upladsHelp CenterLogin

How to use the Optimize Assistant chat

Last updated July 19, 2026

The Optimize Assistant is a chat that can actually read your Meta ad account. It pulls live numbers with its own tools, answers in plain language, and — when you ask for a change — hands you a confirmation card instead of touching anything on its own.

Opening the Assistant

Go to Ads Manager, open the Optimize tab, then switch to the Assistant tab (next to Overview and AI Optimizer). The chat always works against the ad account selected in the Optimize header.

Two things must be in place:

  • An AI provider key. The Assistant runs on your own OpenAI or Anthropic key. Without one you'll see a Connect an AI provider to use the Assistant card with a link to Configuration — see How to configure AI provider keys.
  • A Growth plan or higher. The Assistant is part of the AI Optimizer feature set. On a lower plan you'll get The AI Optimizer is available on the Growth plan and above.

What you can ask

Four starter prompts sit on the empty state — Where am I wasting budget?, Which ads should I pause?, Scale my top performers, Why did my CPA go up? — but you can ask anything about the account. Behind the scenes the Assistant can read the account overview, ad sets, ads, breakdowns, entity details, ad copy, top ads, the stored optimization audit, the account activity log and daily time series.

It judges each entity against its real optimization goal (including custom conversions), states the time window it analysed, and answers in the language you write in. A single message can be up to 8,000 characters.

Picking a model and reading the cost

The header has a model picker listing every model for the providers you've configured. You can only change it before the first message — once a chat has content the picker locks with the hint Start a new chat to change the model. Start a New chat to switch models.

Next to it, a small figure shows the estimated cost of the current chat. It's restored when you reopen a saved chat, so it reflects the whole conversation, not just this session.

Conversations list

The left sidebar lists the saved chats for the selected ad account, newest first, up to 50 conversations. Click one to reload it — the full transcript, the tool chips, the confirmation cards and their outcomes all come back. Hover a row and click the trash icon to delete it. AI Optimizer runs never appear here; they have their own history rail in the AI Optimizer tab.

How tool calls are shown

While the Assistant works, each read it performs appears as a one-line chip with a wrench icon and a plain label — spinning while it runs, a green check when it succeeds, an amber warning if it fails. You don't have to do anything with them; they're there so you can see which data the answer is built on.

Approving changes

Every account change is gated. When the Assistant decides to act, it doesn't execute — it parks the call and shows an Action card with:

  • the human-readable entity name and what will change, as before → after rows,
  • a text preview when it's an ad-copy rewrite,
  • Confirm and Cancel buttons.

Nothing happens until you click Confirm. The message box is disabled while a card is pending (Confirm the pending action first…), so a decision can't get lost. Cancel marks the action Cancelled and the Assistant carries on from there. Related same-type changes are bundled into one card rather than one card per entity.

The server owns the conversation state: your browser only sends a message or an approve/reject decision, never the proposed arguments. A change can't be altered between what you see on the card and what runs.

Undoing a change

After a card shows ✓ Done, an Undo button appears — but only for cleanly reversible single-entity actions where uplads captured the previous value first:

  • Undoable: status changes (pause/activate), budget changes, bidding changes.
  • Not undoable: ad-text rewrites, applying an optimization audit finding, and the bulk variants (bulk status, bulk budget, bulk bidding).

Undo writes the snapshotted prior value straight back to Meta and marks the card ↩ Undone. It's recorded in the audit log like any other change — see How to view the audit log of workspace changes.

Limits and errors you might hit

  • Tool-step limit. One turn may run at most 12 model↔tool round-trips. Beyond that you'll see Reached the tool-step limit. Ask a more specific question. — narrow the question or split it into two.
  • Monthly AI spend cap. The workspace cap is re-checked before every model call, so a long turn can't overshoot it. When it's reached you'll get Monthly AI limit reached with your spend and cap; raise it in Configuration or wait for the new month. See How to set an AI spend cap.
  • Read-only access. If your team permissions don't allow editing this ad account, write tools aren't offered at all — you get the analysis and exact recommendations to apply yourself.
  • Stop. The square button next to the composer aborts the in-flight response.

Related

Related articles