What Lassie Got Right: The Founders Who Did the Work Before Writing the Code

Published: June 4, 2026 Reading time: 9 minutes Type: Case study / Builder takeaway

TL;DR

Lassie, a startup building AI agents for dental practices, raised a $35M Series A led by a16z. The funding isn't the story. The story is what the founders did before they raised a dollar: they worked inside a real dental office processing payments by hand, then built agents that actually replace admin work — not copilots that create more of it. There's a builder lesson here that matters more than the round size.

What Lassie Does (and Why It Works)

Lassie builds AI agents that run the back office of a dental practice. The agent reads insurance claims, reconciles payments, posts to the ledger, and flags the underpayments and denials that would otherwise slip through.

According to a16z, the product already works with more than 700 practices across 49 states, generates more than $10M in annualized revenue, and provides practices with 250,000 hours of labor a year.

For larger practices, that's more than 100 hours per month in recovered admin capacity. The average is at least 30 hours — roughly the cost of a full-time staffer.

But the metric that matters most, according to the dentists themselves, isn't speed. It's accuracy. Catching the claims that were quietly being underpaid is where the real value lives.

The Part That Actually Matters: How They Built It

This is the lesson, and it's hiding in plain sight.

The founder, Steijn Pelle, was a product leader at Robinhood and Coinbase. His co-founder Frédéric Renken had built product at Superhuman and Uber. Both had every credential you'd expect from a Series A–ready team.

Before they wrote a single line of code, they spent months inside a real dental practice.

It started with Pelle's own dentist, Dr. Kwon, who was running a 900-square-foot practice in Palo Alto. His office manager had just quit. Kwon was staying up until 2am each night mailing hundreds of insurance statements, six days a week, falling nearly $1M behind in payments, and worrying he'd run out of cash. Pelle, his patient, knew there had to be a better way.

So instead of giving advice, Pelle embedded. He worked for a few months as a quasi-employee at the practice — perched on a bar stool near the fridge, learning the workflow firsthand. At night he'd call Renken in Canada, and the two would work out how to automate what Pelle had just done with his own hands.

Five years later, that practice has grown from 900 square feet to 4,000 square feet, with twenty staff seeing nearly a hundred patients a day. And Lassie works with over 700 of them.

a16z general partner Alex Rampell described Pelle's approach as "out-obsessed" the competition. He compared Pelle to a method actor training for a role.

The comparison is fair. Most AI products for small businesses are built the opposite way: founders imagine a workflow, build a generic agent, then try to sell it to people who actually do the work. Lassie was built by people who had done the work — who knew which forms got rejected, which underpayments got missed, which exceptions ate up the night.

Why "AI for Small Business" Usually Fails

The category is full of demos and short on production wins. Most products fall into one of two traps:

1. The copilot trap. The AI drafts a response, suggests a next step, or generates a summary — and a human still has to do the actual work. This is what every GPT-wrapper startup ships first because it's the path of least resistance. The customer ends up with two things to review instead of one.

2. The generalist trap. A "do anything" agent that promises to handle any small business workflow. The pitch is appealing, but the implementation is impossible. A dental practice's insurance reconciliation has nothing in common with a plumber's invoicing workflow. A generalist model hits the same wall a generic ERP hit in 2005: too many edge cases, too much local context, not enough leverage on the actual work.

Lassie avoided both traps by doing something less glamorous: it picked one industry, learned it deeply, and built the integrations required for an agent to operate inside the workflow — not next to it.

The a16z post puts it well: "Lassie is not a copilot that creates more work for the practice to review. It does the work."

How Small Businesses Can Identify Their First AI Agent Opportunity

If you run a small business and you want to find your first automation target, here's a useful filter. Score each candidate workflow against these four questions:

1. Is it structured and repetitive? Insurance claim reconciliation, invoice chasing, schedule confirmation, inventory reorder — yes. Custom client negotiation, novel design work, exception handling — not yet.
2. Is it high-volume? A workflow that happens fifty times a month is a better target than one that happens twice. The agent's per-task cost has to be small relative to the value.
3. Is the cost of getting it wrong measurable but bounded? Missing an insurance claim is bad, but it's recoverable — you notice, you fix it, you resubmit. That's a great candidate. Sending the wrong dosage of medication is a different category of risk.
4. Do you have the integrations? An agent can only operate where the data lives. If your workflow is buried in a paper filing cabinet with no API, no OCR pipeline, and no way to verify output, the agent can't actually do the work — it can only describe it.

Lassie's first workflow — reading claims, posting to the ledger, flagging underpayments — scores well on all four. It's why their agent actually replaces labor instead of just summarizing it.

The general rule: don't automate a workflow you don't understand yet. Spend a week doing it manually first. If you can't, the agent can't.

What Developers Can Learn From Lassie's Approach

The technical lessons here are uncomfortable for a lot of AI founders.

1. Domain knowledge beats model quality. The competitive advantage appears to come more from workflow knowledge, integrations, and operational data than from any specific frontier model. A generic agent with the right PMS integrations and the right claims-clearinghouse connectors would likely get most of the way there. The model is the easy part.

2. Integrations are the moat. A 700-practice install base, each connected to a different PMS (practice management software), claim clearinghouse, and payment processor, is the kind of moat that doesn't show up in a benchmark. It's also the kind of moat a generic model can't replicate by being smarter.

3. "Agents that do the work" requires a knowledge layer that learns from every completed workflow. a16z flagged this directly. Every claim Lassie processes, every reconciliation it gets right, every exception it flags — that data makes the next workflow easier. The longer the product runs in production, the harder it is to catch up to.

4. Founders who do the work earn the right to build the product. Pelle didn't pitch a dental practice from a WeWork. He filed claims. He stayed up until 2am. He learned which paper was which. By the time he had code to write, the spec was already in his head. The bar for AI products is going up; the bar for understanding the underlying domain isn't going down.

What Would This Look Like Technically?

If you wanted to build something like Lassie for a different small-business vertical, here's the technical shape of the system. The model is only one component. The real product is the workflow.

A modern version would likely combine:

• LLMs for reasoning and exception handling — natural-language claims descriptions, free-text notes from front-desk staff, ambiguous insurance correspondence, "why was this denied" summarization.
• OCR and document-processing pipelines — digitized EOBs (explanations of benefits), scanned insurance cards, intake forms. Dental claims arrive in dozens of formats; the OCR layer normalizes them.
• Workflow orchestration — a state machine that tracks each claim through submission, follow-up, denial, resubmission, and posting. This is the backbone. Without it, you have a chatbot, not an agent.
• Practice-management software integrations — read/write access to systems like Dentrix, Eaglesoft, Open Dental. This is where the actual work happens.
• Payment and accounting integrations — Stripe, Plaid, and ledger systems (QuickBooks, Xero). The agent has to post reconciled payments somewhere a human can audit.
• Audit logging — every action the agent takes must be traceable, replayable, and exportable. Regulated industries require this. So does the agent's own debugging.
• Human review queues for low-confidence decisions — when the model isn't sure, it doesn't guess. It routes the case to a human and learns from the resolution.

The pattern matters more than the stack. Most "AI for X" failures aren't model failures. They're workflow failures — the agent has no durable state, no integrations, no audit trail, and no graceful escalation path. The result is a tool that demos well and breaks in production.

The lesson: build the workflow first, then drop the model in. Not the other way around.

A Note on Risk: Not Every Workflow Should Be Automated

AI agents are powerful, but they are not appropriate everywhere.

Tasks involving legal decisions, medical decisions, financial approvals, or high-risk judgment still require human oversight. An agent can prepare a tax filing, summarize a contract, or surface an underpayment — but signing the return, accepting the liability, or committing the funds should stay with a human.

The best AI-agent opportunities share three properties:

• Repetitive — the workflow happens often enough to justify the integration cost
• Measurable — you can tell when the agent got it right and when it didn't
• Reversible — a mistake can be caught and corrected without catastrophic cost

Lassie scores well on all three. A claim that gets underpaid by $40 can be refiled. A patient reminder that goes out twice is annoying, not dangerous. The same cannot be said for a misdiagnosed X-ray or an unauthorized wire transfer.

The rule of thumb: if the worst-case outcome of an agent's mistake is recoverable, automate it. If it isn't, the human stays in the loop.

A Note on the Funding (and Why We're Not Leading With It)

Lassie raised $35M at a reported ~$250M valuation. Co-founders of Superhuman, Plaid, Wise, and Reforge joined as angels, along with Dr. Edward Zuckerberg — yes, Mark Zuckerberg's father, who is also a practicing dentist.

The round is real. The valuation is real. The investors are real. None of it is the point.

The reason this story matters for Build With Abdallah readers is not "another AI startup raised a lot of money." It's that the founders did something specific, repeatable, and unglamorous — and it produced both the product and the round. You can copy the approach. You cannot copy the round.

Source Validation

Confidence: High. Two independent primary sources (a16z + Upstarts), both dated June 3, 2026. All factual claims cross-verified.

What to Read Next

If you build AI products for small businesses, the most useful next move is to pick one workflow you understand deeply and ask: could an agent do this end-to-end, with the integrations I have today, with bounded error I can measure? If yes, you have a product. If no, you have a demo.

That's the Lassie test. It's harder than it sounds.

Build With Abdallah — Practical AI, Laravel, Python, Raspberry Pi, and automation tutorials for builders who ship.

📧 buildwithabdallah@gmail.com