The filing cabinet finally does the work

a16z recently published a conversation with Steijn Pelle and Frédéric Renken, the founders of Lassie, an AI company that runs the financial back office for more than 700 US dental practices. Alex Rampell led their $35M Series A and led the discussion. It's the best hour on AI for small business I've heard this year, and nearly every minute of it maps onto what we're building at Kula for studios and gyms.

Here's what they said, where we agree, and where we've made a deliberately different bet.

Software stored the work. People still did it.

Rampell opens with a history lesson. The origin of software was taking filing cabinets and putting them in databases. Airline reservations became Sabre. HR filing cabinets became PeopleSoft. Accounting cabinets became QuickBooks. Storage went digital. The work stayed human. An HR department in 2000 wasn't meaningfully smaller than one in 1950, the files just moved from a locked drawer to a locked server.

A dentist could see every overdue invoice in the system. Chasing them was still a person's job. Steijn's own dentist, the top-rated practice in his area, was spending 200 hours a month on paperwork.

The claim at the heart of the episode: AI is the first generation of software that does the work instead of storing it. Lassie's numbers back the claim. Their first agent hands a practice back around 30 hours of labor a month, autonomously, with payment posting at 98% automation.

If you run a studio or a gym, this arc is yours too. Your booking platform, your payment processor, your access control system, your website — they digitised the records of your business. Understanding what's actually happening in it — who's about to leave, which classes make money, whether the intro offer converts — stayed manual. Usually it's the owner. Usually at night.

The part of the interview that mattered most

Lassie started in 2020, before the models were good. So they built the two things an agent needs regardless of the model: context, meaning access to all the historical data of the practice, and tools, meaning the integrations to actually act. The intelligence layer wasn't intelligent yet — early on, the founders were literally the humans in the loop, processing payments by hand.

Then the models got good. And because the context and tools were already built, they could swap better intelligence straight in. The product got smarter almost overnight, on someone else's R&D budget.

That is the substrate thesis in a single anecdote. The model is not the moat and it is not the product. Models are rented. They improve on someone else's schedule and they're available to everyone, including your competitors. The durable asset is the layer that makes a model useful: connected systems, one governed definition of the business, and safe tools to act. The substrate wins.

Models don't know your business

The second insight is one we've written about before, so it was striking to hear it from a company in a completely different industry. Frontier models are trained on effectively the whole internet and they still can't do this work, because the workflows aren't on the internet. They live in the heads of office managers and in actual filing cabinets. Frédéric assumed the reasoning models would just know how to bill an insurance claim. They don't.

Worse, every system in the ecosystem defines things slightly differently. Lassie had to build an ontology so that every system agreed on what an insurance claim and a patient payment actually are before an agent could safely touch anything.

Fitness has the identical problem. A "member" in your booking platform, in Stripe, in your access control system and on your ClassPass ledger are four different records with four different IDs and four subtly different meanings. Ask a raw model a question across those silos and it joins them by guesswork. You get a confident answer built on invented joins, confidently wrong.

Kula Intelligence exists for exactly this. Ingest the systems. Define one truthful model of the studio member, visit, plan, teacher, dollar. Apply it as a governed layer that any AI can safely use.

The same truth, two different bets

Here's where the paths split, and I think both choices are right for their markets.

Lassie builds the substrate, the agent and the interface as one closed loop. They take a single job posting insurance payments, to near-total automation, then move to the next job. For US dental billing that's the correct call. The work is rule-bound, the payer policies are documented, correctness is everything, and the buyer wants to hand the job off entirely. The incumbent they replace isn't software. As Rampell put it, the incumbent is Betty, and she quit two weeks ago.

Kula's bet is different: build the substrate, don't build the assistant. A studio owner connects their systems to Kula Intelligence and asks questions in the assistant they already use, Claude, ChatGPT or Gemini. Ask anything about your studio, grounded in your own data. Then act on it. It's self-serve at US$149 a month, and a single-site studio's data loads in about half an hour.

Why the different bet? Because studio work is less rulebook and more judgement. Retention, pricing, the timetable, teacher relationships. The scarce resource in a studio isn't claim-posting labour, it's truthful understanding. So we start there, and automation grows outward from it as skills on the same substrate: the At-Risk screen, the Yield screen, and a growing set built by us and by others. Lassie automates a known job to free the human. We give the human leverage over the judgement calls first, and automate from there.

The people who deliver the craft

The line from the episode that travelled furthest came from Lassie's first customer: the agent isn't replacing humans, it's freeing them from wearing so many hats. The dentist got his evenings back and coaches his kids' soccer team now.

In a dental practice, the owner and the practitioner are usually the same person, so freeing one human frees the craft. A studio isn't built that way. The owner carries the judgement calls, but the craft is delivered by a team of teachers and the deepest knowledge of your members lives in their heads, exactly the way Frédéric found the real billing workflows living in the heads of office managers rather than anywhere on the internet.

So the substrate has two surfaces at Kula. The ask model is the operator's: the owner, in their own assistant, with the truth of the business behind every answer. Kula Tribe is the teachers': the same governed substrate, facing the people who stand in front of your members. Lassie frees the practitioner from admin. Tribe puts the substrate in the practitioners' hands because in this industry, teachers aren't overhead to automate around. They're the product.

Underhyped where it matters

The episode's most quoted line is that AI is overhyped in Silicon Valley and underhyped in Iowa. It's underhyped in Penrith and Paddington too. The hard part was never the model. It's reaching busy, non-technical owners with something that works in minutes, not months, which is why both companies obsess over consumer-grade onboarding, and why both were built by doing the work by hand first. Steijn spent months on a bar stool in a dental back office. Kula was built inside real studios.

Nobody is against cleaning up busy work for people who'd rather be teaching, coaching or on the floor with members. That's the whole point. AI allows businesses to be human again.

Kula Intelligence connects your booking platform, payments and member systems into one governed layer you can ask anything in Claude, ChatGPT or Gemini. kula.digital