SystemsPERSUIT2026

Designing an AI agent lawyers would actually trust

Persi advises Fortune 500 legal teams on whether to keep a matter in house or send it to outside counsel. I led design from experimentation to beta — principles, content design, research, and the system components.

My role: Design lead — principles, content design, user research, design system
Focus: Product design, Design system, Strategy, AI
Outcome: 22 beta customers; a deciding factor in contract renewals

Background

Persi is PERSUIT’s legal operations agent. It helps decision makers at Fortune 500 companies work out when to take a matter in house versus hire outside counsel, weighing subject matter, size, geography, and cost.

I joined as design lead while it was still in experimentation — model development and data architecture were both still being worked out. That timing mattered. Designing alongside an unsettled model means the interface has to hold up whether the underlying system is confident or uncertain, and you cannot wait for the technology to stabilize before deciding what the product should feel like.

The problem

Persi was in its infancy in both look and feel and behavior. It also had to sit inside a platform our users already knew, which pulled in two directions: enough cohesion that it felt like PERSUIT, enough distinction that people understood this was a genuinely new kind of experience.

I worked closely with product, engineering, and design to align on use cases and styling that could carry both.

Principles

I developed the principles with the product owner over three workshops.

Competitive analysis first. I researched adjacent software solutions alongside the leading AI chat experiences, then shared the findings with the team. The point wasn’t to copy patterns — it was to ground everyone in the emerging conventions for agent interfaces before we started arguing about ours. When a team lacks shared reference points, principle-setting turns into taste, and taste doesn’t scale.

Competitive discovery research board comparing six AI products, with screenshots grouped by pattern and annotated in pink, alongside a highlighted key insight panel — Six products, pulled apart by interaction pattern. Shared out before the workshops so the team argued from evidence rather than preference.

Then a working session with the whole build team. I set up a FigJam canvas and ran 11 to 13 participants — product managers, engineers, architects, and designers — through exercises applying those best practices to our own Persi use cases.

We covered a wide range: memory, when human or customer success intervention was needed, and how people navigate between surface areas.

The principles that came out of the three workshops.

The intervention question was the most consequential. An agent that never hands off frustrates people at exactly the moment stakes are highest; one that hands off too readily never earns trust. Deciding where that line sat shaped a lot of what came after.

Design system

I introduced a color palette and new components, which the team reviewed and which we landed after finding a workable compromise between the legacy system and the future proposals. I worked closely with our design team and design system engineering leads to evaluate each component for scalability, durability, and maintenance.

Building for an agent surfaced needs the existing system didn’t have, so each addition had to earn its place against the cost of maintaining it.

Design system proposal board showing colour swatches alongside proposed Persi components — message states, artifact panels, warnings, and result cards — laid out for team review — The proposal the team reviewed. What shipped was a compromise between these and the legacy system.

Specification for the prompt box component showing its states across empty, focused, filled, disabled, and error, with annotations for each — The prompt box, specified state by state. It's the one component every use case passes through, so it had to survive all of them.

With engineering, we also evaluated out-of-the-box data visualization libraries for displaying results — Persi’s recommendations lean on benchmark data, so how that data appears is part of whether the recommendation reads as credible.

Use cases

Each feature squad contributed the use cases relevant for Persi to take on: matter intake uploads, approvals and governance, e-billing invoice status updates, among others.

Persi user flows organised by squad, with a horizontal lane per squad and each step captured on a yellow card across panel management, data, firm pricing, e-billing, and intake — Every squad's flows on one canvas. Seeing them together is what made the shared interaction patterns visible.

As the lead on Persi I needed to be across most of them, because the interactions and common prompts only become clear once you can see the whole range at once. A pattern that works for an invoice status check falls apart on a governance approval.

Grid of condensed chat states across different squads and phases, showing how a Persi conversation appears at each stage of several different use cases — The same conversation model, stress-tested across use cases.

Rules and governance screen where a user edits business rules in plain language on the left and Persi generates the structured matter routing, RFP creation, and in-house handling rules on the right — Governance is the clearest case for an agent here: the user writes the rule in their own words, Persi produces the structured version, and the structured version is what's shown back for approval.

De-risking through content design

A key part of my work was de-risking through content design — building trust and understanding between our clients and Persi.

This is where most of the real design decisions lived. When an agent tells a general counsel that benchmark fees for a matter are a particular range, the words around that number determine whether it gets acted on, ignored, or escalated. Content design is the trust surface.

Persi confirming a set of business rule updates, with the conversation on the left and an artifact panel on the right listing the new RFP requirements, in-house handling rules, and billing and engagement terms, with a collapsed warning at the top — Changes are summarised back in an artifact panel rather than buried in the conversation, so what the agent actually committed to is reviewable in one place.

The same artifact panel with its warning expanded, listing the specific rules the user must confirm before the changes take effect, each stated in full — The warning expands to state every rule in full. Making the user open it is friction on purpose — this is the moment where an unnoticed mistake becomes a governance problem.

Outcomes

22 customers on the Persi beta program
A deciding factor in renewing their contracts with PERSUIT

Product and I ran research sessions with a subset of beta customers to understand their most common prompts. Two things came through clearly: they were glad to reach benchmarking data more easily, and they wanted a partner in reaching the final in-house versus outside-counsel decision — not a system that made it for them.

That distinction validated the intervention line we drew in the workshops. The value wasn’t automation of the judgment. It was removing the work that was getting in the way of it.