Why do some AI projects increase revenue while others don't?

Because the technology matters less than where you point it. The size of the gap between what your business currently does and what AI can deliver determines whether you see a return.

The Same Technology, Wildly Different Results

A 2026 study from Columbia Business School and Peking University tested this directly. One company deployed AI across seven different business workflows using the same technical team and the same models. Results ranged from a 16% increase in sales to a 4.5% decrease. Same company. Same AI. Same engineers. The only variable was which workflow received the upgrade.

The pattern was clear. Where the business had a significant capability gap, AI filled it and revenue went up. Where the existing process was already competent, AI added nothing or made things worse.

Where It Worked

The biggest gain came from a customer service chatbot. Before AI, customers asking questions before buying got an automated message saying nobody was available. That was the standard practice because the company couldn't afford multilingual human agents for those inquiries. AI filled a complete void: 24/7 availability, every language, context specific answers. Sales increased 16% for customers who interacted with it.

Search improvement added 3% to sales. Product descriptions added 2%. Both meaningful at scale, both filling a real gap (language barriers and missing product information on a cross-border platform).

Where It Failed

Google ad title optimization went negative. The existing human titles were already decent, and the AI model wasn't tuned for advertising. It produced generic titles that Google's algorithm actually ranked lower. The result was fewer views, fewer clicks, and lower sales.

The lesson is that deploying AI where a competent process already exists, without domain specific tuning, can actively hurt performance.

The Mechanism Behind the Gains

Across every workflow that produced positive results, the gains came from more people buying, not from people spending more per purchase. Conversion rates increased between 1% and 22% depending on the workflow. Cart values didn't change at all.

AI removed friction that was preventing purchases. It didn't upsell, it didn't manipulate, it didn't change buying behavior. It answered questions that weren't getting answered, translated queries that weren't being understood, and described products that had no descriptions.

Return rates and customer satisfaction stayed the same or improved, which rules out the possibility that AI was tricking people into purchases they'd regret.

How to Apply This

Before choosing where to deploy AI in your business, map your capability gaps. For each customer facing workflow, ask:

1. What happens right now when a customer needs help or information at this stage?

2. Is the current solution constrained by cost, language, availability, or volume?

3. If you could provide unlimited, instant, personalized support at this stage, would it change the outcome?

If the answer to the third question is yes and the current solution is weak, that's your highest value deployment. If the current process is already performing well, AI might not improve it and could introduce new problems if not properly tuned for the domain.

The research showed that a hybrid approach (AI first, human escalation when needed) outperformed both AI alone and humans alone by 11.5%. Full replacement isn't always the goal. Sometimes the highest value play is using AI to extend what your team can already do.