I Surrendered My Work to AI for 7 Days—The Results Left Me Speechless

When researchers and companies put AI to the test at work, the results were a mixed bag. Some teams moved faster. Others worked harder without getting more done. And a few found that AI created more problems than it solved.

On the positive side, ChatGPT cut the time workers spent on writing tasks by 40%. Output quality went up by 18%. Weaker-skilled workers gained the most from using AI. And after just two weeks of use, real-job usage doubled among people who’d been exposed to AI tools.

ChatGPT slashed writing time by 40%, boosted output quality by 18%, and helped lower-skilled workers the most.

But quality wasn’t always reliable. Human experts rated AI outputs on a scale of 1 to 9. AI scored a 7 or higher — considered minimally sufficient — in 65% of about 11,000 tasks. However, it never hit a top score of 9 in more than half of its attempts. Researchers project that by 2029, AI could reach 80 to 95% minimally sufficient results.

In marketing and experimentation, AI agents handled nearly 59% of Optimizely’s experimentation usage. Teams ran 78.7% more experiments and launched 24.1% more personalization campaigns. Win rates improved by 9.3%.

Still, not everything ran smoothly. A team using top AI models — including GPT-5, Claude Sonnet 3.7, and Gemini 2.5 Pro — spent 30 hours over two weeks running experiments. The AIs designed good experiments but executed them poorly. One major error: recruiting 39 participants but forgetting to include an experimental condition.

A Berkeley study found that AI caused three types of work intensification. Workers lost downtime, took on wider job scopes, and juggled more tasks at once. An eight-month study across different fields confirmed that people were working harder, not just smarter. The study also found that AI reduced task friction, leading workers to engage in work tasks even during personal breaks like lunch.

The ROI picture was also troubling. A report highlighted that 95% of organizations saw no measurable return on their AI investments. High activity didn’t mean high value.

On the brighter end, narrow AI tools helped some workers free up focus time. Human-AI collaboration produced strong results in some cases. Seven experiments over three years showed that AI could amplify human capabilities — when used well. MIT researchers analyzed 41 different LLMs, including Claude, Gemini, and ChatGPT, to evaluate whether AI could produce acceptable outputs across thousands of tasks without human edits.