Behind the Launch: Coheso Agent for Slack
Sep 30, 2025
Written by
The best software disappears. You don't "use" your email client—you just email people. You don't "use" Slack—you just talk to your team. But legal software has always demanded attention: log in here, fill in this form, check back later.
We decided to rebuild from a different assumption: what if legal help just... appeared when and where you needed it?
The Core Insight
People don't want to "use legal software." They want answers. They want help.
In-house lawyers don't want to introduce business friction to gain legal efficiency. They want to meet business stakeholders where they are, without pulling focus away from substantive work.
For teams that live in Slack, legal often faces a flood: quick questions, substantive requests, and urgent needs all stacked together in one unstructured feed. Slack is efficient for sending a request, but legal teams thrive when work flows through systems designed for legal.
Coheso embraces both realities. Business gets the speed and simplicity of Slack. Legal gets the structure and visibility they need to work effectively.
What We Built
Our MVP started with two simple commands:
/ask for questions.
AI searches your legal documents and policies, returns an answer with a direct link to the exact source language, and offers to loop in legal when human validation is needed.
/request for substantive work.
AI interprets your message, determines the right intake form (NDA review vs. contract draft vs. general question), directs you to complete it, captures conversation context in comments, and creates an actionable request. Legal gets notified, and two-way sync keeps everyone updated from their preferred channel.
That’s already simple. But our goal has always been to make change management for business users as close to zero as possible.
Now, both commands can be triggered just by messaging our agent.
@Coheso Agent understands your intent:
If your message can be answered by AI, the agent initiates the
/askpath.If it’s substantive work, the agent initiates the
/requestpath.
For users who want more control, the /ask and /request commands remain available.
With this update, business users don’t have to remember commands and legal teams still get structured, actionable intake. Business teams continue working from Slack while legal operates in systems designed for legal.
The Details That Matter
We obsessed over things that might seem small but aren't:
Source Attribution in Every Interaction
Whether in the product UI, in an email, or even in Slack, every AI response cites the exact paragraph or clause it’s relying on.
Not: “Based on company policy…”
But: “Paragraph 2 of the Employee Handbook says… [link].”
This means:
Always verifiable. Users can click through and confirm the language themselves.
Always transparent. No paraphrasing or “trust us.” The original words are front and center.
Always consistent. No matter the channel, every interaction carries the same trust layer.
We believe trust in legal AI comes from informed verification, not from blind acceptance.
Thread-based memory
With Coheso Agent in Slack, the conversation is persistent. If you start with a question, get a quick answer, and then realize you need legal to review something, the entire context carries forward. There is no need to explain again.
Two-way sync
When the legal team responds in Coheso, comments appear in the synced Slack thread. Replies in Slack are synced to the requests in Coheso.
What We Got Wrong (And Fixed)
We successfully auto-classifed whether a message indicates a question or a request. Initial messages don't always come with a lot of context. While AI works well most of the time, it's operating in a casual environment where context is often sparse, leading to enough wrong paths that we needed escape hatches. We kept the intelligent classification—you can still just mention @Coheso Agent and it will determine the correct path, but /ask and /request commands are available as smart overrides.
We initially wrapped all evidence text in markdown formatting for aesthetics. While this looked clean in our test, we discovered Slack's message length limits break in weird ways, and suddenly evidence sections were splitting mid-sentence. We rewrote the entire chunking logic. Nobody will notice, which means we got it right this time.
The validation flow was particularly tricky. When a document has a "human must review this" rule, we need to block the AI answer but still make the user feel helped, not blocked. The copy went through probably 15 iterations before we landed on something that felt right.
What's Actually Hard About This
Building AI that doesn't hallucinate isn't the hard part anymore (though it's not easy). The hard part is building systems that help people maintain situational awareness. The user needs to know:
Did this come from a document or did the AI make it up?
Should I trust this or should I escalate?
Where can I verify this myself?
What happens next?
We're teaching people to work with AI, not just use it. That's the real product design challenge.
Why This Matters
Legal is one of those functions that everyone needs but nobody wants to need. It's often seen as the department of "no," the place where deals go to slow down. But legal teams aren't trying to be obstructionist—they're trying to protect the company while moving fast. The friction isn't their fault; it's a systems design problem.
What if you could ask your compliance question and get an answer in 60 seconds? What if raising a request took 30 seconds instead of 15 minutes? What if legal could focus on complex judgments instead of answering "can I share this client data with our vendor?" for the 40th time?
That's the world we're building toward.
Try It
If you're a Coheso customer and haven't connected Slack yet, talk to your account manager. If you're not a customer and this sounds interesting, let's talk. I'm easy to find.
Building tools that actually help people is hard. But it's the only thing worth doing.
Shipped with care by the Coheso team
