Chat with AdWitch AI

The AI Agent at /chat is a conversational media buyer for AdWitch. Describe the outcome you need; it can combine the platform’s campaign, creative, analytics, account, automation, reporting, memory and integration operations.

How to Use

1. Open AdWitch AI at the top of the sidebar — the app opens on it.
2. Start a new conversation or continue one from history.
3. State the goal, account, period and constraints in ordinary language.
4. Answer any compact form the agent presents for missing information.
5. Review confirmation cards before actions that publish, activate, delete, spend money or enable automation.

Every model call is metered against the team balance. Live-ad tasks also require a connected Meta ad account; the agent will surface missing requirements rather than pretending the action succeeded.

Useful Requests

• “Show my campaigns and explain which ones need attention.”
• “Create a draft campaign for this offer with a $50 daily budget.”
• “Import my existing Meta campaigns.”
• “Review performance for the last 30 days and name the best and worst movers.”
• “Analyse my active creatives for fatigue.”
• “Propose a budget reallocation plan across the portfolio.”
• “Search the Meta Ad Library for active meal-delivery ads in Germany.”
• “Set up Autopilot with my target CPA and daily action budget.”
• “Save this approved fact about our audience to team knowledge.”

Create creatives in chat

Ask the agent for a Meta creative at /chat, or press Generate via agent on /creatives. It studies the client profile, goal, placements, current ads and winning creatives, then asks only for missing details. It proposes 2–4 concepts in a card: each shows the format and placements, generation-model chain, required materials, estimated price and time. Attach images, videos or MP3/WAV audio (up to 50 MB) in the composer; use photos of real people only with their consent.

The agent states the price before paid runs. A chain of several runs or a run above the platform’s configured amount requires an explicit yes; included generations in your plan are used first, and the agent tells you how many remain. Chat shows queued and in-progress states, galleries for batches, playable videos with sound, and the model and mode used for each result. If a run fails, the failure shows the refunded amount. Ask for variations, edits, extensions or placement sizes (9:16, 1:1, 4:5, 16:9). Everything is saved in the creative library.

It can make text banners, product shots, lifestyle scenes, series with a consistent character (training requires consent), UGC-style videos with voice, multi-shot stories, first/last-frame transitions, edits and extensions of your own footage, motion transfer, object swaps and cinematic spots. Do not request fake testimonials; AI people must follow Meta’s rules. See Creative Library for the full workflow.

What the Agent Can Operate

The agent chooses registered operations itself. Representative capabilities include:

• Campaigns: list and inspect campaigns; create and edit campaign, ad-set and ad drafts; import from Meta; publish after confirmation; pause or activate; clone or launch across accounts.
• Creatives: list assets, generate images, create videos and carousels, upload to Meta and request placement previews.
• Analytics: account and campaign metrics, live Meta insights, period reviews, audience research, creative analysis, delivery diagnosis, change history and budget planning.
• Ad accounts: check and switch the active Meta ad account, inspect account status, pixels, Pages and spending limits.
• Automation: configure Autopilot and manage tests or Meta-native rules through chat.
• Memory: recall team knowledge, save a fact with consent, manage structured client profiles and scan a website.
• Research and reports: search the public Meta Ad Library, search the web for current outside information with cited sources (see Web Search in Chat) and generate reports.

When troubleshooting, the agent can also read this documentation. You normally do not need to know or type internal tool names. Creative generation is built in; it does not require connecting a separate generation service.

Models and thinking

The agent uses Claude models for chat. Routine turns may start on Sonnet and escalate to Opus for more complex or consequential work; the turn receipt shows the model used. While it works, the activity timeline can show a Thinking… row with its progress and elapsed time.

Conversations and Control

• Conversations are saved in history and can be renamed or deleted.
• Separate conversations keep work for different clients or topics apart.
• Editing or retrying a message creates branches; branch controls let you move between alternatives.
• While the agent is responding, use Stop to interrupt the stream.
• Forms and confirmation cards keep structured or consequential actions explicit.

Voice and Attachments

Use the microphone in the composer to record a request. Stop the recording to transcribe it, review the text, then send it. You can discard a recording or cancel transcription. Images and videos can also be attached to a message where relevant.

First Conversation

Try: “Check which Meta ad account is active, show my campaigns for the last seven days, and recommend the next action.” Then ask a follow-up using the returned campaign name. If you request a draft, provide the objective, audience, budget, schedule, placements and creative—or let the agent ask for what is missing.

Tips

1. Specify account, currency, dates and target metrics.
2. Distinguish between asking for a plan and asking the agent to apply it.
3. Keep durable business facts in Agent memory (/knowledge); client profiles live on its Profiles tab (/knowledge/profiles).
4. Check every preview and confirmation before publishing.