Observation beats specification.
People can't accurately describe their own workflows. Decades of business process management research confirms it, and our recordings prove it on every engagement. We build from evidence, not interviews.
AI can't automate what your company can't describe.
Latch was founded in 2025 by Stefan Kalb and Jared Kofron to solve the prerequisite that every failed AI rollout shares — the gap between how work is documented and how it actually gets done. Headquartered in Seattle and backed by $8M in seed funding led by FUSE, with participation from Y Combinator CEO Garry Tan, Liquid 2 Ventures, and Soma Capital.
The reason your AI
strategy keeps stalling.
Every company wants to adopt AI. Most can't.
95%
of enterprise generative AI pilots fail to produce measurable P&L impact.
— MIT NANDA, 2025
It's not the agents.
It's not the models.
- The agents aren't the problem.
- The models aren't the problem.
- The problem is that you can't automate a workflow nobody can describe.
Agents need instructions, not prompts.
AI agents need instructions. Not prompts — instructions. Step-by-step descriptions of what happens, in what order, with what exceptions, using what systems. That's what an agent runs on. And almost no company has it.
What companies have instead.
What companies actually have is the opposite: workflows that evolved over years inside inboxes, spreadsheets, and tribal knowledge. The person who knows how the order intake process really works left two years ago. The new hire learned it by shadowing someone who learned it by guessing. The Confluence page is from 2019.
Every existing fix fails the same way.
When you hand that to an agent, the agent fails. When you hand that to a consultant, you get a six-figure invoice for a process map that's wrong by the time it's delivered. When you hand that to a software vendor, you get a tool nobody uses.
Stefan saw this at Shelf Engine, where deploying AI inside grocery operations meant first untangling years of undocumented ordering decisions. Jared saw it across Pioneer Square Labs, Flux, Rover, and Glowforge — engineering teams repeatedly asked to automate workflows that no stakeholder could fully articulate.
Four principles that shape every engagement.
These aren't marketing values. They're the operating constraints we use to decide what to build, what to refuse, and how to work with customers.
.svg)
The Latch Session is the asset.
A Latch Session is a structured, machine-readable representation of how a workflow actually operates. Think of it as the JPEG of work — a portable format that any system, any AI, or any team can read. The automation is downstream. The format is the foundation.
.svg)
Compounding leverage over one-off wins.
A single automation is a project. A growing library of Latch Sessions is operational infrastructure. Every workflow we capture makes the next one faster to model, simulate, and deploy.
.svg)
Forward-deployed, not handed off.
Our engineers stay embedded with the customer. No transition documents. No support ticket queues. The person who built your workflow is the person who improves it next quarter.
Operators who've lived the problem.
Latch was founded by people who spent years inside operationally complex businesses — and got tired of watching good AI ideas die against bad process documentation.
Stefan Kalb
Co-Founder & CEO
Stefan is a Seattle entrepreneur whose previous company, Shelf Engine, used AI to reduce food waste in grocery stores by predicting optimal ordering for perishable goods. Before that, he founded Molly's, a food company supplying salads and sandwiches to Seattle-area cafes and hospitals — an experience that exposed him to the operational friction inside traditional businesses.
The idea for Latch emerged after he kept fielding calls from non-technical operators trying to figure out where AI actually fit in their companies. “I would have dreamed of having Latch back then,”
he said.
Jared Kofron
Co-Founder & CTO
Jared was a principal software engineer at Pioneer Square Labs and previously built systems at Flux, Rover, and Glowforge.
Talk to us about your operations
The fastest way to understand whether Latch fits is a 30-minute discovery call. We'll ask about one workflow you'd most like to see automated, and tell you honestly whether it's a fit.