Some of the most useful AI systems I've built for marketing don't create anything.
They watch.
They track when an ad starts losing steam, which competitor messages continue running and what keeps stopping buyers from moving forward. Then they tell the team what changed and what needs a decision.
This changed how I think about AI in marketing. Producing more is useful. Noticing the right thing earlier can save the budget you were about to waste.
Signals arrive daily. Attention doesn't.
Monitoring is still manual in most marketing teams.
Someone checks the ad account when performance drops far enough to become obvious. Competitor activity gets reviewed when a campaign is being planned. Sales objections reach marketing through a meeting, if they reach them at all.
The information exists. But nobody is watching it consistently enough to catch the change early.
3 AI early-warning systems for marketing
Each of these systems does one job: it checks a signal regularly and brings you the few things that need attention. You can connect the source to Claude, ChatGPT, Codex or another AI assistant through MCP or an official connector. The official MCP guide is a useful starting point.
Keep the connections read-only at first. Let the AI find the issue and recommend the next move. A person should still decide what changes.
1. Find creative fatigue before the weekly review
Most teams find creative fatigue after the numbers have already dropped. This system checks every active ad against its own recent performance and flags the change earlier.
What you need: Connect your ad account through an ads MCP or the platform's official API – Meta Marketing API, Google Ads API or LinkedIn Marketing APIs. You also need a short rules file covering the metrics and thresholds that matter in your account. If you want alerts in Slack, Teams or email, connect that channel to the same AI assistant.
The system pulls active ads, compares a recent window with a longer baseline and returns only the creatives that crossed your fatigue rules.
For example, you might ask it to flag an ad when CTR has fallen for three days, CPA has moved beyond your acceptable range and the ad has enough spend for the change to mean something.
The baseline matters. A 15% fall may be normal for one campaign and serious for another. Compare each creative with its own history before comparing it with the rest of the account.
Build it for yourself
Create a simple fatigue-rules.md file with the primary metric for each campaign type, the minimum spend or impressions required before judging an ad, the comparison windows and the point at which you want an alert. Run the system manually for the first week and compare its flags with the decisions you would have made yourself.
Once the rules are working, turn it into a routine or cron job in Claude Code or Codex. Daily is enough for most accounts. For high-spend campaigns, run it every 6–12 hours.
Use this prompt:
Review all active ads in the connected ad account for signs of creative fatigue.
Use fatigue-rules.md as the decision criteria. Compare each creative's recent performance with its own longer-term baseline before comparing it with other ads.
For every creative:
1. Pull only the metrics available from the connected platform.
2. Check that it has crossed the minimum spend or impression threshold.
3. Compare the recent window with the baseline window.
4. Mark it as Healthy, Watch or Action needed.
Return only Watch and Action needed items in a table with:
- Campaign and ad name
- Metrics that changed
- Recent value versus baseline
- Why it was flagged
- Recommended next move
Do not pause ads, change budgets or publish anything. If the source does not provide a required field, say so. Do not estimate missing data.
If nothing needs attention, return: No creative-fatigue action needed today.
2. Track what competitors are changing
Competitor ad libraries are useful, but checking them manually soon becomes another task that gets skipped. This system checks the same set of competitors, records what is live and shows you what changed since the last run.
Use the official public libraries: Meta Ad Library, LinkedIn Ad Library and Google Ads Transparency Center.
What you need: A competitor list, browser access inside your AI assistant and a folder where it can maintain an Excel file. If browser automation is not available, ask Claude Code or Codex to set it up for you.
The Excel sheet can capture the useful details from every ad and keep them ready for comparison.
On Meta, you can reliably collect the advertiser name and page, library ID, active status, start date, visible active time, primary text, headline or description when present, CTA, media type, video duration when present, destination URL and visible creative versions.
That gives you enough to compare changes in copy, formats and CTAs across every check. The history becomes more useful each time the system runs.
Build it for yourself
Ask the AI to create competitor-ad-log.xlsx, with one row per ad. Keep the source URL and the time checked in every row. On later runs, it should add new ads, update ads it has seen before and mark ads that are no longer returned by the active-ad search as No longer visible. That wording is deliberate – absence from a search does not prove when or why an ad stopped.
Run it two or three times a week. Daily makes sense during a major launch or when you are watching a fast-moving category.
Use this prompt:
Check the connected official ad libraries for every company in competitors.md.
Use the advertiser or page record where the library supports it. Do not rely on broad keyword matches when a direct advertiser record is available.
Create or update competitor-ad-log.xlsx with one row per ad and these columns:
- checked_at
- market_checked
- competitor
- advertiser_name
- advertiser_page
- ad_library
- library_id
- status
- started_running
- visible_active_time
- primary_text
- headline
- description
- CTA
- media_type
- media_duration
- destination_domain
- destination_url
- multiple_versions
- visible_version_count
- source_url
Use only information shown by the official library. Leave a cell blank when the source does not provide the field. Never infer spend, targeting, conversions or performance.
Compare this run with the existing workbook. Add new ads, update matching ads and mark previously active ads that no longer appear as No longer visible. Do not mark them Stopped or Ended unless the source confirms that status.
After updating the workbook, give me a short note covering:
1. New ads since the last check
2. Changes in message, offer, format or CTA
3. Ads with multiple visible versions
4. Ads that are no longer visible
Separate what the source shows from your interpretation. Do not call a long-running ad a winner or make performance claims from its duration alone.
3. Review an asset through buyer perspectives
This one was inspired by Ole Lehmann's buy-or-bounce panel, where simulated buyer archetypes review an asset before it goes live. The basic idea is useful. The problem is that generic archetypes usually produce generic feedback because the AI has to invent too much about the buyer.
I changed one part: each buyer gets a separate persona file built from real customer context. The same personas then review every landing page, email, ad or sales asset. This makes the feedback more consistent and stops the panel from turning into five versions of the same imaginary buyer.
What you need: The asset you want reviewed and accurate buyer-persona files. These can sit in Google Drive, Docs or a local project folder connected to your AI assistant. If you want the review delivered through Slack or email, connect that channel too.
Each persona file should cover the person's role, priorities, buying trigger, concerns, proof expectations and how they would judge this type of purchase. Keep each file focused on one buyer. The AI can infer the offer and positioning from the asset it is reviewing.
Build it for yourself
Start with three to five personas that genuinely take part in the buying decision. Save each one as a separate Markdown file in a personas folder. Then save the prompt below as a reusable skill or command in Claude Code, Codex or Cursor.
Run this when an asset is ready for review, before the final edit. It does not need a daily schedule.
Use this prompt:
Review the supplied marketing or sales asset using every buyer-persona file in the personas folder.
Treat each persona as an independent reviewer. Use only the information in that persona's file and the asset itself. Do not invent objections, company facts, customer evidence or buying criteria.
For each persona, return:
- Verdict: Buy, Continue evaluating or Bounce
- What they understood the offer to be
- The strongest reason they would continue
- The main point that would make them hesitate or leave
- The exact line or section that caused that reaction
- The one change most likely to improve the asset for this persona
After the individual reviews, add a short combined view:
1. Issues raised by more than one persona
2. Conflicting needs between personas
3. Missing proof or information
4. The three changes to make before publishing, in priority order
Do not rewrite the asset unless I ask. Do not smooth over disagreement between personas. If a persona file does not contain enough information to make a judgment, mark the verdict Insufficient context and explain what is missing.
What an early-warning system needs
Connecting the source is usually straightforward. The useful work is deciding what deserves attention and what the system should do when it finds something.
Each of the three systems above needs five things:
- A reliable source. The system should read directly from the ad account, public ad library or asset folder. Keep access read-only until you trust the output.
- The right comparison. Creative performance needs its own history. Competitor ads need the previous workbook. An asset review needs accurate buyer-persona files. A signal only becomes useful when the system has something relevant to judge it against.
- Clear decision rules. Write down what should trigger a flag, how much data is enough and which changes can be ignored. Otherwise the system will keep sending you summaries instead of decisions.
- A recommended next move. Every flag should show what changed, the evidence behind it and what the team could do next. The recommendation can be rejected, but it gives the reviewer a useful starting point.
- A human gate. Let the AI check, compare and recommend. Keep budget changes, campaign edits and publishing decisions with a person.
The quality of the system depends more on these five inputs than on the AI model you use. If the output is noisy, tighten the comparison and decision rules before changing the tool.
Improve the judgement over time
Save every flag and add what happened after a person reviewed it. A simple log can capture the issue, the recommendation, the decision taken and the reason for that decision.
Also record issues a person spotted that the system missed. Otherwise you will only reduce false alerts without finding its blind spots.
Review the log once a month. If the same weak flag keeps appearing, change the rule. If useful signals are being missed, add the missing condition or context. If a buyer persona no longer reflects how the customer buys, update the persona file with verified information.
Keep these changes human-approved. The system should not rewrite its own rules or persona files based on its previous output.
The saved history becomes useful beyond individual alerts. It can show which creative types tend to weaken first, how competitor messages change over time and which issues keep appearing across different assets. That gives you better inputs for the next campaign, not just a faster warning on the current one.
These three systems are a starting point, not the whole map — the right signals to watch depend on your accounts, your competitors and your buyers. If you want to design that around your business, let's talk — reach out to AIxGrowth.