Macmillan killed its workplace AI assistant 21 months after a launch party

WIRED reports that Macmillan Publishers held a party in January 2024 for a new workplace AI assistant called NOA. The outlet lists purple balloons, light-up marquee letters, and a plate of doughnuts spelling "NOA IS HERE". The vendor is Chaptr, a German software company "backed" by Holtzbrinck, Macmillan's parent. WIRED reports that Macmillan staff learned via a company-wide email in October 2025 that Chaptr would be sunsetting the tool.

The piece is behind a paywall. This item carries three facts from the reporting. NOA was launched at Macmillan with a party. The vendor was Chaptr. NOA was sunset about 21 months later. The outlet reports much more, including authors, other Big Five publishers and named concerns about AI in publishing. A reader who wants the full account should open WIRED directly.

The pattern this story names

A reader of this item is in or near a decision to buy a workplace AI assistant. The Macmillan case has one shape many of these projects share. A vendor sells a general-purpose assistant. The buyer runs a launch event. Training sessions warn that the model can hallucinate. Staff are then encouraged to use it for the exact work it was warned about. Months later the tool is quietly turned off. The money is not refunded. The time is not refunded. The next vendor arrives with a very similar pitch.

The failure mode here is not the vendor's. It is the lack of a measured decision. A party is a company commitment. A kill decision made in email is the admission that the first commitment never had a success test.

Three steps a buyer can take

First, name a kill date at the launch. A pilot with no end date drifts. Nobody at the vendor wants to be the one who says "stop". Nobody at the buyer does either. So nobody does. A date, written on the launch slide, forces a review. The review decides whether the tool graduates or whether the licence is cancelled.

Second, measure use, not sentiment, in the first quarter. Weekly active users of the specific features the vendor pitched. Not survey scores. If the vendor pitched email drafting and manuscript review, those are the two counts to track. A dashboard inside the tool, read by the buyer every Monday. Three months of flat use is the kill signal, regardless of the balloons.

Third, keep the tool scoped to the task the vendor can prove. "Draft emails" and "evaluate a manuscript" are two different jobs. One rewards early AI, because the drafts are short, structured, and a human reviews them. The other does not, because the model's output is a decision about work the firm's whole business depends on. Buying one tool for both tasks because they fit in one chat box is where the cost shows up. The cost shows up as hallucinated marketing copy.

Why this will keep happening

The Macmillan pattern is not special. Workplace AI assistants are sold on a general chat demo. They are bought on a company-wide seat count. Nobody holds the seat count against a measured task. The buyer's quarterly licence renewal is on autopay.

The one thing WIRED's piece shows is the gap between a launch event and a decision. Any buyer about to run a launch event this quarter can borrow the second half of the Macmillan story. Set the kill date. Pick the two tasks the vendor must beat. Then throw the party.

Source

Macmillan Staff Say the Publisher Threw a Party for a Workplace AI Assistant That Was Later Killed, WIRED. Paywalled.