A Daily AI Research Agent — Proposal & Approval Workflow

Introduction

"I really should check the analytics" — and then a week goes by without opening them. Sound familiar? When you run an EC site and physical stores, the data you should be watching keeps growing while the time to watch it does not.

So I built an AI research agent that starts automatically every morning. It investigates analytics, search data, store sales, EC, and CRM across the board, then registers improvement proposals. All a human has to do in the morning is glance at those proposals and approve or reject them. This article walks through the whole picture.

Handing the Morning Data Review to an AI

The Limits of Manual Monitoring

Regular data monitoring is the classic "important but not urgent" task. Open GA4, check Search Console, open the store sales CSV, review EC orders and CRM deals — done properly, it eats nearly an hour every morning.

And the hard part is not "looking" but "noticing." Staring at numbers is meaningless on its own; value appears only when you catch a change like "traffic to a certain category has been sliding since last week." This job — comparing the same viewpoints day after day and noticing change — is actually something AI agents are quite good at.

Where the Agent's Job Ends

The agent handles investigation and proposals only. It does not execute.

Every morning it investigates each data source and drafts proposals like "the search ranking for this product page has dropped, so I suggest revising the title," then registers them to the internal business hub. But it never changes the site or settings on its own. A proposal is executed only after an administrator approves it in the admin screen.

Borrow the AI's investigative and analytical power, but keep judgment and responsibility with humans. This boundary is the foundation for using an agent in real operations with confidence.

Overall Architecture

The whole system splits into two parts: an agent running on a local PC, and a business hub in the cloud.

Running the agent itself on a local PC is the key choice — it makes sense in terms of cost, authentication, and access to at-hand data. The sub-articles cover the details.

A Day in the Life — From Launch to Registration

Scheduled Launch and Headless Execution

Each morning, the local PC's Task Scheduler starts Claude Code in headless mode (an automated mode with no human dialogue). The prompt passed at launch already defines which data sources to investigate, which viewpoints to look at, and the output format for proposals.

Even with no human at the PC, the agent autonomously fetches data, analyzes it, registers the results, and exits. By the time you arrive at work, that day's findings are already lined up in the business hub.

Investigating Across Multiple Data Sources

This agent's value lies in its ability to cross-reference multiple data sets.

For example, "search traffic hasn't dropped but sales have" points to a cause inside the site — a line of reasoning that a single tool's report never surfaces. Some insights appear only when you look across sources.

Structuring and Registering Proposals

The agent's output is not free text but structured data with a fixed schema. It emits title, category, proposal body, supporting data, expected impact, and priority as JSON, then registers them to the business hub's API.

Fixing the format enables list views and filtering in the admin screen, and curbs the "proposal quality varies day to day" problem. The prompt and output schema are, I'd say, the heart that determines this system's quality.

Mechanisms for Human Judgment

An Admin Screen to Approve or Reject

Registered proposals appear in a list on the Next.js business hub's admin screen. The administrator reviews each proposal's content and supporting data and chooses approve or reject. Only approved proposals advance to execution; rejected ones are recorded with a reason.

This design of "always passing through human review" is called human-in-the-loop. Since AI proposals sometimes miss the mark, keeping the final call with a human is the condition for running operations with peace of mind.

Slack Notifications and Q&A

Proposals are sent to Slack at the same moment they are registered to the hub. You don't have to open the admin screen deliberately — you notice new proposals while doing your morning Slack check.

What's more interesting is the Q&A flow. Mid-investigation, the agent can ask questions only a human could answer via Slack, like "is this dip in sales due to the sale ending, or is there another factor?" Humans reply asynchronously whenever they have a moment, and those answers feed into future investigations.

Notification Design That Avoids Crying Wolf

The scariest failure mode of automated notifications is sending so many that no one reads them — "crying wolf."

We constrain the agent via the prompt to "only propose high-confidence items" and "don't repeat the same point." Notification trustworthiness deserves as much design attention as the system's technical quality.

Summary

We covered the whole picture of an AI research agent that runs automatically every morning. Three key points:

1. AI investigates and proposes; humans decide — run it safely with a human in the loop
2. Cross-source investigation — notice changes invisible to any single tool
3. Structure and a review path — schematize proposals and run them smoothly via the admin screen and Slack

Each element is explored in the sub-articles below.