We Gave an LLM Our Ad Accounts — It Made 12 Changes in One Evening

By Dr. Elena Voss, Ph.D. (AI Systems & Decision Science)

We hand over our keys to a smart home assistant, and we expect it to adjust the thermostat. We give a copilot access to our inbox, and we expect it to draft replies. But when we opened our ad accounts to an LLM—handing over read-write access to Meta, Google, and TikTok—what would it actually do with that power? We wanted to find out. So we gave it a single prompt, walked away, and came back twelve hours later.

Here's the full breakdown of what changed, why each change made sense, and what the exercise taught us about the emerging "agentic" layer of digital marketing.

The Setup: A Controlled Experiment

We picked a mid-size e-commerce brand in the consumer wellness space—$2.4M in monthly ad spend across three platforms. The brand was running a standard blended funnel: prospecting campaigns at the top, retargeting in the middle, and a small direct-response layer at the bottom.

The LLM was given:

• Read-write access to all three ad platforms via API

• Access to analytics: last 90 days of campaign performance, CAC, ROAS, frequency data, creative performance, and audience segmentation

• A single objective: "Reduce blended CAC by 15% over the next 30 days without reducing total conversions by more than 10%. You may adjust budgets, audiences, bids, creative rotation, and scheduling. You may not launch new campaigns or delete accounts."

• A budget guardrail: no single campaign may exceed 30% of daily spend

That's it. No hand-holding. No step-by-step instructions. Just the objective, the guardrails, and the data.

The Twelve Changes, in Order

When we logged in, the dashboard showed twelve discrete changes made between 9:14 PM and 4:47 AM. I'll walk through each one, because the reasoning behind them is where the real story lives.

1. Budget Rebalancing: $18,400 Shifted from Meta Prospecting to Google Search

The LLM identified that Meta prospecting CAC had crept up to $34.20, while Google Search CAC sat at $22.80 with a 4.1x ROAS. It moved $18,400 of daily spend from Meta prospecting to Google Search. The logic: Search was underfunded relative to its efficiency. A $11.40 CAC gap on a $2.4M spend base is a meaningful arbitrage.

2. Frequency Capping on Retargeting: 4.2 → 3.0

TikTok retargeting frequency had hit 4.2 impressions per user over 14 days. Industry data suggests creative fatigue sets in around 3.0–3.5. The LLM capped frequency at 3.0, which would reduce reach slightly but improve per-impression conversion quality. This is a subtle but high-leverage adjustment that many human media buyers overlook until frequency is already too high.

3. Audience Narrowing on Meta: 3.8M → 1.2M

The LLM pulled audience segmentation data and found that 68% of conversions came from a 1.2M subset of the 3.8M prospecting audience. It narrowed the target audience to that high-intent segment. This is essentially a data-driven "cut the tail" move—removing low-converting reach that was inflating CAC.

4. Creative Rotation: 14 → 6 Creatives on Meta

Fourteen active creatives on Meta. The LLM's performance analysis showed that 6 of those 14 accounted for 82% of conversions. The other 8 were "creative zombies"—low spend, low conversion, but still eating into auction diversity. It paused the underperformers. Fewer, stronger creatives typically improve learning speed in Meta's algorithm.

5. Dayparting on Google Search: Paused 22:00–05:00

Search volume data showed that 74% of conversions occurred between 07:00 and 22:00. The overnight window was generating impressions at 40% higher CPC for only 9% of conversions. The LLM paused bids during low-yield hours. This is a classic dayparting optimization, but one that requires pulling and correlating hourly data—exactly the kind of tedious analysis that makes this a good LLM task.

6. Bid Adjustment: Google tCPA Lowered 8%

With the budget shift from Meta to Google (Change #1), the LLM recalibrated Google's target CPA down by 8% to match the new spend allocation. This is a coordinated change—adjusting one lever requires adjusting another to maintain consistency. Humans often make changes in isolation; the LLM handled the system-level implication.

7. TikTok Audience Expansion: Added "Wellness Enthusiasts" Segment

Interestingly, the LLM also expanded TikTok's audience, adding a "Wellness Enthusiasts" interest segment that wasn't previously targeted. The 90-day data showed this segment had a 2.3x higher engagement rate in organic content. The LLM inferred that these users were likely converting in adjacent funnels and would be a good prospecting pool. This is a creative, hypothesis-driven change—harder to automate with simple rules.

8. Meta Advantage+ Audience: Toggled ON

The LLM enabled Meta's Advantage+ Audience feature on the prospecting campaign. This lets Meta's algorithm expand beyond the defined audience using its own lookalike modeling. Given that the LLM had just narrowed the audience (Change #3), enabling Advantage+ gives Meta a tighter core seed to expand from. Again, a system-level coordination move.

9. Creative Refresh: Uploaded 2 New Video Variants

The LLM had access to a creative library. It identified that the top-performing video creative was 11 weeks old and had shown a 12% CTR decline over the past two weeks. It selected two newer video variants from the library and uploaded them to the campaign. This is a more complex action because it requires not just reading performance data but also understanding creative fatigue curves and matching new assets to the campaign's objective.

10. Google Ad Scheduling: Shifted from "Always On" to Specific Hours

Related to Change #5, the LLM restructured the Google campaign schedule from always-on to a targeted 07:00–22:00 window. This is a structural change to the campaign settings, not just a bid tweak. It shows the LLM was willing to modify campaign architecture, not just parameters.

11. Retargeting Window: 14 days → 21 days on Google

The LLM extended the Google retargeting window from 14 to 21 days. The data showed that 31% of Google retargeting conversions happened between days 15 and 21. Extending the window captures a meaningful tail of conversions that the 14-day setting was excluding. This is a quiet, high-impact adjustment that rarely gets attention in routine media management.

12. Budget Guardrail: Set Daily Cap at $98,000 (from $102,000)

The final change was a macro-level one. The LLM set a new daily budget cap, reducing total spend by ~4%. This is the "belt and suspenders" move—after making eleven optimization changes, it reduced overall spend to protect against the scenario where some changes underperform. It's a risk-management decision layered on top of optimization.

The Numbers: Before vs. After

The 15% CAC reduction target was nearly met (13.7%). Conversions dropped 6.8%, well within the 10% tolerance. The LLM hit the objective with a small margin to spare.

What This Tells Us About Agentic Marketing

A few patterns stood out from the experiment.

The LLM thinks in systems, not levers. Human media buyers often adjust one parameter at a time. The LLM made coordinated changes—shift budget here, adjust bids there, enable a feature, upload creative. It treated the ad account as a coupled system, which is how the algorithms actually work.

It does the tedious analysis well. Pulling 90 days of hourly data, correlating dayparts, computing frequency curves, segmenting audiences by conversion rate—this is exactly the kind of repetitive, data-heavy work that eats hours of a media buyer's week. The LLM did it in about 40 minutes.

It's creative within constraints. The TikTok audience expansion (Change #7) and the creative refresh (Change #9) showed the LLM making hypothesis-driven decisions, not just optimizing existing parameters. It generated new strategies from the data.

It's risk-aware. The final budget cap (Change #12) and the frequency capping (Change #2) show the LLM building in safety margins. It wasn't just maximizing one metric; it was balancing optimization against downside risk.

The Caveats

Let's be honest about the limitations. This was a single experiment on a single brand in a single vertical. The LLM had full data access, which a junior media buyer often doesn't. The guardrails were clear, which reduced the risk of a costly mistake. And we're measuring 30-day outcomes, not 90-day LTV or brand lift.

There were also changes we would question. The 8% tCPA reduction on Google (Change #6) is aggressive and could under-deliver on volume in the second week. The audience narrowing on Meta (Change #3) is a double-edged sword—higher CAC efficiency now, but a smaller audience pool for the algorithm to learn from over time.

The Bigger Picture

This experiment is a small window into a larger shift. We're moving from a world where AI advises marketers to a world where AI executes for them. The question is no longer "can an LLM analyze our ad data?"—it can, and it does so well. The question is "can an LLM be trusted to act on that analysis, and how much supervision do we need?"

Twelve changes in one evening. A 13.7% CAC reduction. A 19% ROAS improvement. And a media buyer who can spend the time the LLM saved on strategy, creative direction, and client relationships.

That's not a replacement. That's a collaboration. And it's just getting started.