Human-in-the-Loop PPC: Approval Without Bottleneck

"A human approves every AI action" is the sentence that makes stakeholders relax and workflows die. Uniform approval fails in both directions at once: trivial actions drown the approver until clicking yes becomes a reflex, and material actions wait in the same queue behind them. The fix is not more approval or less — it is placing approval where judgement actually earns something, and letting the agent run at full speed everywhere else. Tier the actions by blast radius, and the loop gets both safer and faster.

The two failure modes: rubber-stamping and bottlenecking

Watch a uniform-approval workflow for a month and you will see both failures, usually in sequence.

Rubber-stamping arrives first. When the queue mixes "read last week's search terms" with "raise this budget", the approver's attention is priced by the average item, not the important one. By week three the yes-click is muscle memory — which means the one approval that mattered got the same half-second as the forty that did not. A control everyone bypasses on autopilot is worse than no control, because it also manufactures false confidence.

Bottlenecking is the mirror image. If the approver is diligent, the queue backs up; the agent that could have fixed a leak on Tuesday waits for Thursday's review block. The team quietly learns the AI is slow — when what is slow is a process that made a human read forty reads.

Both failures share one root: the workflow priced every action the same. Blast radius is the correct price. You can diagnose which failure you have from the log alone: approval times collapsing toward zero means rubber-stamping; proposals ageing unactioned for days means bottleneck. Either reading is the trigger to re-tier, not to abandon oversight.

Tiering changes by blast radius, not by tool name

The question for any action is not what is it called but what does it cost if wrong, and can we get back. Three tiers cover a Google Ads agent's entire action space:

Tier	Actions	Human involvement	Why this level
1 — log only	Reads, reports, drafts, paused additions, negatives	Weekly log review	Wrong costs nothing or is trivially removable
2 — confirm first	Budgets, bids and targets, pause/enable, campaigns going live	One click per action, evidence attached	Wrong costs money — and can be reverted
3 — never automated	Deletion, structural surgery, billing and access	Human act in the interface, if ever	Wrong cannot be undone

Tier 1: what runs on a log, not a prompt

Everything whose worst case is "remove it later" belongs here: every read, every report, drafted RSAs, keywords and campaigns that arrive paused, and negative keywords — whose blast radius is excluded traffic you were probably paying to waste. The governing artefact is the log, not the prompt: each action recorded with tool, account and timestamp, reviewed in batch.

The counterintuitive entry is negatives. They do change serving, but their mistake mode — excluding a good query — is caught by the same weekly review and reversed by deleting one row. Teams that push negatives into tier 2 usually stop adding them at the pace waste demands, which is the more expensive error.

Tier 2: the confirmation that earns its click

Spend-affecting, reversible changes get the approval prompt: budget moves, bid strategy and target changes, pausing or enabling live entities, and the moment a drafted campaign switches on. Two properties keep this tier fast. Evidence rides with the request — the agent proposes "raise budget on Brand — Search, here is the pacing math", not "trust me" — so judging takes seconds. And volume stays low because tier 1 never enters this queue. In practice that is a handful of decisions on an ordinary day, each backed by analysis you can skim. The client-side mechanics — how the prompt surfaces the exact tool call — are covered in the approval workflow tutorial.

Tier 3: what never gets automated

Anything without an undo. Deletion is the canonical member, and the architecture should enforce it rather than promise it — a system that cannot delete beats a system instructed not to. In AdCopilot's case the four remove tools are simply never exposed, and REMOVED-status mutations are refused server-side. Teams extend tier 3 to taste: billing, access grants, cross-account structural moves. The tier's definition is worth writing down for your own account, because it is also the honest answer to "what is the worst this system can do?" — and "a reversible mistake" is a very good worst case.

Writing your own tier policy: borderline calls, decided once

The table above is a default, not a law, and the value of a tier policy comes from deciding your borderlines once instead of re-litigating them per action. The classic borderline cases, with the reasoning that usually settles them:

• Geo and schedule changes — reversible and rarely urgent; most teams put them in tier 2 because they reshape delivery in ways that surprise stakeholders.
• Negative keywords — tier 1 by default, but a brand-side negative (blocking your own brand term account-wide) has tier-2 consequences; scope the exception, not the whole class.
• Budget changes below a threshold — some teams tier by size: small moves within a campaign's normal range run tier 1 with the log, larger moves confirm first. Honest thresholds come from your account's history, not a round number.
• New campaigns — the draft is tier 1; the enable is tier 2. Splitting the action at the moment spend becomes possible is the cleanest line in the whole policy.

Write the policy down — one page, three lists, the thresholds named. The document is not bureaucracy; it is what lets a second teammate, or a future you, approve consistently at speed.

The weekly log review: the habit that replaces anxiety

Tiering only holds if tier 1 is genuinely reviewed — otherwise "log only" decays into "unwatched". Fifteen minutes, weekly, same agenda:

1. Volumes — actions per tier this week; a tier-2 spike or a tier-1 flood is a process signal, not just activity.
2. Refusals — anything the server declined, and why. Refusals are the system's immune response; read them.
3. Spot-check five tier-1 actions at random against the account. Trust the log by verifying the log.
4. Regrets — any approved change you would not approve again. That is your calibration data, and the input to reconciling both ledgers.

The deeper effect of the ritual is on the humans. Anxiety about AI in an account is mostly unbounded imagination; a log turns it into a short list of facts reviewed on Fridays. After a clean month, most teams do the opposite of what they feared they would: they move actions down a tier — deliberately, with evidence — because the record showed the judgement was already sound.

That is the quiet payoff of tiered oversight: it is not a brake on autonomy, it is the instrument that tells you when more autonomy is safe. Start a free pilot and run the first week with the tiers exactly as written here.