What to hand to an agent, and what to keep

AI is excellent at some messages and quietly bad at others; decide which is which before you hand anything over.

The jagged frontier

In a field experiment with 758 consultants at Boston Consulting Group, researchers gave people realistic tasks with and without AI. On tasks inside what they called AI's "jagged frontier", consultants using AI completed 12.2% more tasks, 25.1% more quickly, at more than 40% higher quality. On a task just outside it, w…

Usually safe to hand off

Scheduling, confirmations and routine logistics.Summaries of a long thread.First drafts of an ordinary reply, which you then edit.Translations of a message you can sanity-check.

Keep for yourself, or review closely

Bad news, apologies, condolences.Anything involving money, commitments or legal exposure.A first message to someone who matters.Anything new, where nobody has done it before and there's nothing to copy.

Start simple, add autonomy slowly

Anthropic's guidance on building agents recommends the simplest approach that works, adding complexity only when it's needed, and letting agents pause for a person at checkpoints. For messaging that means: let the agent draft and you send. If you later let it send on its own, give it the narrowest permission that work…

The short version

AI's abilities are jagged: great on some tasks, worse than you on similar-looking others.Hand off routine, low-stakes messages and first drafts.Keep bad news, money, commitments and first impressions for yourself.Default to draft-and-review; add autonomy narrowly and with approval for risky sends.

Try it in Kanvo

In Kanvo the agent drafts and you press Send. If you want an agent to send on its own, grant it a mandate (Mandates → Grant a mandate) limited to the apps and daily volume it needs; risky drafts wait under Approvals.

Sources

With 758 BCG consultants: on tasks inside AI's frontier, those using AI completed 12.2% more tasks, 25.1% more quickly, at more than 40% higher quality; on a task outside it, they were 19 percentage points less likely to be correct.