Most marketers still use AI as a chat box: ask, receive, copy. It works, but the real gain is elsewhere. When you put AI inside the workflow, repetitive work moves forward while you are away. Here is what the automated part of my week looks like.
The architecture: three layers
The whole setup rests on separating three roles. The data layer carries facts: Google Ads scripts, GA4 exports, platform reports. The reasoning layer interprets: Claude, inside a project that knows the brand's persistent context, summarises, clusters, drafts. The transport layer moves things: automation tools like Make/Zapier/n8n carry data from source to reasoning, and output from reasoning to its destination (email, CRM, archive). The separation matters because the most common setup mistake is piling all three jobs onto one tool: making a chat assistant fetch data, or making an automation tool do the interpreting.
Monday 08:00 — the report draft
A scheduled task pulls last week's channel performance (Google, Meta, organic), summarises it, flags anomalies and writes a draft in client language. When I sit down I don't write the report; I add judgement, write the next step and send it. In Claude I build this with Projects (the brand's persistent context: goals, tone, past reports) and Scheduled tasks (self-running at a set hour); the output arrives as an Artifact — a dashboard or report.
Daily — search term cleanup
A Google Ads script lists wasteful search terms; AI clusters them and proposes negatives. Nothing is applied without my approval — but the decision shrinks from a 40-minute review to a 5-minute sign-off.
Weekly — creative variants
Ten variants of the winning creative are generated (copy + visual direction), screened against the brand guide and entered into testing. Human approval is mandatory: brand consistency and usage rights are not things to delegate to AI.
Real-time — the lead flow
A form or WhatsApp lead lands in the CRM, gets pre-qualified, a notification arrives; a reply draft sits ready. Here Make/Zapier-style tools work together with AI.
How to choose your first automation
Whoever tries to automate everything at once finishes nothing. I use three filters: the job must be frequent (repeating at least weekly), rule-bound (if you can describe what to do in writing, it can be automated) and low-risk (if the output is wrong, the result should be fixable, not embarrassing). Typical first candidates that pass all three: the weekly report draft, search-term clustering, competitor price/message monitoring digests. Typical wrong first candidates that fail them: budget changes, client communication, bid strategy.
Designing the approval gates
Every automation's output falls behind one of three gates: inform (nobody approves, it is just read — like an anomaly alert), approve-to-apply (the suggestion waits ready; I apply or reject with one click — like negatives) and draft (a human works on it and sends it — like the client report). You pick the gate by the risk of the job; an automation that earns trust over time can move up one gate. What's wrong is the gateless automation: the "let AI change the settings itself" urge turns into an expensive lesson at the first silent error.
The maintenance truth: automations break too
The part no brochure mentions: APIs change, formats break, a source quietly stops sending data. Silent breakage means a wrong report — and a wrong report is more dangerous than no report. Two habits manage it: every automation's "last successful run" time stays visible, and outputs get a sanity check (an unexplainable cliff versus last week means suspect the setup first). A ten-minute monthly maintenance lap catches most surprises.
The 30-day setup plan
- Week 1 — inventory: For one week, note every repetitive task you do and how long it takes. At the end, pick the first candidate that passes the "frequent + rule-bound + low-risk" filter; for almost everyone it is the weekly report draft.
- Week 2 — one automation: Build only that. Connect the data source, give the reasoning layer the brand's context (goals, tone, sample outputs), drop the output behind the draft gate. Don't perfect it; make it run.
- Week 3 — the side-by-side period: Compare the automation's output with your hand-made version. Differences get folded into the prompt and the context. This week is the automation's job interview.
- Week 4 — handover and the second candidate: If the output has earned trust, hand the job over and measure the time reclaimed. Only now start the second automation. One a month makes twelve by year's end — each of which actually works.
The three questions I ask when picking a tool
The answer to "which automation tool" is not a brand but three criteria: does it have ready connectors to your data sources (otherwise every integration becomes a mini software project), is there someone on the team who will keep it alive (an automation that dies when its builder leaves costs more than one never built) and how loudly does it fail (the flow that stops silently is the most dangerous flow). If all three hold, the difference between Make and n8n is a matter of taste.
What I don't automate
Budget decisions, the final text that goes to a client, and anything carrying policy risk. AI buys speed; the decision and the responsibility stay with me. Draw that line clearly and automation stops being frightening and starts giving your week back.
Short questions, short answers
Do I need to code for this? For report drafts and clustering, no; a chat interface and ready connectors are enough. Servers, APIs and custom scripts are the higher gear — shift when the need appears.
Where is the line on feeding client data to AI tools? Personal data doesn't reach the reasoning layer without anonymisation; performance data (campaigns, costs, conversions) is the job itself. Checking data-processing clauses in contracts comes before the setup.
The team asks "will it take our jobs". My experience is the opposite: automation takes writing the report, not defending it. The reclaimed hours go to analysis, testing and talking to clients — the un-mechanisable, value-producing part of the work.