Until very recently, running an AI model on your ad account without supervision would have been genuinely reckless. One developer just published an account of doing exactly that for 30 days with Claude Code.
The experiment used Claude Code, Anthropic's AI coding assistant that can write and execute code, read files, and take actions inside real computing environments, to autonomously manage advertising campaigns. Rather than asking Claude to draft ad copy for a human to review and publish, the developer gave it execution access to actually run campaigns without sign-off on each decision.
This is meaningfully different from using AI as a writing aid. It is AI as an operator: reading performance data, making budget decisions, creating or pausing ad variations, and iterating through all of it without a human approving each step.
What Autonomous Ad Management Actually Requires
For Claude Code to run ads end-to-end, it needs access to ad platform APIs (the technical interfaces that let software connect to services like Google Ads or Meta Ads Manager), the ability to read analytics data, and logic for acting on what it finds. In practice, this means writing and executing code that queries performance metrics and takes action based on the results.
Claude Code is designed for exactly this kind of agentic workflow, where each step depends on the output of the previous one. Advertising fits the pattern neatly: check performance, write a new ad variation if click-through rate is low, pause what is underperforming, shift budget toward what is working.
The harder question is reliability. Autonomous agents are prone to specific failure modes: taking irreversible actions they should not, getting stuck in loops, misreading data, or making decisions that are locally sensible but globally wrong. Running ads unsupervised amplifies all of these risks, since a misread budget figure or a flawed decision rule can mean real money spent on campaigns that are not working.
The 30-day timeframe is what makes this experiment credible. Short tests, a few hours or a single day, do not surface the slow-drift failures that emerge when an agent runs continuously. A month is long enough to reveal whether the system holds its course over time or gradually veers into problematic territory.
For marketers and small business owners watching the AI agent space, real-world tests like this are more useful than benchmark scores. Benchmarks show what a model can do under ideal conditions. A month of autonomous ad management shows whether it holds up under real data, real variation, and real consequences for getting things wrong.
Claude Code is available as part of Anthropic's Claude Pro and Claude Max subscriptions, with API access for developers building their own agent workflows.