In January 2024, publisher Macmillan rolled out an internal AI assistant with balloons, doughnuts, and a name-the-bot contest. Leadership was thrilled. The staff — the people actually supposed to use it — called it terrible. Usage flatlined. By October 2025 the tool was quietly retired, and the follow-up evaluations of its replacement were reportedly described as “useless” too.

This is the most common AI failure in business, and it has almost nothing to do with the technology. It’s a rollout failure wearing a launch party.

Where It Went Wrong

• Launch first, usefulness later. The celebration happened before a single employee had proven the tool on real work.
• Built for the announcement, not the job. Staff needed help with their actual tasks; they got a mascot with a search box.
• No honest feedback loop. When the people using it daily say it’s failing, that’s data. It was treated as resistance.

The Rollout That Actually Works

1. Pilot quietly with real staff. Five to ten people, real tasks, no balloons. If it can’t survive contact with a Tuesday, it can’t survive a launch.
2. Measure use, not vibes. Log-ins, completed tasks, time saved. If usage decays after week two, the tool is the problem — fix or kill it fast.
3. Let the users name the wins. The best internal AI stories spread peer-to-peer: “this saved me an hour.” No doughnut competes with that.
4. Scale what they invent. Your team’s workarounds and prompts are the real product. Formalize those, then expand.

The Bottom Line

Macmillan didn’t buy a bad tool; they threw a party for an unproven one, and the party made it harder to admit the truth. Every business is one launch event away from the same mistake. Pilot quietly, measure honestly, and let the balloons wait until the staff asks for them.

Stardust Fallout rolls AI into client workflows the boring way that works — small pilots, real tasks, measured results. Want that instead of a launch party? Book a consultation.

Quick Questions

Why do internal AI tools fail so often?
Usually rollout, not technology: tools get launched top-down before staff prove them on real work, and nobody measures actual usage afterward.

What’s the right size for an AI pilot?
Small — a handful of real users on real tasks for a few weeks. Expand only what they demonstrably keep using.

How do I know if our AI tool is working?
Sustained voluntary usage and completed tasks. If you have to mandate it or promote it constantly, that’s your answer.