I asked Cowork to "find restaurants that were praised as delicious on our company Slack," and it was useful but a bit scary.

Hello, this is Taisa from Dejima.

Classmethod's internal Slack has many casual chat channels, including some where restaurant recommendations are shared. Posts like "That restaurant was delicious," "You absolutely must try this ramen," "Spicy and amazing." But there are too many to sort through, and searching manually often doesn't yield the information 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 search across company posts and extract restaurant information
• In 20 minutes, narrowed down from 70 hits to 9 useful recommendations with scoring
• However, I realized it could extract "who ate where and when," raising privacy considerations

What I wanted to do

Classmethod's Slack has various 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 look across all of them.

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

What I actually did

My instruction to Cowork was simply this:

Please search Slack and extract restaurant recommendation discussions.
Also provide brief feedback and categories, starting with February posts

Cowork proceeded with the following steps:

Step 1: Keyword search

Used Slack MCP's slack_search_public to search with multiple keywords four times. Searching for "recommendation restaurant," "was delicious," "Tabelog," and "hotpepper" returned a total of 70 hits.

Step 2: Scoring with Python

Since all 70 hits would include 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.

Incidentally, I encountered an issue where the Slack MCP search results were returned in Unicode-escaped format. Japanese text appeared as \u3042\u3044\u3046 and needed to be converted using codecs.decode(). (This was a subtle point where one could get stuck.)

Step 3: Organizing into human-readable format

From 32 items, I extracted 9 that clearly mentioned restaurant names and impressions. I compiled these with categories (ramen, yakiniku, izakaya, etc.) and brief feedback.

Results

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

I discovered that misc channels often contain posts with Tabelog or Google Maps links, providing immediately useful information. It turns out that channels I usually just skim through contained quite practical information.

...and then I got a bit scared

This might be the main point of my post.

After completing the work, I suddenly realized that using this method makes it easy to see "who" ate "when" and "where." Could it even reveal social relationships and food preferences?

You could get the same information by manually searching Slack. But when you delegate to an AI agent, it can instantly aggregate large amounts of information across platforms. This feels like a qualitatively different risk than manual searching.

For instance, if a person makes their information public, technically it would be possible to:

• Create a list of restaurants a specific person frequently visits
• Analyze "who often eats together with whom"
• Infer private preferences from posting patterns in hobby channels

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

I think this is an unavoidable topic for organizations like Classmethod that use AI agents internally. The same issues arise not only with Slack MCP but also with Google Drive and Notion integrations.

From this experience, I feel we at least need to consider:

• How much access should AI agents have to search (only work channels? casual ones too?)
• Should times-* and misc-hobby channels be excluded from AI search by default?
• Guidelines for saving and externally sharing search results
• Whether employees are aware that "AI agents can cross-search internal Slack"

What's technically possible isn't the same as what should be done. Balancing convenience and privacy is a matter of operational rules, not just tool settings. (To be clear, the search only covers public channels and private channels/DMs you're part of. If an administrator did this... it could be 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, I'm left with the question: "Is it okay to continue using this without rules?" If we're going to use AI agents internally, I think we should establish guidelines early on regarding search scope and data handling.

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