Macmillan's NOA Chatbot Flops: Lessons for Enterprise AI Infrastructure Buyers

In January 2024, Macmillan Publishers rolled out NOA — short for "No Ordinary Assistant" — with balloons, doughnuts, and light-up marquee letters at its Lower Manhattan headquarters. The chatbot, built by German vendor Chaptr (backed by Macmillan's parent Holtzbrinck), was pitched as a closed-loop enterprise tool that could safely ingest proprietary manuscripts and internal data without third-party leakage.

Staff quickly discovered the gap between launch fanfare and daily utility. Multiple employees described NOA as prone to hallucination, tone-deaf in its adjustable personas ("sarcastic," "regretful," "pessimistic"), and ultimately unused. One editor called it "terrible." By October 2025, Macmillan sunset the product and shifted toward Google Gemini — only to find Gemini's evaluations of agent submission letters "useless," according to an editor.

Why it matters for GPU / AI infrastructure

The NOA episode underscores a recurring pattern: enterprises procure or build LLM-powered tools without first validating model quality, retrieval accuracy, and workflow integration on representative workloads. For infrastructure buyers, this means GPU capacity planning must be coupled with rigorous offline evaluation — latency, throughput, and cost per useful token — before committing cluster spend. A closed-loop deployment on dedicated hardware still fails if the underlying model cannot meet domain-specific quality bars.

Macmillan's pivot to a general-purpose API (Gemini) also illustrates the hidden cost of vendor lock-in versus self-hosted control. Teams running their own model-serving stack on owned or leased GPUs retain data sovereignty and can iterate on fine-tunes, guardrails, and evaluation sets — critical for publishing workflows where factual fidelity and author trust are non-negotiable.