#207 — Building vertical AI & AI-native services

Mirage PMF is the illusion of product-market fit caused by rapid revenue growth and high retention that is powered by hidden human labor rather than true software automation. You can diagnose it if gross margins remain flat or decline (below 50%) as revenue grows, if headcount scales 1:1 with customer onboarding, or if ARR per FTE matches legacy service benchmarks instead of expanding software multiples.

Unlike traditional SaaS, vertical AI and AI-native service businesses must classify all foundation model inference costs, API tokens, fine-tuning compute, and human-in-the-loop review labor directly inside Cost of Goods Sold (COGS) rather than hiding them in R&D or OpEx. Accurate gross margins typically start at 40-50% during early human-guided workflows and should expand past 70-80% as machine autonomy increases.

HURT (Human Review Time) is a north-star product metric that tracks the average minutes of human labor required to review and verify an AI-generated output without degrading accuracy or quality. Pioneer companies like Crosby Legal track HURT on every document processed; as HURT approaches zero, operating margins mathematically converge toward high-margin software economics.

Targeting smaller or mid-market customers enforces productization discipline through workflow homogeneity and lower ACVs, making it impossible to throw unscalable manual labor at delivery. For example, insurance AI startup Harper targeted Main Street businesses (daycares, restaurants) over Fortune 500 accounts, allowing them to standardize document schemas, minimize edge cases, and scale to 5,000+ businesses in 13 months.

Outcome pricing charges directly per discrete deliverable (e.g., $150 per processed insurance claim or $500 per audit binder) rather than per seat or per hour. For ongoing, variable-volume work, use normalized workload credits that map to task complexity, decoupling your top-line revenue from labor hours while capturing 10-20% of the customer's total operational labor savings.

Ensure client MSAs and engagement letters grant your platform explicit contractual rights to retain and train models on anonymized operational data, human corrections, and workflow edge cases. In companies like Harper, every customer interaction and underwriter feedback loop compounds into proprietary matching graphs and domain models that competitors cannot replicate.

Instead of relying on conceptual slide decks, live-demo your internal operator workbench and autonomous agent pipelines directly in front of buyers. Enterprise legacy modernization startup Mechanical Orchard cut sales cycles by over 50% by demoing their internal 'Cursor for COBOL' code-refactoring engine live on discovery calls, proving real automation capability to skeptical buyers.

Never acquire a legacy firm early in your lifecycle, as it introduces manual operational bloat and threatens an AI-first engineering culture. M&A should only occur after your core AI platform achieves proven workflow leverage and high gross margins; at that stage, acquisitions serve as an inorganic customer acquisition and distribution accelerant rather than a crutch for missing technology.

The 80-customer rule states that founders should avoid rejecting early custom enterprise requests until reaching roughly 80 live clients. As Crosby Legal CEO Ryan Daniels observed, before 80 clients you cannot reliably distinguish genuine platform rules from one-off exceptions; past that threshold, recurring workflow patterns emerge, allowing you to confidently say no to bespoke custom work that does not productize.

Staff initial customer onboarding with a specialized 'SEAL team' adept at extreme operational ambiguity and hands-on migration hurdles, rather than rolling general project teams onto pilots. Once implementation stabilizes and data schemas are validated, the SEAL team hands off the account to steady-state delivery operators focused on long-term automation integration.

Partner with mid-sized legacy service providers that have executive buy-in and urgency, rather than slow-moving market giants (e.g., Prosper AI partnering with Firstsource for revenue cycle management). Crucially, startups must maintain direct contractual ownership of the end customer to ensure their proprietary training data flywheels and client relationships are never disintermediated.

Automating modular tasks rather than entire people makes product adoption palatable to risk-averse enterprise buyers and dramatically simplifies machine evaluation benchmarks. As Strala CEO Timon Gregg noted, focusing on specific task units—like extraction, reconciliation, or draft generation—enables process-oriented engineers to ship reliable, high-precision evals far faster than attempting end-to-end role replacement.

Instead of running slow batch evals every few weeks, sit domain practitioners (such as attorneys or underwriters) directly beside software engineers in real-time pairing sessions. At Crosby Legal, lawyers provide granular feedback every few hours, allowing engineers to tweak system prompts and context orchestration in real time, drastically accelerating model performance improvements.

Startups can build recurring SaaS-like stickiness by becoming the foundational system of record or infrastructure layer for the client. For example, AI-native fund administrator Hanover Park built an internal ERP ledger through which all client financial data runs, creating multi-year contract stickiness and prohibitive switching costs.