Work

Case-y

An always-on AI teammate for case teams, in Slack and email. It knows what the whole group is allowed to know.

Role
Created and architected it; led the build
Team
Grew to five
Started as
A pilot of “OpenClaw for case teams”
Lives in
Slack and email
Can
Research, build decks, run long tasks, draft replies
Status
Launched at BCG

Most AI assistants work for one person. Consulting work is done by teams.

I proposed Case-y as a pilot of “OpenClaw for case teams”: one AI teammate a whole case team could hand work to. I designed it and led its build. The goal was an agent that:

  • does real work, beyond chat
  • is always on
  • has its own identity, so the team treats it as a teammate and nobody has to sign their name to a first draft from AI
  • has the context of the firm, from both personal and firm systems
  • shows up where the team already works: Slack and email

The name is a play on “case,” and it sounds like a person’s name too.

The hard question: what should it know?

The easy answer is “whatever is in this chat.” That isn’t how a case runs. A quick chat can fade out, or turn into the main thread for a big piece of work. Work starts in a meeting, moves to Slack, and ends with a file sent by email. People are on several cases at once, and teams overlap.

So we asked a second question: how many Case-ys should there be? One per person? One per case? One for everybody? None of those worked. Instead, Case-y shows up fresh for each conversation, with exactly the context that conversation should have.

One rule

Case-y can see anything that all members of a group can see.

If Case-y and I talk one-on-one, that stays between us. If Maya joins, the Case-y in that chat can’t see our private conversation. It isn’t keeping a secret. It never had the information.

Sometimes you want Case-y to share more. You can tell it that something is fine to share with the group. You can also ask a colleague. Say Sam made a great page and I want to know how. My Case-y asks his. Sam sees the request and what would be shared, and he decides.

The same idea works when I don’t know who has the answer. I can ask Case-y whether anyone has solved this problem before. If someone has, they get a request. If they say no, I never learn they were asked. Private conversations stay private until their owners say yes.

We didn’t plan this feature. It emerged from the rule above, and it became one of the most distinctive things about Case-y.

Built for the lethal trifecta

An AI teammate brings together three things that are risky in combination: it reads private data, it takes in untrusted content, and it can send messages. That mix is sometimes called the lethal trifecta, because it is how agents get tricked into leaking information.

I designed Case-y’s security around it. Case-y runs in isolated containers, and its credentials sit behind a proxy it can’t read. The one rule above is enforced in the data layer, not by asking the model to behave. And every outbound email needs explicit human approval.

Working like a teammate

In a group chat, Case-y follows along and speaks up only when it helps. It can research, build decks, and work through long tasks overnight. It also checks in on its own. When a client email arrives, it may research an answer and ask whether I want to send its draft. More often than not, the draft is good.

Case-y doesn’t live on my laptop, so I can message it from a plane and it keeps working. It can also draw on caseOS for the context of a case.

Where it went

Once we had a server, Case-y was working in Slack and email within three weeks. The team behind it grew to five, and Case-y has since launched at BCG.