AI Model & System Changelog
A running log of all significant changes to AskBiz's AI model, system prompting, confidence scoring, and business intelligence methodology.
About This Changelog
We publish a log of all significant changes to AskBiz's AI system. 'Significant' means any change that could affect the answers you receive, the confidence levels shown, or the behaviour of Business Tools. Minor infrastructure changes and security patches are not logged here.
2026
April 2026 — Pulse v3.2 deployed
Updated Business Pulse scoring model. Changes: B2B customer health thresholds recalibrated (addresses known limitation for B2B businesses). Seasonal decomposition window extended from 90 to 180 days for businesses with sufficient history. No change to dimension weights.
March 2026 — Churn model v2.1
Churn prediction model updated. Changes: cohort comparison signal weight increased from 15% to 22%. Category breadth signal added. Backtesting accuracy improved from 81% to 89%.
February 2026 — Claude model update
Upgraded from Claude Sonnet (Feb 2025 version) to Claude Sonnet (Jan 2026 version). Benchmark results: calculation accuracy +0.4%, factual accuracy +1.2%, hallucination rate -0.3%. Recommendation accuracy unchanged.
January 2026 — Confidence scoring v2
Confidence scoring model updated. Changes: source directness weighting added as a fourth factor. Previously confidence was calculated from completeness, recency, and specificity only. This change primarily affects answers that use inferred data from secondary sources — these now more accurately receive Medium rather than High confidence.
2025
Q3 2025 — Financial domain fine-tuning layer
Deployed custom post-processing layer for financial calculations. Reduces currency conversion errors by 34%, date range interpretation errors by 28%. This layer sits between Claude's output and AskBiz's response delivery — it validates numerical outputs against your source data before delivering the answer.
Q2 2025 — Anomaly detection v2
Migrated anomaly detection from simple threshold model to seasonal decomposition model. False positive rate reduced by 61%. Alert accuracy improved significantly for businesses with strong seasonal patterns.
Q1 2025 — Launch
Initial deployment. Claude Sonnet (Dec 2024 version). Business Pulse v1.0. Basic anomaly detection. Churn model v1.0.