Latch (fka SuperLabs) is a pre-seed startup building a B2B product to capture how work actually happens, structure it into machine-readable workflows, and enable AI systems to automate them safely.
Company
Latch AI
Product
Latch
My role
Product Designer
Team
4 (2 designers, 1 PM, 1 researcher)
Status
Shipped • Spring 2026
Challenge
Teaching a non-technical user to articulate their own processes clearly for a machine to understand them.
The scope:
(1) design the UI for the first-time recording experience
(2) develop principles to shape voice-based interactions between user and agent.
What we did
6 user interviews across construction, CPG, and consulting. Competitive analysis of 10 tools across 4 categories. Two prototypes at opposite ends of a spectrum — one ambient and minimal, one Zoom-like and explicit — tested with real users and merged into a single final direction.
What we found
Users don't choose between control and simplicity. They want simplicity during recording and control at review. The biggest blocker is the inability to capture enough context to build it reliably.
The way in: show the AI how you do it. It builds the automation.
We were asked to design the dashboard where users monitor their automations.
But you can't monitor automations that were never properly captured. The real blocker was upstream.
Secondary data analysis + competitive analysis
What we did
We used ATUS/BLS data to identify which industries carry the highest manual workflow burden, then evaluated 10 competitors to understand where existing tools fall short.
Key moment
Vercept was acquired by Anthropic the same week we downloaded it. That made the pace of this industry very concrete.
Decision it drove
We scoped to SMB employees in high-manual-workflow industries and moved away from designing for executives. The data told us where the real gap was.
6 participants across diverse industries
What we did
We recruited across industries intentionally, our target user wasn't one type of worker. We looked for patterns across very different contexts.
Key moment
Every user had a completely different mental model of what automation should do for them. That range was the core design challenge.
Decision it drove
We shifted our research questions entirely to the employee side, focusing on trust and control rather than metrics and executive dashboards.
Usability testing + prototyping
What we did
We built two prototypes at opposite ends of a spectrum — one minimal, one familiar — and tested both to see where users actually land.
Key moment
Nobody landed cleanly on either end. Users wanted the control of one and the simplicity of the other.
Decision it drove
We combined both directions into one final prototype with full visual polish and design system consistency, grounded in what users showed us.
In Depth
Favored by participants who wanted automation to feel like a quiet copilot—until something broke and they hunted for where Latch was “looking.”
Tension · Legibility vs. calm
Preferred when participants wanted theatrical clarity—clear modes, obvious boundaries—at the cost of feeling “always on stage” during sensitive screens.
Tension · Performance anxiety
Users didn't fall cleanly on either end.
So, we merged them into one direction, keeping the restraint of A and the legibility of B.
Before capture starts, users see exactly what Latch will observe — screen, voice, and clicks each get a clear consent moment so recording feels teachable, not surveilled.
DECISION: Fear of invisible observation was the biggest adoption blocker. Showing users exactly what Latch accesses before anything starts removes that barrier upfront.
Trust and transparency around data access were identified as critical barriers to adoption.
Users are informed of all access permissions before recording begins.
Users retain full control to pause or mute at any point to protect sensitive information during process automation.
The controls live in a compact overlay so Latch never competes with the work happening beneath it.
An observation log lets users peek at exactly what Latch is tracking at any moment. In testing, seeing the log significantly increased confidence.
The overlay is fully draggable so it never blocks the content that matters. Users can position it wherever it's least intrusive.
Users needed to verify what the model captured before committing to automation.
Users wanted to revisit recordings alongside the model's interpretation of their session.
The editable review layer lets users re-record, add context, or upload supporting files.
DECISION: A floating overlay keeps the user's work primary. Latch stays peripheral, present enough to follow along, invisible enough not to interrupt.
DECISION: Keeping users focused on the work, not on managing the AI.
DECISION: Correcting a mistake shouldn't mean starting over. Prior context is preserved so users can fix without losing what was already captured.
DECISION: The review screen is where trust is confirmed. Typing to answer deferred questions — rather than speaking — came directly from testing feedback.
Follows along as you record, tracking what's been covered and reflecting it back to confirm understanding.
Grounds itself in what's on screen, noticing when you switch views or take action to stay in context.
Doesn't interrupt mid-thought. It reads natural pauses in speech to confirm, recap, or ask one focused question.
Every question is tied to your workflow, probing for branches and exceptions, never filler.
Lives in the corner of your screen. Mute it, dismiss a question, pause for privacy, or turn it off entirely. You're always in control.
Latch captures screen activity, audio, and behavioral patterns. The risk is that users in high-pressure workplaces may feel coerced into being observed.
Mitigation: explicit consent before every session, privacy pause at any moment, and no data retained beyond the active session unless the user saves it.
When one person's way of doing a task becomes the captured workflow, it quietly becomes the default for everyone who runs that automation after them. If that person's approach reflects individual bias or inefficiency, it scales.
Mitigation: the review step lets users challenge and edit what was captured before it's built.
When an automation fails, it's not always clear whether the user, the AI, or the system is responsible. Non-technical users are especially vulnerable to accepting outputs they can't evaluate.
Mitigation: every automated step is traceable, irreversible actions are flagged as manual steps, and humans remain accountable for the final output.