Latch records real work to capture tribal knowledge
Stefan Kalb is Cofounder & CEO of Latch, which records people doing real work and uses an AI interviewer to turn those sessions into a structured knowledge graph. The company focuses on capturing exceptions and the judgment calls that never make it into process documents, so agents can query the knowledge people actually use rather than idealized SOPs.
Automations fail when they are built from sanitized process documents; capture work while it happens and structure those sessions into owned, reviewed knowledge so agents can handle real exceptions. Prefer pilots that demonstrate exception handling and avoid large documentation initiatives that rely on experts narrating their own work.
What’s the quick origin story of your business, and what makes what you do genuinely different from the other options your customers are looking at?
I’ve started two companies, and both came out of watching smart people work around a problem everyone had stopped noticing.
The first time it was food. I founded a healthy high end ready-meal company in Seattle called Molly’s (like the pret a manger of the west coast). Our customers included grocery stores like Whole Foods, corporate cafeterias like Microsoft, Amazon, and Boeing, and university and hospital campuses. We became a huge part of our customer’s business and I quickly found out about the staggering food waste in this industry. Of the top 10 grocery stores in the US, 3 were our customers, and their fresh food waste was over 30%!
That discovery became Shelf Engine, which used machine learning to forecast demand for perishables and handle ordering for grocers. We went through Y Combinator, raised from General Catalyst and GGV, and sold to Crisp in 2025.
Latch came from the same instinct – just pointed at knowledge instead of food! So much work happening inside companies does not have to be done by a human, but companies are spending real money on AI agents that have no idea how the company actually works. There might be process doc that describe a sanctioned path, but actual job is full of judgment calls, exceptions and workarounds that live in two or three people’s heads and were written down nowhere.
The difference is in what we capture.
Latch records someone doing real work while an AI interviewer asks the follow-ups a good interviewer would ask, things like why you skipped that field, or what happens when this vendor runs late. Then we structure the answers into a knowledge graph that agents query directly over MCP. My favorite way to describe it is that we’re integrating the one data source that never had an API, which is your people.
Where does most of your new business come from today, and where do you wish it came from?
Most of it is me, and we’re early enough that I’d rather say so than dress up a pipeline. Warm intros, my network from the Shelf Engine years, our investors, conversations that start at applied-AI events or in the Slack groups where people are actually building agents. The second biggest source is people who already failed at this once. An internal AI lead or a consultant builds something on top of a process doc, watches it break the first time a real exception shows up, and goes looking for why.
I want the calls to start coming from operations leaders. Today the person who finds us is usually technical, and then they have to go convince the ops side to hand over time with the two experts whose knowledge we need, which is a hard ask when those two people are also the ones holding the week together. Tribal knowledge, retirements, the same exception escalating to the same two people every Thursday. The ops leader lives with all of it and just doesn’t have language for it yet as an AI context problem.
The day a VP of Operations opens a call by telling me their agents don’t understand how the company actually works, the category exists. We’re not there yet.
What’s one thing you’ve actually handed over to AI in the last year, and one thing you tried it on and went back to doing the human way?
Handed over: the first pass at turning a recorded work session into structured knowledge. That’s our own product and we run the company on it. When someone here works out how we do something, like how we scope a pilot or handle a particular onboarding edge case, they record it instead of writing a doc, and the structuring happens on its own. We stopped writing SOPs entirely. Engineering is the other one. A large share of our first-draft code is model-written now, which in Seattle in 2026 is about as controversial as saying we use version control.
Went back to the human way: outbound. We ran AI-written prospecting sequences for about six weeks. The output came back grammatical, personalized-looking and completely inert. What makes a cold email land is a specific observation about how that company operates, and the model only has their website to work from. I’d rather send fifteen emails a week that say something real than four hundred that say nothing.
Have you noticed a change in how customers find you or what they already know before they reach out? Is anyone arriving through AI assistants like ChatGPT yet?
Yes, and the change that interests me is what people already believe by the time they show up. They arrive with vocabulary they picked up from a conversation with AI. They usually also arrive with a half-formed answer, and that answer is almost always to put your documents in a place that’s accessible to your AI. I end up spending the first ten minutes of a lot of calls explaining why documents were the wrong place to look.
On arriving through assistants: some of them, and measuring it is genuinely hard. A referral from ChatGPT or Claude doesn’t land in analytics the way a Google click does. What I have is anecdotal and fairly consistent. People tell me on calls that they asked an assistant how to give an agent context about their company’s workflows, and either our name came up or the concept did and they searched from there.
If you were advising someone in your industry on all of this for the next 12 months, what would you tell them to actually do, and what would you tell them to ignore?
Start by asking yourself what would be transformational for your company. Grow topline? Increase margin? Reduce turnover? Improve training? Automate operational roles? From there you have two choices. Either enable a centralized team to build and execute against that goal, or enable everyone in the company to execute against that goal.
The trade offs are real. On one hand you have everyone in your company utilizing AI, on another hand, most of them aren’t technical enough to do much. Either way, pick a path and invest it and demand results. Sometimes you can do things like weekly demo days. Everyone on Friday at 10AM, show me what you’ve built with AI that’s making an impact toward these goals. Sometimes you need to give clear and hard directives, like “automate 80% of the quoting process such that our team goes from 15 people to 5.”
The risk is that your initiative and the automations fail. This is what’s happening at most companies. Why? Because they don’t automate how people actually work. They start building something that seems ideal but isn’t practical. How do you insure against this? Capture work while the work is happening. A retrospective interview in a conference room gets you the sanctioned version, because exceptions only surface when the exception actually happens. Give every captured workflow a named owner and a review date, since this decays like any other asset and decays faster wherever the underlying process keeps changing.
Once you’re building the solution, be skeptical of any agent demo that stays on the happy path. And I’d stay away from launching a big documentation initiative, because those have been failing since the nineties for a reason, which is that experts make poor narrators of their own expertise. Ask a veteran machinist how she decides a job is finished and you’ll get three bullets. Stand behind her for an hour and the real number is closer to forty.
Thank you to Stefan Kalb and the team at Latch for sharing what actually worked with Leaders Perception readers.
Want to share your playbook?
Leaders Perception publishes founder and operator interviews every week. A few questions, about five minutes, no calls. It is free and everyone we accept gets published.