Air Canada

Information Architecture Is Essential for Reliable AI

The article argues that decades of underfunding information architecture (IA) are now causing visible, repeatable, and costly AI failures — hallucinations, wrong answers, and faulty agent actions — because retrieval and agent systems rely on structured, typed content. It explains that metadata, taxonomies, controlled vocabularies, and a governed semantic layer make documents findable and distinguish authoritative from anecdotal sources. The author recommends auditing what models retrieve before buying another model, typing content prior to indexing, fixing labels and taxonomies before prompts, and giving agents explicit relationship and permission models. The piece cites industry studies and examples (surveys, legal rulings, and vendor research) to show IA’s direct impact on retrieval accuracy, hallucination rates, and business risk.

PM's Guide to Managing AI Debt

This guide reframes AI debt as "options debt": the loss of a product team's ability to respond when AI systems fail in production. It presents a practical control-room metaphor with three debt categories—foundation, drift, and operations—each monitored by a gauge (green/yellow/red) and defended by a corresponding lever (version & replay; shadow & refresh; guardrail & stabilize). The core rule is: never scale when any gauge is red or unknown. The article gives concrete thresholds and runbook guidance (example replay pass-rates, drift/resolution thresholds, latency and cost cutoffs), recommended trial sizes (500–1,000+ conversations and 1–2+ week slices), and low-budget alternatives (spreadsheets for policy versioning, manual shadowing, sampled replay). It cites IBM Security breach cost findings and legal and industry examples (Air Canada, Klarna) to underscore operational and reputational risk.

LLM-Powered Email Support Triage with Agent Mailbox

The article demonstrates how to build an LLM-driven email support triage system by giving the agent its own hosted mailbox using Nylas Agent Accounts (beta). Inbound mail is classified into four buckets (URGENT, ACTION, FYI, NOISE) using a short prompt and limited context (sender, subject, 200-character snippet) with temperature=0 for high accuracy. Classification is inexpensive (example token and cost math for GPT-4o-mini and GPT-4o are provided). Draft generation is gated by knowledge-base confidence thresholds and a risk tiering policy; high-risk topics escalate to humans. Mail-layer rules (spam blocking, folder routing, VIP marking) prevent unnecessary model inputs. Drafts are saved to the agent’s drafts folder and never auto-sent. The post includes rollout advice (start small, batch similar tickets, log everything) and operational constraints like send caps.