What to automate first — designing your automation frontier

“What can we automate with AI?” is the wrong question — because almost everything eventually qualifies. The right question is “what first, and how far right now?” — and the answer isn’t a fixed list but a boundary that keeps moving.

Five criteria for the first workflow

Agent-deployment practice has converged on five traits of a good first candidate: high volume (rare work never pays back its automation cost), clear rules, available data, measurable savings, and low downside risk. Start where all five hold, and the first win becomes trust capital for the next automation.

But avoid the pilot trap

Don’t read “low downside” as “unimportant.” In MIT’s GenAI Divide research, the 5% that captured value were the organizations that integrated into high-value core workflows — not peripheral pilots. The sweet spot is work that is low-downside and attached to a core process: draft generation, classification, and verification-prep stages of the work that actually matters.

The boundary moves every few months

There’s a reason you can’t set the automation list once and walk away. METR’s measurements show that the length of tasks (in human-professional time) that frontier AI agents can complete at 50% reliability has been doubling roughly every 7 months for six years — with follow-up analysis suggesting the doubling has accelerated to around 4 months since 2023. Much of what agents “can’t do” today isn’t can’t — it’s can’t yet. If you’re drawing this year’s boundary with last year’s judgment, the boundary is probably already wrong.

Manage the frontier as a living document

So the form we recommend is not a list but a living document. Split workflows into three columns — automated, semi-automated (human approval gate), still human. Then re-examine the “still human” column quarterly: is each item there because of a capability limit, a trust limit, or a regulatory requirement? Capability limits get re-tested in a few months; trust limits move to the gated column behind an approval gate; regulatory requirements stay. Where to place the gates is covered in designing approval gates.

How we apply this

Notique’s frontier document is public in How we run. Today, code writing, deployment, and monitoring sit in the automated column; releases and publishing sit at gates; contracts and regulatory accountability remain human. That placement fits our scale and risk profile — it isn’t the answer. But the method — three columns, re-examined on a cycle — applies at any scale.