Source assessment — Lassie / SMB AI frontier

Verdict: read it, then file it under GTM/market — not under workflow patterns. Worth the ~6 min. It's a founder manifesto + recruiting post, well-written, with two reusable lessons and one market frame. Don't mine it for agent-orchestration patterns (it has none); mine it for go-to-market.

Why this is in the vault

A clean, credible worked example of the niche + bottleneck targeting discipline RDCO holds itself to ([[feedback_targeting_system_prioritization_filter]]) — dental admin as the niche, claims/billing busywork as the bottleneck — plus the "schlep-as-moat" / domain-immersion thesis that directly informs how the founder should scope the phData / Lionsgate regulated-domain contract work (do the work yourself first to learn what "correct" means before building the grounding evals). Filed as a GTM/market reference and an investing thesis-candidate seed ("vertical full-back-office AI agents for regulated SMBs"), explicitly NOT as a workflow-patterns source.

What it argues (the spine)

SMBs that own the economy's plumbing — dental, medical, plumbing, nail salons — burn ~200 hrs/month and ~$200k/yr on admin staff they can barely hire or keep. The owner ends up doing paperwork at 11 PM.
Prior software only rearranged the work ("a dashboard, a queue, a form... but someone still has to click"). AI is the first thing that can do the job: understand messy context, move across systems, complete the task.
Lassie's claim: autonomous systems that run an SMB's back office. 700 businesses / 49 states, ~30 hrs/mo saved per practice (up to 190), growing by word-of-mouth.

The two reusable lessons (this is the value)

Schlep-as-moat. Their edge isn't a model — it's domain immersion. The team worked inside dental offices for months: reconciled millions in insurance payments, submitted thousands of claims, billed hundreds of patients. That's how they learned what "done correctly" means in a regulated, brittle-integration vertical. Explicitly invokes PG's "schlep blindness." The unglamorous work is the moat.
Collison installation, taken to the extreme. They installed the product in person in the first 100 practices — over-the-shoulder onboarding from Florida to Kansas to Oregon — to learn how non-technical owners connect their systems of record, banks, and data sources. "Do things that don't scale," then encode what you learned into self-serve onboarding.

The market frame

"Vertical AI agents that run the whole back office of a regulated SMB." The article even links Garry Tan's "half the AI-agent market is one category, the rest is wide open." Demand thesis rests on the lump-of-labor fallacy being false — owners reinvest freed hours into growth, not layoffs.

Mapping against Ray Data Co

Targeting-system filter. This is a clean worked example of niche + bottleneck done right (dental admin = the niche, claims/billing busywork = the bottleneck). The schlep-as-moat lesson is the antidote to the "slop cannon" — depth in one vertical beats breadth. See [[feedback_targeting_system_prioritization_filter]].
Direct rhyme with the phData / Lionsgate engagement. Lassie's hardest problem — "dozens of insurance portals, each fails with its own cadence... it has to be done completely and correctly... in regulated industries" — is structurally the same as the Box/Copilot music-rights-contract problem: brittle multi-source integration + correctness/citation in a regulated domain. The "do the work yourself first to learn what correct means" lesson applies directly to scoping the contract-Q&A grounding evals. See [[2026-06-03-copilot-studio-contract-rag-citation-eval-design]].
Investing ideation. "Vertical full-back-office AI agents for regulated SMBs" is a thesis-candidate category worth a future /investing pass, distinct from the infra/memory capital-cycle lane.

What it is NOT

Not a workflow-patterns source — zero orchestration content, so it does not feed [[~/rdco-vault/06-reference/2026-06-04-agent-workflow-patterns-catalog.md]]. Also a recruiting CTA at the end; read the market claims (700 businesses, 30 hrs/mo) as founder-reported, not independently verified.