A small marketing team has agents read six systems and approves what reaches customers

The task

The marketing team has three people: the marketing lead, a creative director, and a part-time contractor. Nine systems each hold part of the answer to whether marketing is working, and each one reports a number without context from the others. The marketing lead used to check the two or three systems they trusted most, decide, and hope the rest did not say something important.

What they did

The team gave AI agents defined roles, such as competitive intelligence analyst, pipeline analyst, and technical content writer. The larger gain came from skills, which are written procedures for one recurring job. Examples are a monthly competitive review, a clause-level compliance mapping across several security standards, conference submissions, and a monthly audit of how leads arrive. The competitive review skill holds the watchlist, the tier a change must reach to matter, and the rule that it goes back to the marketing lead before anything becomes a battlecard. In one pass, an agent reads HubSpot, Google Analytics, Search Console, Semrush, LinkedIn, and call recordings, then returns a single page on what moved and why it matters. Before the monthly inbound journey report, a validation step checks whether the data can be trusted, and every run includes a counterpoint that questions the results. When a report gets something wrong, the team changes the skill instead of correcting the output, so the same mistake does not come back on later runs.

What the source reports

According to the source, the validation step failed on its first run: Google Analytics 4 showed more new users than sessions, and month-over-month traffic had stayed flat within a few percent for half a year. According to the source, Semrush showed a very different organic visit count that looked more realistic. According to the source, another report showed barely 7% of contact records with an entry URL, but contacts created through a form had one 99% of the time once the records were split by how they were created. The source states that setting up the agents and skills took days, while the checks and approval guardrails have taken months and are still unfinished. According to the source, the author lets a workflow run for at least a month, with several rounds of improvement, before trusting its data and process.

Where the approval sits

The agents monitor data, analyze it, and draft. A person reviews, approves, and publishes anything customer-facing. The competitive review skill returns to the marketing lead for approval before a finding becomes a battlecard. The author describes this review step as the design, and as the only reason for being comfortable running the agents at all.

Worth knowing

This is the company's own blog post, written by its marketing lead, so it is one person's account. It reports no measured time savings, and the author says the guardrails are still being built. The post does not name the AI model or product that runs the agents. The author notes that the agents have credentials into the customer records, analytics, and Slack.

Prompt for your agent

More in this category

• Marketing operations

  A marketing lead has Claude Code draft six weekly meeting docs and reviews every line

  A marketing lead at a startup runs one Claude Code command each Sunday evening to prepare the next week's meeting documents. The lead reviews every line, and nothing reaches a live document until the lead signs off.

  Read the entry
• Marketing operations

  A marketing workspace queues agent work for approval and puts paid calls behind a gate

  A marketer keeps procedures, checklists, and house rules in one Obsidian folder that Claude Code works from. Finished work waits on a board for accept, ask, or deny, and paid API calls stop at an approval note first.

  Read the entry