Back to Insights
AI & Bias10 min read5 June 2026

TARA in Practice: How Voice AI Transformed 250 Performance Conversations at a Rural Social Enterprise

When an organic cotton social enterprise in Andhra Pradesh deployed TARA for their field workforce, no one expected the findings. TARA flagged recency bias in 61% of sessions, affinity bias in 38%, and caught multilingual sentiment shifts that a human reviewer would have missed entirely.

When an organic cotton social enterprise in Andhra Pradesh deployed TARA for their field workforce review cycle, the initial goal was modest: reduce the administrative burden on HR and bring some structure to conversations that had previously happened over WhatsApp calls and informal site visits. What happened instead was a lesson in what AI-assisted fairness actually reveals when you apply it to a real, complex, multilingual organisation.

The enterprise manages 250+ employees across rural agency areas — field extension officers who work with smallholder farmers on organic cotton cultivation, quality coordinators who travel between ginning facilities, and support staff based across multiple locations in Andhra Pradesh and one entity in the Netherlands. Reviews had never been standardised. Managers rated from memory, in whatever language felt natural, with no structured prompts and no bias checks.

What TARA Found

Across 250+ TARA-facilitated review sessions: recency bias flagged in 61% of sessions, affinity bias in 38%, halo effect in 29%, and significant language-linked sentiment divergence in 44% of Telugu-English code-switching conversations.

The recency bias finding was not surprising — it is consistent with what we see across all deployments. What was striking was the affinity bias pattern: managers who shared a regional background or language with an employee rated them, on average, 0.4 rating points higher on qualitative competencies (not on measurable OKR outcomes, where the pattern disappears). In a 5-point scale, 0.4 points is not trivial. Over a career, it compounds.

The Language Problem Nobody Had Named

The most unexpected finding was in multilingual conversations. Field review conversations in this organisation often switched between Telugu and English mid-sentence — managers and employees both moving fluidly between languages depending on what felt more natural for a given concept. TARA's sentiment analysis caught something no human reviewer had flagged: when conversations switched from Telugu to English, the expressed sentiment about employee performance became measurably more positive, even when discussing the same events.

This is a known phenomenon in code-switching research — people tend to use their second language for more formal, professional framing, while their first language carries more emotional authenticity. In performance reviews, this means that an employee discussing a difficult situation in Telugu was being heard as more negative than an employee discussing the same situation in English. TARA flagged this divergence so HR could account for it before ratings were finalised.

The HR Response

The organisation's HR team used TARA's findings not to override manager ratings, but to have structured conversations with managers before ratings closed. In every case flagged for recency bias, the manager was shown the documented OKR record from Q1 and Q2. In most cases, the manager voluntarily adjusted their rating after seeing the full-year evidence. No rating was changed by the system — all changes were human decisions, made with better information.

TARA does not make performance decisions. It gives managers and HR leaders better information before those decisions close. The social enterprise's 91.1% core retention rate and 74.1% field retention rate — both above sector averages — are outcomes driven by people, supported by evidence.

See these principles in action

TalentSpotify applies evidence-led performance management, AI bias detection, and structured OKR cascades to your real workforce — in India and the GCC.

Book a Demo
TARA in Practice: How Voice AI Transformed 250 Performance Conversations at a Rural Social Enterprise — TalentSpotify Blog