ExpertsFMRe-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee
No Priors: Artificial Intelligence | Technology | Startups · September 24, 2026
5 Selected Lessons:
case · 2 priority · 1 framework · 1 heuristic · 1
Three criteria for choosing which industry/company to target
Source
No Priors: Artificial Intelligence | Technology | Startups — Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee (September 24, 2026) · at 17:46 ▸
When this applies
when prioritizing which market or target to commit scarce, highly concentrated resources to
The takeaway
When resources are intentionally scarce and concentrated, rank targets by compounded filters (size, structural durability, technology fit) rather than optimizing any single dimension.
Why it works
Because the firm can only execute roughly one deal per year, each criterion filters out most opportunities, ensuring the rare deal chosen has maximum combined upside from market size, incumbency durability, and AI-fit.
One level up
Scarcity of deployable attention/capital should drive stricter, multi-dimensional filtering, not broader diversification.
Where it stops holding
This stops working when deal flow or resources are abundant enough that diversification across many smaller bets is preferable to concentration on singular large bets.
Detail
hierarchy:
- 1. Scale of the market — must be large enough to build a $100B+ company, since the firm does very few deals
- 2. Incumbent-advantage industry — the incumbent must have durable advantages (brand, retention, regulation, relationship-centricity) that make it hard for startups to compete
- 3. Overlap between what AI is good at and what the organization actually does — can a special product/experience be built from that overlap
rationale: Insurance brokerage satisfied all three: a $2T+ premium market, 90% gross retention with customers not directly paying brokers (startup-resistant), and clear overlap between AI capability and the back-office/underwriting-adjacent work brokers do.
In their words
And then, three, is if we think about what the organization does and what AI is good at today, and we think about the overlap, can we build something special?
Atlas four-layer platform architecture for reusable AI infrastructure
Source
No Priors: Artificial Intelligence | Technology | Startups — Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee (September 24, 2026) · at 20:17 ▸
When this applies
when designing infrastructure meant to be reused across multiple business units or future acquisitions with partly shared, partly unique needs
The takeaway
Architect reusable infrastructure in layers ordered from most general (data/ontology) to most specific (application), so each new deployment only needs to build the thin vertical-specific layer on top of a shared foundation.
Why it works
Separating the homogeneous 80% of a business's operations (data definitions, agent logic, orchestration, application layer) from the vertical-specific 20% lets a platform generalize across industries while leaving room for customization at the top layer.
One level up
Separate the generalizable 80% from the specific 20% in any multi-instance system to maximize reuse without sacrificing fit.
Where it stops holding
This stops working when the businesses being served have less than 80% operational overlap, or when the vertical-specific portion is actually the dominant source of competitive differentiation.
Detail
procedure:
- Layer 1 - Data ontology: encode the business's entities and relationships in code (make the business legible to models)
- Layer 2 - Agent builder: build high-performance agents grounded in that ontology's ground truth
- Layer 3 - Lattice (orchestration engine): instrument workflows using the built agents
- Layer 4 - Artifacts (application builder): build end-user applications on top of the stack
rationale: 80% of a business, broken to atomic units, is homogeneous across companies; 20% is vertically specific — build the platform to capture the 80% once and reuse it.
In their words
if you break down the business to its atomic units, 80% of it is largely homogeneous and 20% of it is vertically specific.
Human engineering is harder than technical engineering in transformation
Source
No Priors: Artificial Intelligence | Technology | Startups — Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee (September 24, 2026) · at 20:17 ▸
When this applies
when deploying any technology that changes how employees do their jobs
The takeaway
Treat change management and emotional buy-in as a separate, often harder problem than the technical build, and budget deliberate attention to it.
Why it works
Employees feel anxiety about job security and identity when automation enters their workflow; successful adoption requires explicit framing (elevating the job, removing rote work, making people feel the organization is winning) independent of the technical capability itself.
One level up
The hardest part of deploying powerful new capability is rarely the capability — it's the people whose roles it touches.
Where it stops holding
This stops working when the workforce has no job security concerns or discretion over whether they use the tool (e.g., brand-new teams with no prior workflow to disrupt).
Detail
rule: Explicitly manage employee anxiety and framing alongside technical deployment of AI agents.
threshold: Stated as: 'We have two problems here at Sequence. We have an engineering problem and a human engineering problem. And the human engineering problem is much more difficult than the engineering problem.'
rationale: Without addressing anxiety about what change means for employees, even a technically excellent system will face resistance or failure to adopt.
In their words
I often joke internally that we have two problems here at Sequence. We have an engineering problem and a human engineering problem. And the human engineering problem is much more difficult than the engineering problem.
Bank South as a cold-start pilot that validated the business model
Source
No Priors: Artificial Intelligence | Technology | Startups — Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee (September 24, 2026) · at 13:08 ▸
When this applies
when a new venture faces a cold-start/chicken-and-egg problem between capital, proof, and talent
The takeaway
When you can't get the resources to prove the model at full scale, find a low-commitment real-world engagement with a trusted relationship to generate proof, then convert that proof into the bigger commitment.
Why it works
By first operating a services engagement for a friendly customer (no capital commitment required upfront), the firm built proof points and engineering credibility that later converted into a full ownership investment once trust was established.
One level up
Sequence the stages of trust-building (service → proof → ownership) rather than trying to skip straight to the end state.
Where it stops holding
This stops working when there's no trusted relationship willing to grant the initial low-commitment engagement, or when the pilot's learnings don't actually transfer to the scaled model.
Detail
situation: Sequence had capital and a handful of engineers but no deal; couldn't raise large capital without a deal, couldn't get a deal without proving capability.
decision: Took on a friend's family bank as a services customer (not yet an investment) to prove the model from August to November, then converted to a full ownership investment closed in March after regulatory approval.
outcome: The family asked to become a permanent partner; the regulated, centralized nature of the bank turned out to be ideal for agent deployment, validating the business model before scaling to Baldwin.
lesson: An unregulated or decentralized business would have been much harder to prove the model on; regulated, centralized operations with excellent data hygiene are a feature, not a bug, for early-stage AI agent deployment.
In their words
The holding company model is a weird business. Like without money, you can't do a deal. If you don't have a deal, you can't hire engineers. If you can't hire engineers, no one wants to give you money.
Bank South underwriting transformation results
Source
No Priors: Artificial Intelligence | Technology | Startups — Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee (September 24, 2026) · at 31:11 ▸
When this applies
when redeploying automation to reallocate skilled staff from repetitive processing to high-judgment exception handling
The takeaway
When automating a process, measure success not just by time-per-task but by whether the freed capacity can absorb a real surge in demand without quality degradation or headcount growth.
Why it works
Automating routine data-entry-like underwriting work freed staff capacity, which combined with a doubling of incoming loan volume (external, unrelated) to prove the system could absorb significantly more volume with the same or smaller headcount, without loosening underwriting standards.
One level up
The real test of an operational transformation is its performance under an unplanned stress test, not just its baseline efficiency metrics.
Where it stops holding
This stops working when demand doesn't grow to stress-test the freed capacity, so the gains remain theoretical rather than demonstrated, or when the quality/exception judgment can't be reliably automated at all.
Detail
situation: Community bank's underwriting headcount historically grew linearly with loan volume; staff spent most time on manual data entry rather than complex exception loans.
decision: Built a system to handle all consumer loan underwriting and roll out commercial loan underwriting automation, reallocating staff to focus on exceptions and front-office work.
outcome: Average consumer loan underwriting time decreased 94%; average commercial loan time dropped from 30 days to 11 days; when loan volume doubled in Q2 (unrelated to their work, per speaker), the bank absorbed it without loosening standards and with a smaller underwriting team (one retirement, one internal transfer).
lesson: Freeing staff from rote work and routing them to higher-value tasks lets an organization scale volume without proportional headcount growth, and the real proof came from an external demand shock the new system could absorb.
In their words
We functionally have taken the average loan at the bank, which used to take 30 days to go end to end to 11 days.