How to build an AI growth marketing agent (n8n + Claude, step by step) Start with one boring job the agent can do every morning.

an AI growth marketing agent is a program that pulls your numbers, reasons about them with a model like Claude, and hands you a short answer. the best first build is a daily growth report: every morning it fetches yesterday’s spend, installs and day 1 retention, compares them to targets, and posts a summary plus anomalies to Slack or email.

Pick one job before you build anything

pick one job that is boring, repeated daily, and read-only. a daily growth report fits all three: you already check these numbers by hand, the agent only reads data, and a mistake costs you a confusing message, not money. agents that try to do five jobs at once are hard to test and easy to break.

i draw the same line in my own work. i automate research, tracking and reporting, never the voice. anything posted in a community is written by a person who read the thread. my full set is in my n8n + Claude growth automations playbook. for the bigger picture of where agents fit in growth, see agentic growth marketing.

The architecture in plain words

the agent is a straight line of five steps: a timer starts it, plain data nodes fetch the numbers, a model reads them against your targets, a check decides whether anything is unusual, and a message goes out. the key design choice is that the numbers come from data nodes, never from the model’s memory.

here is the flow, left to right:

1. trigger: n8n’s Schedule Trigger runs the workflow every morning, say 7am in your timezone.
2. fetch: three HTTP Request nodes pull yesterday’s spend (for example from Meta’s Ads Insights API), installs (from your attribution tool or store console export), and day 1 retention (for example Mixpanel’s retention query).
3. compare: a Google Sheet holds your targets. a small Code node merges actuals and targets into one clean JSON object.
4. reason: n8n’s AI Agent node with the Anthropic Chat Model sub-node reads that object and writes the summary.
5. deliver and log: a Slack message via an incoming webhook or a Gmail node, then one row appended to a log sheet.

The nodes, step by step

build the workflow in the order the data flows and test each node before adding the next. every node in n8n shows its output after a test run, so you can see exactly what the model will receive. budget two to four hours for a first version if you already have API access to your ad and analytics tools.

step	node	what you set	time
1	Schedule Trigger	daily, 7am, your timezone	5 min
2	HTTP Request (spend)	read-only token, date = yesterday	30 min
3	HTTP Request (installs)	attribution or store export	30 min
4	HTTP Request (retention)	cohort = day before yesterday, day 1 return	30 min
5	Google Sheets (targets)	one row per metric: target, alert threshold	15 min
6	Code	merge into one JSON, compute % vs target	20 min
7	AI Agent + Anthropic Chat Model	system prompt below, low temperature	20 min
8	Slack or Gmail	channel or inbox for the report	10 min
9	Google Sheets (log)	date, inputs, output, run status	15 min

note: yesterday’s installs have no day 1 number yet, so report day 1 retention for the cohort from the day before. if these metrics are new to you, app growth metrics explains each one.

The system prompt to copy

the system prompt tells the model its one job, the exact shape of the output, and the rules it must never break. keep it short and strict. the most important rule is that it may only use numbers present in the input, because a model asked to explain a gap will happily invent a reason if you let it.

you are a growth analyst for a consumer mobile app. every morning you receive one JSON object with yesterday’s spend, installs, cost per install, and day 1 retention for the previous cohort, plus a target and an alert threshold for each.

write a report of at most 8 lines:

1. one line headline: on track, watch, or off track.
2. one line per metric: actual, target, % difference.
3. an “anomalies” section listing only metrics beyond their alert threshold, with one plain-language possible cause each, labelled “possible cause”.
4. one line “suggested check for today” that a human can do in under 15 minutes.

rules: use only numbers present in the input. never estimate or fill in a missing number; write “missing” instead. never recommend changing budgets, pausing campaigns or posting anything; you only report. if the input looks broken (zero spend, negative values, empty fields), say “data problem” in the headline and stop.

Tools the agent needs

this agent needs very few tools, and all of them should be read-only. in n8n, a tool is a sub-node the AI Agent can call, and the node requires at least one. for this job the data is fetched before the agent runs, so the only tool it truly needs is a lookup on your targets sheet.

• targets lookup: a Google Sheets tool so the agent can check a threshold it was not given.
• no write tools in version one. no budget changes, no ad edits, no posting.

if you would rather build this in code, Claude’s tool use works the same way: you define a tool, Claude returns a tool call, your code runs it and sends back the result. the Claude Agent SDK wraps that loop with permissions and hooks, and MCP is the open standard for plugging data sources into agents.

Guardrails that keep it safe

guardrails are the rules that stop a helpful agent from doing damage. for a growth agent there are three that matter most: read-only credentials, a human approval step before any action that spends money or posts publicly, and a log of every run. set them up on day one, before the agent does anything useful.

1. read-only credentials. create API tokens with read scopes only. on Meta, the ads_read permission reads performance data without edit rights. if a token can edit, the agent can edit.
2. human approval before spend or posting. when you later add an action tool, turn on n8n’s human review for tools. the workflow pauses and sends you the exact tool call in Slack, email or chat to approve or deny.
3. logging. append every run’s inputs, output and status to a sheet, so you can see whether the data or the model was at fault.
4. a sanity check. an IF node that stops the run if spend is zero or retention tops 100%.

What it costs to run

this agent is cheap because it runs once a day on a small input. the main costs are n8n hosting, model tokens and your setup time. n8n’s self-hosted community edition is free software; n8n Cloud is a paid plan. all numbers below are illustrative, so check current prices before you budget.

item	illustrative monthly cost	note
n8n self-hosted on a small server	$5 to $10	or a paid n8n Cloud plan instead
Claude model tokens	under $1	about 10k input and 1k output tokens a day on Claude Haiku 4.5 at $1 and $5 per million tokens, per Anthropic’s pricing page
your time	2 to 4 hours once, 15 min a week	the real cost

How to test it before you trust it

test the agent by running it on days you already know the answer to. pick five past days, including one bad day and one day with broken data, and compare its report with what you would have written. only switch on the daily schedule once all five match your own read.

• week 1: run it manually next to your normal check. note every wrong line.
• week 2: feed it a broken input on purpose (zero spend, a missing field). it should say “data problem” and stop.
• week 3: turn on the schedule. read the log every friday for 10 minutes.

what to measure: how many mornings the report matched your own read, how many anomalies it caught that you missed, and minutes saved per day.

Three agents to add next

once the daily report has run cleanly for a few weeks, add agents that follow the same pattern: read-only data, a strict prompt, a human who decides. each one below is a new workflow, not a bigger version of this one, so a failure in one never breaks the others.

1. creative fatigue watcher. flags creatives whose cost per install keeps rising, so you brief new hooks early.
2. review digest. groups new App Store and Google Play reviews by theme and quotes the phrases users repeat.
3. community research log. logs which topics come up in your subreddits. it only reads; a person writes every reply.

for a wider map of what agents can and cannot do in marketing, read AI marketing agents: what they can do and where they fail.