Candles Not Lightbulbs
Summary: Ship a useful but imperfect analytic solution now rather than delaying for a polished one — a poor light source still beats darkness, and can be upgraded once the customer can act.
Sources: Handwritten.md; LinkedIn Posts.md (“Candles, not lightbulbs,” posted 2026-03-03) (original content, authored by Paul)
Last updated: 2026-07-07
Amid a wide-ranging conversation about ethics, security, cost, and scaling in analytics and AI, Paul anchors on a single priority: all the data, algorithms, and workflow integration ultimately exist to help a human make a better decision. He frames insight/inference as illumination — if the customer is sitting in the dark, the job of analytics is to bring light to the choice (source: Handwritten.md).
The anxiety that shows up around this — “what if we get it wrong,” “did we examine all the corner cases,” “will it scale,” “how do we sustain trust when the error rate isn’t zero” — is where the phrase Candles Not Lightbulbs comes from: the job is to illuminate, and there’s no good reason to make a customer navigating in darkness wait for the (better, but not-yet-ready) lightbulb if a candle is available now (source: Handwritten.md).
Candles are admittedly a poor light source with real drawbacks, but the difference between operating in shadow and having any light source is profound. Once a customer can make informed decisions, the analytics can be upgraded over time (source: Handwritten.md). The tradeoff is explicitly between accuracy and speed — getting to a better answer sooner catalyzes learning, both for the model and, more importantly, for the operational implications on users. Jump-starting that learning curve accelerates the path to an integrated, trusted tool, provided it’s coupled with a genuine commitment to continued evolution (source: Handwritten.md).
The LinkedIn version grounds the principle in Paul’s own insurance-industry vocabulary — naming an underwriter, claims handler, clinician, or risk officer as the “human operator” navigating in the dark, and reframing the test not as “is this the perfect model?” but “is this enough light to make today’s decision meaningfully better than yesterday’s?” It also names where genuine defensibility comes from: early deployment under “candlelight” surfaces edge cases, failure modes, workflow friction, and trust thresholds that “no amount of whiteboarding will fully anticipate” (source: LinkedIn Posts.md).