Your ad account holds more data than you will ever look at. What it does not have is an analyst.
Every Monday someone exports a report from Meta Ads Manager, exports another from Google Ads, pastes both into a spreadsheet, colours a few cells and sends it on. Two hours of work that repeats identically every week. Next week, the same.
That is work an agent can take over. Since April 2026 it genuinely can, because Google and Meta finally opened the door.
What Claude Code is, if you are not a developer
Claude Code is Anthropic’s agent. It runs in a terminal and has access to your files. It was written for programmers. Marketing picked it up for a simple reason: it is not a chatbot that hands you text, it is a tool that opens the file, runs the script, calls the API and gives you the result.
The difference is in what you get back. ChatGPT tells you how you would calculate ROAS by creative. Claude Code performs that calculation on your data and returns the numbers.
If the terminal puts you off, Anthropic shipped Cowork in January 2026, the same agent in a desktop interface with no command line. It is built on the same framework and aimed squarely at teams like marketing and operations. Everything below applies to both.
What changed in April 2026
Until last year, connecting an AI agent to an ad account was a developer project. Then both platforms released official MCP servers, interfaces through which an agent talks to the ad system directly instead of clicking through a UI.
Google Ads MCP shipped on 28 April 2026, open source under Apache 2.0, maintained by the Google Ads API team. It is deliberately read-only. It runs GAQL queries against your account, returns data, and changes nothing. It will not pause a campaign or move a budget. You host it yourself with your own credentials.
Meta Ads MCP became available to developers in July 2026. This one reads and writes: it creates, edits and deletes campaigns, ad sets and ads, manages custom audiences and catalogues, and runs A/B tests. Authentication goes through Facebook Login for Business. Agencies managing other businesses’ accounts have to submit their app for review to get Advanced Access.
That asymmetry matters. Google gave the agent eyes. Meta gave it hands. What follows is advice on handling the hands carefully.
Five things worth building in week one
Start with reporting. The work is repeatable and a mistake costs nobody anything.
A weekly report that writes itself. A script pulls the last seven days from both platforms, calculates ROAS, CPA and conversion rate by campaign, compares against the previous week and writes a summary. Not an “AI report”, just an ordinary script the agent writes in half an hour and you then run every Monday.
An account audit. The agent walks your account structure looking for the boring things humans skip: ad sets with no conversions in 30 days, overlapping audiences, campaigns with no budget cap, missing UTM parameters. This is work nobody enjoys doing by hand, which is exactly why nobody does it.
Creative analysis. Join ad performance to creative metadata, meaning format, video length, whether a face appears, message type. Only then can you say what works, instead of listing which ad won.
Anomaly monitoring. A script runs every morning and tells you when something moves outside its normal range: CPA jumps 40 percent, spend flatlines, conversions drop to zero. The value is not the analysis, it is finding out on Tuesday instead of the Monday after.
Your guidelines, written down once. Claude Code has the concept of skills, folders of instructions the agent follows every time. Into one of them you put brand tone, banned phrases, metric definitions, report format. Anthropic’s reference brand guidelines skill is public and works as a template. You write it once and the team never repeats it again.
The rule you do not break
The agent does not spend money without a human.
Easy to write, quickly broken, because Meta’s MCP can write by default. An agent that misread one line of a query can pause the wrong campaign or double a daily budget, and it needs no bad intent to do so, just bad context.
In practice that means three things.
Separate read and write access. Use a read-only token for analytics. Keep write access apart, for moments when you deliberately reach for it.
Let the agent propose and the human execute. Have it write what it would change and why, with numbers attached. Approval is a click, not a debate.
Use the permissions the tool already has. Claude Code ships with allow and deny lists, approval prompts and a sandbox. Configure them before the agent touches the account, not after.
Where this breaks
APIs move. Meta updates its Marketing API regularly. A script that worked in March throws an error in October. This is maintenance, not a one-off project.
Metrics look alike and are not. A Meta conversion and a Google conversion do not measure the same event, and attribution windows differ. If you do not tell the agent which definition you use, it will pick one, and your report will look credible while being wrong.
Small budgets are noisy. At seven conversions a week, every “anomaly” is statistical noise. Plenty of teams advertise at spend levels where a week-on-week comparison says nothing. Report on 14 or 28 day windows and save yourself the false conclusions.
Unsupervised runs get accounts locked. A script with no error handling that retries on every failure will hit platform rate limits fast. Build in backoff and a hard ceiling.
Why this matters for a small market
Digital advertising in Slovenia reached 106.7 million EUR in 2025, up 13.6 percent, taking almost 45 percent of the total ad market. According to IAB Slovenia’s 2026 panel, 81 percent of participating companies advertise primarily or exclusively in digital, allocating on average 61 percent of their budget to it.
Most of those companies have no data analyst. Media is bought directly from the platforms and reporting lives in a spreadsheet somebody maintains on top of everything else.
That team is the one that gains most. Not because an agent advertises better than a person, but because the work that used to be too expensive for a small team, meaning in-house analytics, weekly audits and anomaly monitoring, now costs an afternoon.
How to start
Pick one report you build manually every week. Just one. Have the agent automate it end to end, from export to final summary, and run it alongside the old process for a month. When the numbers match, retire the old report and pick the next one.
Do not start with write access and do not start with optimisation. Start with reading, because that is where a mistake shows up without an invoice attached.
If you would rather not spend a month learning this, get in touch. At WebZone we build exactly this kind of automation, which we also covered in our post on workflow automation with n8n.