FDE Foundations · Module 6: AI Decisions
Baselines Before Models
A baseline is the number every model must beat and the fallback when the model is down. Skipping it is how teams ship complexity they cannot justify.
10 min reading
Objectives
- Build the dumb baseline first and measure it
- Use the baseline as the accuracy bar and the fallback
- Decide with numbers whether the model earns its complexity
The dumb baseline
Before any model, measure the simplest workable approach: keyword matching, last value carried forward, the human process timed, majority class. It takes an afternoon and produces two things: an accuracy number and a fallback. If a clever approach cannot beat the baseline on your labeled sample, you just saved weeks.
Beat it honestly
Evaluate candidate and baseline on the same held-out sample with the same metric. State the metric that matters to the business: precision on flagged fraud (false alarms cost analyst hours), recall on missed invoices (misses cost money). A model that wins on accuracy but loses on the metric the business feels has not won.
The fallback is a deliverable
When the model is unavailable or below a confidence threshold, the system falls back to the baseline path and says so in the output. Confidence thresholds and fallbacks turn a research demo into a production system.
Keep the comparison alive
Re-run baseline against model quarterly with fresh samples. Drift is not only a model problem: sources change, and the baseline is your canary for whether the model still earns its keep.
Quick check
An optional 2-3 question self-check. Answers never leave your device, are not stored, and never count toward any assessment.
Exercise
Define the baseline for a fictional email-triage problem: the dumb approach, the metric, an imagined baseline score, and the rule for when an LLM approach replaces it.
Pass criteria
Dumb baseline named and measurable, one business metric chosen with a reason, a concrete threshold to beat, and an explicit fallback behavior.