, the kind that requires budget sign-off at multiple levels, tolerance for a long period of uncertainty, and a champion inside the client organization willing to absorb the friction. At one week, it starts to resemble a product feature rather than a professional services engagement. The sales conversation, the risk calculus, the implementation cost: all of them change when the timeline compresses by that magnitude.

The Harvey context matters here. Harvey operates in legal work, a domain where the stakes attached to model outputs are high and where clients are, by professional training, skeptical of claims that have not been tested and retested. Cutting onboarding time in that environment is harder than cutting it in a context where a model that produces something imperfect is simply corrected on the next iteration. Legal AI customers are not running experiments. They are evaluating whether a tool is fit for production use, and they are doing so carefully. Getting from three months to one week while maintaining that standard of scrutiny suggests the underlying tooling has matured, not just that the team has gotten faster at the same work.

Malde frames the outcome in terms of the flywheel, and that framing is worth taking seriously. A flywheel, in this context, means that each successfully onboarded customer makes the next onboarding cheaper and faster, partly through refined tooling and partly through a growing body of evidence about what works. The compression Trajectory.ai has achieved is not just a historical fact about how long past engagements took. It is a structural advantage that accumulates. The organization that has already done this work ten times is operating on a different cost curve than one doing it for the first time, and the gap widens with each additional engagement.

The honest caveat is that Malde’s account is a single data point from inside Trajectory.ai, reported without independent verification of the specific timelines or the conditions under which they were achieved. Different customers, different model complexity requirements, and different data environments could all affect what “one week” actually entails in a given case. What the claim establishes, taken on its own terms, is a direction and a rough magnitude. That direction is toward commoditization of what was, not long ago, a genuinely novel and slow capability. Whether the timeline compression holds across all use cases or only the most straightforward ones, the pattern Malde describes is one that the broader legal AI market will not be able to ignore.