How to choose the first process to put AI to work

Almost every company has a list of ideas for AI. The hard part is choosing the first one, and the first one matters more than it seems: if it goes well, it opens the door to the next; if it stalls halfway, it closes that door for years. The research of the last two years can help you choose better.

The jagged frontier

In 2023, a team from Harvard, Wharton and MIT ran one of the most cited field experiments on AI at work. It involved 758 BCG consultants, around 7% of the firm’s consulting workforce, split between those working with AI and those working without it (Dell’Acqua et al., 2023).

On tasks within AI’s reach, the results were striking:

With AI, the consultants completed 12.2% more tasks, finished 25.1% faster and delivered results of 40% higher quality.

But the researchers included one task deliberately designed to fall outside AI’s reach. There, those working without AI got it right 84% of the time. With AI, the rate of correct answers dropped to between 60% and 70%. The authors called this the “jagged frontier”: AI is excellent at some tasks and confidently wrong at others that look similar.

Who gains most

A second study, published in 2025 in the Quarterly Journal of Economics, followed more than five thousand customer support agents at a real company. With an AI assistant, they resolved on average 15% more cases per hour. Among the least experienced staff, the gain reached 30%. Among the most experienced, there were small gains in speed and small losses in quality (Brynjolfsson, Li and Raymond, 2025).

The reading matters: AI passes on to less experienced people the know-how that the best performers already have. It works best where the work is repetitive, what a good result looks like is well understood, and the volume is high.

Four criteria for your first process

These studies yield practical criteria:

1. Inside the frontier. Tasks where you already know what a good result looks like: reading, classifying, extracting, summarising, comparing, drafting. To begin with, avoid tasks that call for ambiguous judgement.
2. Volume and repetition. A 15% gain per hour only shows if there are plenty of hours. A process that runs every week pays back visibly and quickly.
3. A verifiable result. If an error is easy to spot, the jagged frontier stops being a risk and becomes a list of exceptions.
4. An owner on the inside. Someone who feels the pain and will stand up for the tool when it needs adjusting.

A test before you decide

Before you commit to a process, test it. Pick twenty real cases, some easy and some hard, and see where AI gets it right and where it fails. It is the cheapest way to find out which side of the frontier your problem sits on before you invest in it.