Summary

Lassie’s wedge is not dental software but an understaffed labor budget: roughly 160,000 U.S. dental practices each spend about $200,000 annually on administration. Dr. Quan, despite being the number-one-rated doctor on Yelp, was spending 200 hours a month on paperwork; Lassie already charges five figures for an agent doing about 30 hours. Olivia frames the gap as “AI is overhyped in Silicon Valley but underhyped in Iowa.”
The product thesis is that software must perform work, not merely digitize the filing cabinet. Alex Rampell argues legacy systems stored records while leaving headcount broadly intact; agents can now edit those records, chase invoices, explain benefits, or complete onboarding. Fintech enlarged software markets through payments—his Toast example implies a 2% take on a $5 million restaurant creates $100,000 of revenue—but charging for labor makes the opportunity “orders of magnitude bigger.”
Lassie reached roughly 98% automation by first having its founders perform the work and then “automating away our own problems.” Starting in 2020, it built the context layer and tools before reasoning models were capable enough, then upgraded the intelligence as models improved. Frédéric Renken targets roughly 95%-plus automation before launching a job, accepting a small exception queue rather than waiting for an impractical 100%.
The defensibility comes from replacing an absent worker rather than adding an AI feature to an incumbent platform. As Alex puts it, “The incumbent was named Betty, and she quit two weeks ago”; reproducing Betty requires read-and-write integrations, a shared ontology across inconsistent systems, historical workflow data, and agents trusted to act autonomously. His enduring rule remains: “The battle between every startup and incumbent comes down to whether the startup gets the distribution before the incumbent gets the innovation.”
Distribution and implementation—not model access—may be the binding constraints in bringing agents to mainstream businesses. A dentist may not be on LinkedIn or in conventional SaaS databases, and may abandon the product if it does not work within a couple of months. Lassie is therefore pushing onboarding toward self-service: connect the bank, practice-management system, insurance portals, and business details, then configure the agent under the hood. Its consumer-product benchmark is faster: at Robinhood or Superhuman, users had about 48 hours to see core value.
The expansion plan is dental first, another underserved medical-office category likely second, and eventually every small business. Steijn Pelle sizes dentistry alone as a roughly $1 billion recurring-revenue market, then sees reusable primitives across verticals: systems of record, customers, appointments, payments, and communications. The end state is a business agent interacting with consumers’ personal agents and counterparties’ agents—“every small business should run itself.”
The remaining technical frontier is proprietary workflow knowledge plus the digitization of stubbornly physical operations. Large models still do not know payer-specific procedures or the tacit expertise held by office managers, while Steijn says roughly 70% of payments remain on paper. Federal requirements for direct-deposit options and digital file formats now combine with better models to make automation possible—and could expand capacity where demand for dentists, plumbers, and primary-care doctors exceeds available supply.

Deep dive

1. Two hundred hours of paperwork revealed the real product

After six years in Silicon Valley, Steijn was at Robinhood looking for a “hard problem” when his dentist, Dr. Quan, invited him behind the front desk. The number-one-rated doctor on Yelp was spending 200 hours a month submitting claims, billing patients, and covering vacancies himself.
Steijn initially wondered whether Dr. Quan was simply an anomaly: the practice was highly rated and already used modern technology. Work inside other offices—including for a gastroenterologist in Scranton, Pennsylvania—showed the same manual processes and suggested hundreds of thousands of small businesses were still doing this work by hand.
The earliest demand signal was unusually intimate access. Despite the founders lacking billing experience and facing questions about HIPAA, security, and access to sensitive information, Dr. Quan told them to work in his office; another doctor offered access to whatever they needed or to teach them “like my son.”
Doctors would discuss the problem for hours because it was not a “vitamin.” Administration kept them awake, made them consider quitting, and displaced the patient care that had drawn them into medicine—the conditions for selling completed labor rather than another tool.

2. Software becomes valuable when it edits the filing cabinet

Alex traces software’s origin to digitized filing cabinets: SABRE replaced airline reservation files, while PeopleSoft, LexisNexis, QuickBooks, and NetSuite did comparable work elsewhere. Green screens made records easier to change, but “people still had to do the work.”
His deliberately exaggerated test is headcount: HR departments serving similarly sized companies did not necessarily shrink between 1950 and 2000. Guards protecting paper records were replaced by IT departments and CISOs protecting digital ones; storage changed more than the underlying workload.
Agents turn that passive database into an actor. An HR system can run a background check or explain benefits; an accounting system can identify overdue invoices and prompt a call for payment. “The filing cabinet should be doing” the next action because the work around stored information is orders of magnitude larger than storage itself.
Fintech supplied an earlier market-expansion template. A restaurant grossing $5 million would not have bought $100,000 of MS-DOS software in 1985, Alex argues, but Toast can effectively earn that amount through a 2% payments take. Labor automation surrounds that already-large financial-services opportunity with a much larger economic circle.

3. Labor scarcity makes automation additive rather than merely substitutive

Steijn’s first dentist, Ronald Sloop, retired partly because he lost the employee who handled the books and other work and could not face replacing her. He sold the practice to a junior practitioner and left dentistry.
That story changes the displacement framing: “It’s not like AI is going to take the jobs. In many cases, you can’t find somebody.” Alex describes this as demand to the right of the supply-demand equilibrium—valuable services never supplied because a specialized worker, such as a Dutch-speaking dental receptionist, would cost too much for rare usage.
Steijn says there are roughly 160,000 U.S. dental practices spending around $200,000 apiece on administrative costs. Lassie’s first agent supplies about 30 hours of monthly labor for a five-figure price, against as much as 200 hours of work; acute pain makes adoption quick and drives dentist-to-dentist referrals.

4. Trust came from doing the job before automating it

Lassie began in 2020, before reasoning models existed in their current form. Frédéric’s team still built the durable prerequisites: historical context from patient and practice records, plus tools capable of acting on those systems. As models improved, Lassie could replace the intelligence layer and “just get smarter over time”—a tailwind he admits involved some luck.
The founders initially became “the humans in the loop,” taking responsibility for the work and automating whatever caused them pain. That immersion mattered because an SMB often has nobody available to operate a new tool; handing a doctor another dashboard merely leaves the owner working at night.
Frédéric wants approximately 95%-plus automation before selling a new job, not necessarily 100%. A practice that once updated 200 patient ledgers and checked insurance portals and bank records 200 times weekly may retain only a handful of exceptions, still saving 10–20 hours.
Autonomy raises the correctness bar: reconciling insurance payments or billing patients cannot simply leave a human in the loop while software engineers decide what gets deployed. Lassie reports roughly 98% automation, while thousands of staff members provide feedback on the long tail; Dr. Quan’s practical result is time to coach his children’s soccer teams and attend their games.

5. The best incumbent may be a departed office manager

Alex’s TiVo lesson is that an innovative feature attached to someone else’s distribution earns a “control discount.” A cable company can copy it, license it, or acquire it without preserving much of the startup’s economics; startups should often own the boring pipe first, as he realized when he wished TrialPay had instead built Stripe-like payment processing.
His rule survives AI: “The battle between every startup and incumbent comes down to whether the startup gets the distribution before the incumbent gets the innovation.” AI may help mediocre engineers at large companies become competent faster, increasing risk where platforms such as Workday already own customers.
Labor categories offer a different contest because no software incumbent performs the job. “The incumbent was named Betty, and she quit two weeks ago,” Alex jokes; alternatives are the doctor, a billing agency, or an outsourced worker. Market labels obscure the convergence: “AI is software. Software is AI.”
The missing incumbent also creates technical work. Lassie had to build read-and-write integrations across every system Betty touched, reconcile inconsistent definitions of claims and payments through a shared ontology, then build agents on top. Steijn argues that this years-long “schlep” makes the product harder to build and more defensible.

6. Consumer-grade onboarding is the last-mile infrastructure

Olivia’s implementation pushback is decisive: this is not an app a dentist downloads and finishes configuring alone. She and Alex sharpen the distribution problem with “AI is overhyped in Silicon Valley but underhyped in Iowa”—the potential user is busy, nontechnical, and difficult even to reach.
Steijn imports a consumer-product constraint from Robinhood and Superhuman: those products had roughly 48 hours to deliver core value or lose the user. Doctors are similarly demanding, though Steijn says a practice may give Lassie only a couple of months if the product does not work. Lassie needs to be plugged in and perform the job.
Lassie’s nearly self-service flow connects the bank account, practice-management system, and insurance portals, confirms doctors and business information, and configures the operational details underneath. Steijn says one or two pieces remain, with an almost self-serve setup potentially months away.
Robinhood provides his reference model. Opening a brokerage account once required visiting an office; its product absorbed KYC and bank linking into a consumer flow without a human operator. Lassie is attempting the equivalent across insurers and fragmented practice systems.
Go-to-market requires reconstructing a market conventional SaaS databases barely see: map every dentist, owner, system, and intent signal such as an Indeed job posting. Dr. Sloop may not be on LinkedIn or in the relevant databases, so Lassie must develop messages and channels that cut through to thousands—and eventually hundreds of thousands—of scattered owners.

7. Vertical focus is the route to a horizontal small-business agent

Steijn’s three-step plan begins with dentistry: 160,000 practices, about $200,000 of administrative costs each, and what he sizes as a roughly $1 billion recurring-revenue market. The likely next move is another underserved medical-office type with a large TAM and a need for consumer-simple onboarding.
At an abstract level, small businesses share primitives: a system of record, customers—patients in healthcare—appointments, payments, and communications. The eventual business agent could interact with a consumer’s personal agent and with agents at insurers or other counterparties, then extend to “salt-of-the-earth people” from Iowa and Paducah to Amsterdam and Hamburg.
Frédéric keeps the customer profile narrow because selling labor creates a stronger promise than selling functionality. An incompatible practice leaves Lassie choosing between manual fulfillment and offboarding; onboarding is therefore treated like a measured “story or a playbook or like a movie,” with value checkpoints reached on schedule.
Product design similarly begins with eliminated work. Traditional patient-billing software sends statements and accepts payments, but staff time goes into validating the amount and answering “Why do I owe this?” Lassie considers the job shipped only when the office no longer needs to spend meaningful time on those fuzzy communications.

8. Easier operations could expand capacity before they compress margins

Steijn’s hiring standard remains conventional at the top end: steep learning curves, ambition, and top-five-percentile engineering or sales ability. The additional filter is AI fluency—whether candidates expect coding, finance, and organizational design to change—and the company hopes to ship twice as much, move four times faster, and generate four- or fivefold output.
Alex’s pushback is that materializing “Betty” for everyone could make small businesses easier to start but harder to remain. If labor accumulation is part of a local operator’s moat, removing it might crowd the market, erode margins, and produce the Yogi Berra paradox: “It’s so crowded nobody goes here anymore.”
Steijn’s current answer explicitly depends on demand not being capped. He sees perhaps twice as much demand as dentists and plumbers can currently supply; America might support half a million dentists, or existing dentists might treat twice as many patients, while better access could provide more of the community primary-care experience he describes from the Netherlands.
The optimistic mechanism is more time for craft, not fewer craftspeople: baking pies, polishing nails, cleaning teeth, or treating patients instead of processing forms. Lassie’s launch resonated, Steijn says, because few people oppose removing busy work that prevents small-business owners from serving customers.

9. Workflow knowledge and paper remain the technical frontier

Frédéric’s surprise is that enormous models trained on vast datasets “actually don’t really know how to do any of this work.” Payer-specific SOPs and office-manager knowledge are often absent from the internet; Lassie must collect documents and infer procedures from historical ERP data rather than assume general reasoning supplies them.
Steijn says Lassie initially assumed the latest reasoning models would know these workflows because they were trained on so much data, but found that they lacked many of the intricacies. Historical data lets the company infer workflows, while staff feedback helps handle cases the agent cannot yet do.
Alex asks what happens when “the marginal cost of arguing goes to zero”: insurers can deny repeatedly while customers can contest repeatedly. Steijn’s response is that dental reimbursement is regulated and documented—a crown submitted with the required X-ray and narrative should be covered—and insurers also need good in-network dentists to retain employer plans.
The more immediate obstacle is physical infrastructure. Steijn once opened envelopes containing roughly $100,000 of checks and manually reconciled their itemized statements; he says the statistics are that roughly 70% of payments remain on paper. The federal government has mandated that the industry offer direct deposit as an option and create digital file formats for itemized invoices, creating a tailwind while Lassie productizes the insurer-by-insurer conversion that would otherwise consume about 50 hours of a doctor’s time.

Is Offense or Defense Dominant? FAR.AI's Adam Gleave on the AI Security Leaderboard Jensen's Open-Weights Letter | Google Cloud Grows 82% But The Market Tanks