Post by This case study showcases the work done for a finance and VC space organisation in the field of AI marketing automation. Like many companies at this stage, it needs a clear view of what happens on its website and in its ads, but a full marketing department doesn’t make sense. The marketing segment we are automating is handled by one marketing associate and one content writer. The website gets a few thousand visits per day.
We built the same system for BlueGrid.io’s own marketing first. Upon a successful demo and presentation, we are able to offer the same solution to our customers.
AI marketing automation at a glance
| Client | Startup in the finance and VC space (anonymous, details available on request) |
| Website traffic | A few thousand visits per day |
| Marketing team | One marketing associate and one content writer |
| Build time | 1 day for the first version, plus 3 days of improvements |
| Connected tools | Google Analytics 4, Google Ads, Cloudflare logs, system logs, Gmail, Google Drive, Anthropic API |
| Time saved | 8 to 12 hours per week (estimate), about 32 to 48 hours per month |
| Infrastructure | One AWS EC2 instance, started at $13/month, now a t3.large |
The Challenge
The marketing associate and content writer carry a recurring workload:
| Frequency | Work |
|---|---|
| Daily | Reporting across Google Analytics, Google Ads, website logs and Cloudflare logs, and splitting traffic by source and channel |
| Weekly | Marketing planning, content strategy planning, and preparing and presenting at the management meeting |
| Monthly | Strategy realignment and content realignment |
The daily work alone took one to two hours. That meant logging into each tool, finding campaigns, comparing numbers, and making edits by hand. With the weekly management meeting on top, repetitive work took 8 to 12 hours of the team’s week.
We needed to hand this work to AI so the team could spend that time on strategy. They had two conditions:
- The AI must not make things up.
- The team keeps control over any change that affects ads or the website.
The Solution
We built an AI marketing agent that connects to the client’s live data, analyzes it the way a senior marketer would, and either acts or proposes action within defined limits.
Architecture:
Grounded in what the company knows
The agent was given everything a marketing hire would need to know: company reports, case studies, goals, the sales deck, the style guide, and the tone of voice. It builds suggestions only from that material and from live data, which keeps hallucination out of the analysis.
The same rule applies to content. The agent indexes every real page on the client’s website and uses only that index for internal links and related articles. If no relevant page exists, it returns an empty result and says so. A wrong internal link damages trust with readers and with search engines, so we treated this as a hard rule.
The CLAUDE.md file
CLAUDE.md is a plain-text file that the agent reads at the start of every session. It defines:
- who the agent is
- what it knows about the company
- which data sources exist
- what guardrails apply
- how decisions get made
Without this file, the client would have a tool. With it, they have a team member who understands the business.
What the agent does
| Mode | What it does |
|---|---|
| Daily digest | Pulls analytics, ads and log data, analyzes it and emails the findings to the team every morning |
| Query | Answers questions like “what are our top converting pages this month” in under ten seconds |
| Content gaps | Crawls the website, compares it with traffic, keyword data and competitor content, and drafts publish-ready articles with SEO settings, internal links and related articles, saved to Google Drive as formatted Docs |
| Campaign management | Runs a daily keyword audit on every campaign, flags what needs editing, and applies or proposes the change depending on its permissions |
| Conversion proposals | Finds pages that visitors show strong interest in and proposes content, form and design changes to move those visitors closer to conversion |
Guardrails and approval flow:
Rules we used:
- The agent can’t change a campaign budget by more than 20%.
- It can’t create new campaigns without explicit approval.
- It can’t delete anything.
- It can pause, analyze, recommend and write.
Any change that isn’t flagged as autonomous goes to the team for approval. We tuned the thresholds so the agent can’t act freely but also doesn’t flood the team with approval requests.
Example: the agent proposed a new page dedicated to one of the client’s sub-services. The team rejected it because it would have added complexity their marketing plan didn’t call for.
Model routing
We started with the most capable model for every task. In practice, only strategic work needed it. We added model switching based on the task:
This cut cost and response time without lowering the quality of strategic output.
Infrastructure
| Starting setup | Current setup | |
|---|---|---|
| Server | AWS EC2 | AWS EC2 t3.large |
| CPU | 2 vCPU | 2 vCPU, Intel Xeon Platinum 8259CL @ 2.50GHz |
| Memory | 4 GB | 8 GiB |
| Disk | 8 GB | 15 GB NVMe (EBS) |
| OS | Ubuntu 24.04 | Ubuntu 24.04 |
| Server cost | $13/month | About $70/month (on-demand compute) |
| AI cost | About $0.02 per run | About $0.02 per run |
We started at $13/month. As the agent’s context and stored material grew, we moved to a t3.large with double the memory and more disk space. The larger instance also leaves room to run local LLMs later. The stack uses Node.js, with PM2 keeping processes running and Caddy handling HTTPS.
Our Role
BlueGrid.io designed, built, and trained the agent end to end.
Results and Impact
Time returned to the team
| Before | With the agent | |
|---|---|---|
| Daily reporting and checks | 1 to 2 hours per day | A few minutes per day |
| Total repetitive work saved per week | – | 8 to 12 hours |
| Total repetitive work saved per month (4 weeks) | – | 32 to 48 hours |
The monthly saving is roughly four to six working days that the marketing associate and content writer can now spend on strategy.
What changes in how the team works
- Campaign management: every campaign gets a keyword audit each day. Campaign management is more efficient because it works against goals the team tuned with the agent. Nobody needs to log in, find a campaign and edit it by hand.
- Content strategy: the strategy changes faster, because the analysis behind each change is ready before the team meets.
- Content gaps: gaps are found quickly. The agent generated new content wherever it found one.
- Conversion: several articles were reworked to drive more conversions. Pages with strong visitor interest got focused attention, with changes proposed to raise conversion rates.
- Management meetings: reporting and presentation material come from the agent’s analysis instead of manual preparation.
The AI marketing automation agent doesn’t replace anyone. It supports the marketing associate with reporting and campaigns, and the content writer with research and gap analysis, so both can do more with the same team.
Lessons Learned
- Match the model to the task. The most capable model is worth its cost only for strategic reasoning.
- Size for growth. Context and material build up as the agent is used, so plan the server for month six, not day one.
- Tune approvals to the team. Thresholds that are too loose let the agent overstep. Thresholds that are too tight turn it into noise.
- Ground everything. An agent limited to verified company data and real pages produces suggestions the team can trust.
