upladsHelp CenterLogin

How to run the AI Optimizer and approve its proposals

Last updated July 19, 2026

The AI Optimizer is the autonomous mode of the Optimize tab: a senior-marketer persona works through your ad account in six phases, thinking out loud as it goes, and records every concrete action it wants to take as a proposal you approve afterwards.

Starting a run

Open Ads Manager → Optimize and pick the AI Optimizer tab (or click Start AI Optimizer in the Optimize header). The start screen shows the six phases, a model picker listing every model for the AI providers you've configured, and a green badge that states the deal plainly: Read-only analysis. Nothing changes without your GO. Click Start AI Optimizer.

Requirements are the same as for the Assistant: an OpenAI or Anthropic API key in Configuration, and a Growth plan or above (The AI Optimizer is available on the Growth plan and above.). The run writes in German when your browser locale is German, otherwise in English.

The six phases

Before phase 1 the server prefetches a live snapshot — 30-day and 7-day overviews, ad sets, the last 14 days of account activity and the stored audit — shown as a Prefetching account snapshot chip. If that stored audit is missing or older than 7 days, a fresh audit runs first (Running a fresh audit), so the run isn't blind to the change journal and prior outcomes.

  1. Account understanding — what kind of account this is: objectives, conversion mode, tracking confidence, structure, budgets, recent human changes. Ends with a focus plan.
  2. Performance pulse — daily series and 7-versus-7 comparisons hunting CPM inflation, CTR decay, frequency creep, conversion gaps. Diagnosis only; proposals just for emergencies.
  3. Structure & setup — fragmentation versus consolidation, learning-phase math, budget distribution, bid strategies, placements, naming hygiene.
  4. Ads & creatives — spend concentration per campaign, underfed winners, clear losers, fatigue, video hook and hold, copy quality on the top spenders.
  5. Budget & scaling — spend share versus result share, where money is misallocated, which winners have headroom, the concrete budget moves.
  6. Action plan — the deliverable: executive summary, a priority-ordered table of every proposal, non-automatable advisory recommendations, and a 7-day watchlist.

Each phase has its own step ceiling (6, 8, 8, 8, 8 and 4 model turns). Hitting it ends that phase gracefully and moves on — it doesn't fail the run.

Watching it work

The centre column streams the model's reasoning in real time under a phase stepper, with tool chips for each read. A footer shows tool reads, proposal count, elapsed time and an estimated cost for the run. Use Jump to latest if you've scrolled away.

Controls sit in the run header:

  • Pause stops the run after the current phase (Paused after phase n/6); Resume picks it back up.
  • Stop aborts the in-flight request — it's a pause, not a delete.
  • If a phase errors you get Retry phase; after a clean stop you get Resume.

Resumability is derived from the run's own message log, so a crash, a closed tab or a provider blip resumes at exactly the right phase rather than skipping or repeating one. Finished runs are listed in the left rail with their proposal count (executed/total) and can be reopened read-only; Run again starts a fresh one.

One run at a time

Only one optimizer run per ad account can be in progress. A second start returns An optimizer run is already in progress. A run that stops being touched for 2 minutes is treated as crashed and stops blocking new starts. When resuming a phase that was interrupted, uplads requires the run to have been quiet for 15 seconds — until then you'll see Run is busy — try again in a few seconds. (a failed phase releases that lease immediately, so Retry phase works right away).

Proposals: nothing runs without your GO

The run cannot change your account. When it identifies an action, the write call is validated, snapshotted and parked as a card in the Proposed actions tray, and the model is told to carry on. Execution only ever happens later, through the same confirmation path the Assistant uses.

Each card shows a severity dot (critical, high, medium, low), a category icon, the entity name, a one-sentence reason, before → after rows, a confidence letter (high, medium, low) and an expandable Evidence section with the KPIs and the expected impact. Clicking a proposal marker in the stream jumps to its card.

Card actions and statuses:

  • GO executes that one proposal. It moves pending → executing → executed (or failed with the error).
  • GO all (n) appears when two or more are pending; it asks Execute n actions? first, then runs them in severity order.
  • Dismiss marks the card Dismissed and nothing runs.
  • Undo appears on executed cards for the cleanly reversible single-entity actions only — status, budget and bidding changes where a prior value was captured. Ad-text rewrites, applied audit findings and bulk actions can't be undone.

With read-only access to the ad account, GO and Dismiss are disabled and the run records no proposals at all — it writes the recommendations out in prose instead.

Guardrails on what it may propose

The proposals are deliberately conservative:

  • Scaling winners moves budgets +20–80%, never more than 2× in one step.
  • Cuts to underperformers are −15–30%; a pause is reserved for what is clearly dead against the account's own benchmark.
  • One budget or status proposal per entity per run — no stacking two moves on the same ad set.
  • A 3-day cooldown: an entity a human changed in the last three days gets no budget or status proposal, only a note on what to watch.
  • Total-budget ceiling: the sum of all budget proposals shouldn't lift the account's total daily budget by more than about 25% without saying so explicitly.
  • Restructures, creative briefs, tracking and landing-page fixes never become proposals — they land in the advisory section of the action plan.

The workspace monthly AI spend cap is re-checked before every model call, so a long run can't overshoot it; if it's reached, the run stops with Monthly AI limit reached.

Related

Related articles