How to Leverage Domain Expertise — Chris Lovejoy, Notius Labs

Chris Lovejoy, founder of Notius Labs and former medical doctor, presents a framework for building what he calls a "domain native AI organization" — arguing that winning in vertical AI is fundamentally an organizational challenge, not a model sophistication challenge. Drawing on his background as a physician at Cambridge and the NHS, and experience as the first employee at Anterior (a prior authorization AI startup), Lovejoy contends that the system for incorporating domain expertise into AI products is more important than the underlying models. He cites Gartner data showing roughly 50% of generative AI projects were abandoned in the prior year, attributing this failure rate largely to insufficient embedding of domain judgment.

The core of the talk is a three-archetype framework for how domain experts can be embedded in an AI organization:

1. The Oracle — the domain expert directly assesses AI outputs and makes improvements themselves (e.g., tweaking prompts, adding documents, tools). Both assessment and improvement happen in the same person's hands. Lovejoy uses Granola, the AI meeting notes startup recently valued above $1 billion, as the primary case study: their first employee Joe, a writer and journalist, wrote all the original prompts and still serves as the primary gatekeeper of AI quality. Lovejoy argues this works because there is no objectively perfect meeting note — quality requires human taste rather than a measurable metric — and because the output volume at Granola's scale remains amenable to direct human review.

2. The Evaluator — the domain expert defines quality metrics and builds measurement systems, then hands improvement work to engineers. This archetype suits situations where quality can be quantified. Tandem, described as the largest clinical AI product provider in the UK by adoption, illustrates a scaled version: they started with a single clinician-oracle (Roy, a doctor-turned-McKinsey consultant), but as specialties and geographies multiplied, moved to a "decentralized oracle" model hiring doctors across specialties and countries to own prompt variations for their respective sub-segments.

3. The Architect — the domain expert designs self-improving systems that learn from usage with minimal ongoing human-in-the-loop involvement. Lovejoy's own arc at Anterior illustrates all three stages sequentially: he began as the clinical oracle reviewing prior authorization AI decisions, then built a clinician review dashboard and defined failure modes (evaluator), and finally designed automated improvement mechanisms (architect) to handle the large variation in how different insurance organizations interpret their own policies.

Lovejoy provides a decision heuristic: if quality cannot be measured in metrics, default to the oracle archetype. If it can be measured and manual iteration is fast enough, an evaluator suffices. If manual iteration is too slow and automation can be designed, pursue the architect pattern. He notes organizations commonly start as oracles and evolve, and that evolution requires forethought in hiring.

On skills, Lovejoy distinguishes the oracle (needs direct-use-case expertise, prompting instinct, taste, customer empathy), the evaluator (adds data science intuition, statistical skills, industry connections for building review teams, product management to collaborate with engineers), and the architect (adds LLM product experience and engineering skills to implement automated improvement systems).

Three organizational principles close the talk: define a single principal domain expert who is accountable for AI quality rather than distributing responsibility across a committee; give that person genuine ownership and include them in decision rooms rather than treating them as a consultant; and hire for breadth of adjacent skills beyond raw domain knowledge, so the person can grow from oracle to evaluator or architect as the product scales. A cited failure mode is hiring two senior clinicians in co-advisory roles with no clear principal, which produced slow iteration and both eventually leaving after 12–18 months.