Latch AI Case Study

Building Trust, Privacy, and Control Into an AI Agent That Learns How You Work

Organization
Latch AI
Timeline
January–June 2026
Role
Product Researcher
Team
Product Researcher, Project Manager, and 2 Product Designers
Skills
Figma Claude Cursor

Get the highlights.

Company overview Formerly SuperLabs

Latch is a seed-stage startup that raised $8 million to bring AI automation to companies that operate outside a traditional engineering culture.

Objective

Latch’s early onboarding depended on interviewing employees about how they work. Those conversations took more hours than the company could give, and they still failed to show how each person actually did the job.

Our brief Find a faster way for one employee to describe their work in enough detail for Latch to automate it.

Research question

How might Latch accurately gather enough information from a non-technical employee to build out an automated workflow?

What was done

Open a phase to see the methods used.

Discovery
  • Secondary research
  • ATUS / BLS data analysis
  • Competitive analysis
Research & Insights
  • Semi-structured interviews
  • Affinity mapping
  • Thematic analysis
Problem Definition
  • Persona development
  • User flow mapping
  • Problem framing
Concept Development
  • Divergent prototyping (low- and high-fidelity)
  • Rapid prototyping in Cursor
Validation
  • Moderated usability testing
  • Concept testing
  • Iterative synthesis
Delivery
  • Convergent prototyping
  • Design system development
  • Technical specification
Users

Non-technical employees at small and mid-sized businesses whose daily work involves repetitive, time-consuming tasks.

Shasha Workflow Example

Shasha is an office manager at a 30-person law firm, handling all of the firm's invoicing, billing, and vendor payments by hand, with no technical support to fall back on.

Pull invoice from inbox
Rename by client and date
File into correct folder
Update master billing spreadsheet
Kick off payment
Repeat, every day, for every invoice that comes in
Takes 2 hours
Phase 1 Market Research Data Visualization

Goal Determine how much of the non-technical business employees' work is still manual.

We used the American Time Use Survey / Bureau of Labor Statistics (ATUS / BLS) dataset to determine how much manual, repetitive work exists in an average workweek across occupations and industries. Findings showed that non-technical business employees lose a measurable share of their time to it, a cost that lands on both their workload and the company's operating costs.

Insight Manual administrative work costs the 155,282 surveyed workers about 104 million hours and $2.1 billion in wages every year, time and money spent on tasks that could be automated.

155,282
workers surveyed, 2003 to 2024
0
hours a year, about 671 per worker
$0
in wages a year, about $13,300 per worker

American Time Use Survey (Bureau of Labor Statistics), 155,282 employed workers, 2003 to 2024, by their actual occupations in the survey. Share modeled from McKinsey Global Institute’s national figure of 33% of working time on collecting and processing data (A Future That Works, 2017). Wages are survey hourly wages, not adjusted for inflation, so the total is conservative.

Compare by
Show as
Timeframe
Count
Repetitive manual work Other work Sorted by repetitive manual work ↓
Where this data comes from

Measured (ATUS / BLS, 2003–2024). Average weekly work hours by occupation and industry, and average hourly wages by occupation, from the American Time Use Survey. ATUS does not say what people do inside work time — so the repetitive share cannot come from ATUS alone.

National anchor (McKinsey Global Institute, 2017). About 33% of U.S. working time is spent collecting (17%) and processing (16%) data. Financial Activities is fixed near McKinsey’s ~50% for finance and insurance. Other groups get relative intensity weights, then calibrated so the population-weighted average equals 33%.

Discover
Configure
Test
Ship
Maintain

What we heard

Across 12 contextual interviews and 3 shadowing sessions, three patterns kept surfacing. Users wanted guardrails, not tutorials — they needed the product to constrain choices until they had enough context to explore.

Mental model gap

"I keep looking for a 'save and run' button. I don't understand the difference between draft and published."

P03, fintech ops lead
Terminology friction

"Trigger, webhook, payload — I had to Google half the labels in the sidebar before I could even start."

P07, HR operations
Confidence threshold

"Once I got one workflow running, I felt like I could build anything. But getting to that first win took me three days."

P11, supply chain manager

How we approached it

Rather than a linear tutorial users would skip, we designed a progressive onboarding layer that met users inside the builder — revealing complexity only after each step was validated.

1

Discover

12 contextual interviews and a competitive audit of onboarding patterns in Zapier, Make, and Retool informed our "first win in 15 minutes" target.

2

Test

Three rounds of moderated usability testing with Figma prototypes — each round narrowed the step count and replaced jargon with plain-language labels.

3

Refine

We cut the guided flow from 9 steps to 5, added inline validation, and shipped a "preview before publish" checkpoint based on the #1 request in testing.

What we shipped

The final experience walks users through trigger selection, action mapping, and a test run — with the full builder available but de-emphasized until the first workflow is live.

What changed

73% Task completion on the guided onboarding flow (up from 31%)
4.2/5 System Usability Scale score in post-launch survey (n=24)
6/8 Users completed their first workflow unaided in final usability round

Looking back

The biggest lesson wasn't about onboarding UI — it was about respecting the anxiety behind the click. Non-technical users don't need simpler tools; they need clearer consequences. Showing them exactly what "publish" does, and letting them undo it, mattered more than any tooltip we wrote.

If I were starting over, I'd involve customer success earlier. Their ticket data surfaced the terminology problem weeks before our interviews confirmed it — and that time could have gone into iteration instead of validation.