I asked Cowork to "search for good restaurants mentioned in our company Slack," and while it was helpful, it was also a bit scary.

Hello, I'm Taisa from Dejima.

Classmethod's internal Slack has many casual conversation channels, including some where restaurant recommendations are shared. Posts like "That restaurant was delicious," "You absolutely must go to this ramen place," "It's spicy and amazing." But there are so many, and when you deliberately search, you might not find exactly what you're looking for.

So I asked Claude's Cowork to "Find restaurant recommendations from Slack." As a result, I was able to extract 9 pieces of information in 20 minutes, but at the same time, I felt "this is a bit scary," which is why I'm writing this.

Summary of this article

• Used Cowork's Slack integration (MCP) to cross-search internal posts and extract restaurant information
• Narrowed down from 70 to 9 useful pieces of information with scoring in 20 minutes
• However, being able to retrieve "who ate where and when" raised privacy considerations

What I wanted to do

Classmethod's Slack has several food-related casual channels like misc-ramen, misc-焼き鳥_yakitori, and misc-wine-house. Employees post "this restaurant was good," but since the channels are separate, it's troublesome to view them comprehensively.

What I wanted to do was summarize "recommendations" from February's posts with Tabelog and Google Maps links. Since doing it manually would require going through each channel one by one, I decided to delegate it to Cowork.

What I actually did

My instruction to Cowork was just this:

Please search Slack and extract restaurant recommendation discussions.
Provide simple feedback and categories, starting with February posts

Cowork processed this in the following steps:

Step 1: Keyword search

Using Slack MCP's slack_search_public, it searched with multiple keywords four times. "Recommended restaurant," "was delicious," "Tabelog," "hotpepper" returned a total of 70 hits.

Step 2: Python scoring

Since all 70 results contained too much noise (including business "recommendations"), I used a Python script to assign scores to food-related keywords and filter the results. This narrowed it down to 32 items.

By the way, I encountered an issue where Slack MCP search results were returned in Unicode-escaped format. Japanese text appeared as \u3042\u3044\u3046, requiring conversion with codecs.decode(). (This was a subtle point where I got stuck.)

Step 3: Organizing into a human-readable format

From the 32 items, I extracted 9 with specific restaurant names and impressions. I compiled them with categories (ramen, yakiniku, izakaya, etc.) and brief feedback.

Results

The process took about 20 minutes. I got a list of 9 restaurant recommendations with categories, poster comments, and links.

I discovered that misc-type channels contained many posts with Tabelog and Google Maps links, providing immediately useful information. Practical information was buried in channels I usually just skim through.

...And the scary part

This might be the main point.

After finishing the work, I suddenly realized that this method easily reveals "who" ate "when" and "where." Could it reveal social relationships and food preferences too?

You could get the same information by manually searching Slack. But delegating to an AI agent allows you to instantly aggregate large amounts of information across channels. This feels like a qualitatively different risk than manual searches.

For example, if someone has made this information public, technically the following would be possible:

• Create a list of restaurants a specific person frequents
• Analyze "who often goes to eat with whom"
• Infer private preferences from posting patterns in hobby channels

How should we think about internal data searches in the AI agent era?

I think this is an unavoidable topic for organizations like Classmethod using AI agents internally. The same issues arise with Google Drive and Notion integrations, not just Slack MCP.

From this experience, I feel at least the following points need consideration:

• How much search access should be allowed to AI agents (only work channels? casual ones too?)
• Should times-* and misc-hobby channels be excluded from default AI search targets?
• Guidelines for saving and external sharing of search results
• Do employees even know that "AI agents can cross-search internal Slack"?

What's technically possible and what's appropriate are different things. Balancing convenience and privacy is about operational rules, not just tool settings. (For the record, the searchable content is limited to public channels and private channels/DMs you're in. If an administrator did this... it could be a bit scary if they could see everything.)

Conclusion

Using Cowork × Slack MCP to dig through internal posts is quite practical. I got the information I wanted in 20 minutes.

However, because it's so convenient, the question remains: "Is it okay to continue using this without rules?" If we're going to utilize AI agents internally, I think we should establish guidelines early for search scope and data handling.

By the way, among the 9 extracted restaurants, there were several outside Tokyo, which made me want to visit Fukuoka.