An HR agent helps screen resumes for technical roles. The team trains a recommendation model on historical hiring data — resumes that "made it through" are positive examples. After a quarter in production, the agent's recommendations are 80% male candidates, despite the applicant pool being 45% female. Legal is alarmed.
What's the right architectural response?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
HR bias from agentic systems is one of the most legally and ethically consequential failure modes in business AI. Historical hiring data encodes past discrimination — training on it propagates that discrimination forward as a "neutral" prediction. The fix is multi-layered: audit the data, measure the distribution outcomes, set fairness constraints, add human review gates. Prompt instructions can't do this work — bias is in the data and the model weights, not in what you ask the model to avoid. This is one of the highest-risk applications of agentic AI in business; treat it accordingly.
"Don't be biased" tells the model what not to be — it doesn't change the data the model learned from or the patterns it's been trained to recognize. Bias mitigation requires changes at the data and architecture levels; prompts are downstream of the actual problem.
"It's reflecting historical reality" treats discriminatory historical hiring as ground truth. It's not — it's an artifact of past discrimination. Building a system that automates the perpetuation of that artifact is exactly the harm fair-AI work exists to prevent. And it's legally untenable in jurisdictions with employment-discrimination law.
Smaller models can be more biased on certain dimensions, not less — they pick up cruder statistical patterns more readily. Capability size isn't the right axis for bias mitigation.