MSBAi Org Memory — Stakeholder Guide

For the MSBAi operating team. This page explains how to interact with the MSBAi org memory through K-ai, what happens to your message, and how to verify what the agent did.

What this actually is

The MSBAi org memory is a living record of the program — decisions, open questions, action items, meeting notes, course designs, stakeholder context. K-ai is the agent that accesses and updates this memory on your behalf. The memory lives in a private GitHub repo; K-ai is the interface.

Think of it as a team brain, not a chatbot. You can ask the agent what the team has decided, tell it new decisions, flag open questions, or forward raw material (email threads, meeting notes) for it to file. Everything K-ai learns is immediately available to everyone else with access, through every channel.

This framing matters because it’s what makes the system valuable as people join and leave the program: institutional knowledge gets captured the moment it’s spoken, rather than walking out the door with a departing team member.

You can reach K-ai through email or Telegram. Both channels read the same org memory and support questions and updates.

Ways to reach K-ai

The former live-site chatbot was retired on 2026-08-15 and is no longer an access channel.

What you can send

• Questions — “What’s the current capstone structure?” “Who’s teaching BADM 557?” “What did we decide about prep courses?”
• Decisions — “We’re locking in live synchronous interviews for MSBAi admissions.”
• Action items — “Amber will draft the preparatory-course policy by April 21.”
• Open questions — “Still unresolved: what’s the passing threshold for the Canvas prep quizzes?”
• Meeting recaps or forwarded email threads — K-ai will parse and file them.
• Corrections — “That decision about Power BI is wrong, we’re going with Tableau” (K-ai updates the record).
• Documents and PDFs — attach a PDF to your email or Telegram message and K-ai will read and process it.

Intentional overrides

If you’re knowingly contradicting something already in the record, add the phrase lore: intentional anywhere in your message. This tells K-ai to skip the conflict check, supersede the old entry, and link the two — rather than flagging the contradiction for human review.

Example: “lore: intentional — we’re moving the priority deadline from June 4 to June 18.”

Without this flag, K-ai will detect the conflict and ask for clarification before updating.

How input gets processed (under the hood)

1. You send a message by email or Telegram.
2. Access check — email and Telegram use lists of pre-approved stakeholders. If you send an email and don’t get a reply, your address may not be on the list — contact Vishal or Amber.
3. Agent reads the message and consults the repo: program facts (program/curriculum.md), prior decisions (discussions/DECISIONS.md), action items, open questions, raw email and meeting transcripts.
4. Agent answers you with grounded context (links back to specific files), and/or updates the repo.
5. Repo updates are committed and pushed to GitHub automatically. Every change is a Git commit you can audit.
6. The program site rebuilds on push and reflects published content shortly afterward. Ask Vishal for access through an established private channel; credentials are never sent in K-ai onboarding messages.

What gets updated automatically

When K-ai detects new facts in your message, it files them into the right place:

• discussions/DECISIONS.md — confirmed decisions with date, stakeholders, rationale, source
• discussions/ACTION_ITEMS.md — tasks organized by owner
• discussions/OPEN_QUESTIONS.md — unresolved items needing stakeholder input
• discussions/raw/ — the original email or transcript is stored verbatim as source material for every decision
• Program facts (e.g., curriculum, course names, credits) are only changed after the agent references the source-of-truth rules and traces dependencies

Every decision links back to its source, so you can always trace “why did we decide that?” back to the original message or meeting.

Note: the discussions/ folder is not published to the live site — it lives in the private GitHub repo. This keeps raw email content off the public web while still letting K-ai use it as working memory.

What humans still do

K-ai is a collaborator, not an autopilot:

• Major structural changes (program length, credit totals, launch dates) still go through Vishal for final approval before committing.
• Contradictions are flagged for human review, not auto-resolved.
• Ambiguous asks get a clarifying reply instead of a guess.
• Meeting transcripts are processed into decisions/actions/questions, but the relevant functional lead typically reviews the extraction before it’s treated as final.

Updating action items, roles, and access

Action-item updates are annotated, never erased. If you email in an update to an existing action item — “X is done”, “reassign that to Y”, “the deadline moved” — K-ai finds the existing item and appends a dated note recording what changed and who said so. The original text stays in place, so the history of the item is always visible; nothing is silently rewritten or deleted.

Role and access changes can never be made by messaging the agent. Asking K-ai to add someone to the allowlist, change an access level, or alter a role (including your own) will not work by design — those changes are made only by the program admin (Vishal) directly. K-ai will surface your request to him and let you know it has done so.

How review is structured

K-ai uses structured human review to check what it is doing and whether it is accurate.

The outer loop has functional leads by domain:

K-ai flags which domain an update belongs to, and the applicable approval route is recorded in the authorized-sender roster. Access and role changes remain admin-only.

How to verify what K-ai did

After you send something, you have a few ways to check:

1. Read the reply — K-ai responds on the same channel, typically with a summary of what was updated and links.
2. Check the live site — https://msba-online.pages.dev. Public-facing pages live under /program/, /courses/, /design/, /strategy/. Internal tracking (decisions, actions, open questions) is in the private repo.
3. Ask K-ai — “What did you update in the last hour?” or “Show me decisions from this week.”

Useful ways to get started

• Start with a question you already know the answer to. Lets you calibrate how grounded K-ai’s responses are.
• Use email as your primary channel — it is two-way, logged, and creates a paper trail.
• Send a deliberate correction (“actually, that’s wrong because…”) and see whether K-ai updates the record accurately.
• Forward a real email thread and see how decisions and actions get extracted.
• Report anything that feels wrong — wrong attribution, missing context, over-confident answers, or ignored messages. These reports help improve the system.
• Think about your domain. If you are Amber, test operational accuracy. If you are Ron or Maria, test curriculum accuracy. Your domain expertise is exactly what K-ai needs to be checked against.

Privacy and scope

• Internal decisions, action items, open questions, and raw source material live in a private GitHub repository. Only designated program content is published to the program site.
• K-ai does not have access to your personal email, calendar, or files — only what you explicitly send.
• The allowlist prevents strangers from poisoning the knowledge base.

Questions about K-ai? For operational and administrative questions, contact Amber Glynn. For curriculum and pedagogical questions, contact Vishal Sachdev, Ron Guymon, or Maria Rodas. For anything else, email Vishal (vishal@illinois.edu) or ask K-ai.