Where AI earns leverage

Not every task wants to be an agent. The rough heuristic: an AI agent earns its keep when the task is repetitive, language-heavy, and tolerant of a small percentage of errors that a human can catch downstream. If any one of those three is missing, the math gets worse fast.

Three dimensions to score

• Frequency — does this happen many times a week, or once a quarter? Compounding wins are weekly or daily.
• Language content — is the task mostly about reading, writing, classifying, or summarizing language? If yes, the model is doing the work it's best at.
• Error tolerance — what happens when the agent is wrong? If the cost is "a teammate fixes it in 30 seconds," you're fine. If the cost is "we lose the customer," you're not — at least not yet, not without heavy guardrails.

The two task shapes to favor

Triage — sorting inbound things into the right bucket and acting only on the simple cases. Tickets, leads, expense reports, candidate resumes. The agent handles the easy 60-80%; humans handle the rest. Wins because the easy cases used to consume real time.

Draft-to-review — producing a first draft of something a human would have written from scratch, then letting that human edit. Status updates, customer emails, code reviews, briefs. Wins because going from blank page to draft is the expensive step.

The shapes to avoid in week one

• Anything where "mostly right" is unacceptable — billing, compliance, medical, legal of-record decisions
• Tasks where the input is unstructured data your team doesn't already capture (you'd need a whole pipeline first)
• Tasks that take you 5 minutes once a month — the build cost outweighs the win
• Tasks where the bottleneck isn't language, it's some other system — slow approvals, missing data, broken integrations