Read, draft, approve, or act: choose the right autonomy level

Read is the safest starting point

Give a new workflow read-only access first: research, monitoring, categorization, and summaries. Nothing it produces leaves the dashboard, so the only thing at stake is your attention. This rung earns its keep by proving the agent understands the work before it is allowed to change anything — you see how it interprets your sources and your rules while the cost of being wrong is a discarded paragraph.

How to judge it: every claim in the output should point to a named source you can spot-check in seconds. If you cannot trace a statement back, treat the whole result as unverified.

Read still has a failure mode, and it is a quiet one: a fluent summary that is confidently wrong. An agent told to “read everything relevant” will also wander, surfacing noise and burying the two lines that mattered. Name the sources it may read. The trade-off is honest — read produces no finished work, so you are still the one doing the doing. That is the price of the safest rung, and it is worth paying until the agent has proven it reads your world correctly.

Draft separates speed from consequence

On the draft rung the agent prepares the real artifact — the email, the update, the record — and a person reviews it in final form before anything is sent. This is usually the best first step for customer communication and public content, because you get the speed of a ready-to-go draft without letting it leave your control. The work is done; the decision is still yours.

How to judge it: count material corrections per batch — facts fixed, tone rewritten, recipients changed. A steadily low count means the drafts are becoming trustworthy; a batch that needs heavy rewriting is telling you the instructions, not the reviewer, need work.

The classic failure is a mislabeled outbox: drafts saved somewhere that looks like “sent,” so a teammate assumes the work already went out — or worse, a gate wired so a “draft” quietly auto-sends. The other failure is human: when every draft has been fine, the reviewer starts approving without reading. The trade-off is that drafting does not remove your review step; it relocates it to the cheapest possible moment, before the words reach anyone.

Approve protects consequential actions

Use an approval gate for the actions that leave your control the instant they happen: sending, paying, publishing, deleting. The agent does all the preparation and pauses at the last responsible moment for a yes. The request itself has to carry its own context — the exact action, the target, the reason, and the consequence — so you can decide from a phone without opening another system. “Approve task?” is not a gate; it is a coin flip.

How to judge it: watch the rejection rate over a few weeks. A gate that is never once rejected is usually in the wrong place — it has become a rubber stamp, not a decision.

The failure here is approval fatigue: gate too many small steps and you train yourself to tap yes blind, which defeats the one gate that mattered. The fix is to batch — one approval for a reviewed set of thirty reminders, not thirty separate prompts. The trade-off is direct: every gate spends your attention, so place it only where saying no still changes something outside the dashboard.

Act is earned, never assumed

Automatic action is the top rung, and it fits a narrow shape of work: repeated, rule-clear, low cost per mistake, and reliably recoverable. This is where time savings actually compound, because the agent does and logs without waiting on you. But you promote a workflow to this rung — you do not start it here — and you promote after several clean drafts and a deliberate look at the edge cases, not after one good week.

How to judge it: promote the single narrow action that earned trust, not the whole workflow around it. Keep a review-after log so “act” never means “invisible,” and set a clear trigger — a spike in corrections — that drops the action back to the approve rung.

The failure that hurts is promoting everything at once on the strength of a calm month, and then an edge case runs silently: a refund to the wrong customer, a message to a churned account. Recovery has to exist before you grant autonomy, not be improvised after. The trade-off is visibility for speed — you trade watching each step for logging it — which is a fair trade only where the log lets you catch and reverse the rare miss.

Match the level to the mistake

When you are unsure which rung a workflow deserves, score two things separately: how bad a mistake would be, and how hard it is to undo. Collapsing them into a single “risk” number is where teams go wrong — it hides the cheap-but-irreversible action (a one-line email you can never unsend) and the expensive-looking action you can reverse in a click. Rate consequence and reversibility on their own, then read the table.

The table is a starting point, not a verdict. Start one rung lower than your eventual goal and let evidence move you up: a workflow you hope to automate begins at draft, proves itself, and earns the promotion. The right level for the same action can also change as the stakes change — a $5 credit and a $5,000 refund do not belong on the same rung, even inside one workflow.