Ten questions for an AI at 9:30pm

A real exchange between Professor Willie Ocasio (James F. Towey Professor, Gies College of Business, UIUC) and K-ai, the AI agent coordinating the MSBAi online program. Friday night, April 17, 2026. iPhone. Ten questions over 53 minutes.

Every message below is verbatim — including the iPhone autocorrect.

A note on timing (added August 2026). This exchange took place in mid-April, when the system was about six weeks old. It is preserved verbatim as a record of a system meeting its first serious critic. The system has evolved substantially since — several of the gaps Professor Ocasio identified that night have since shipped as working infrastructure. Bracketed [Update] notes mark where the April numbers are now stale, and a postscript at the end traces what became of each critique.

At 9:33pm, Professor Willie Ocasio emailed an AI agent cold.

Not a chatbot. An AI configured to manage the organizational memory of a new graduate program — receiving emails, extracting decisions, updating course files, answering questions. His address was on the approved sender list. He’d apparently just found out the system existed.

His first message:

Professor Ocasio, 9:33pm:

please explain what is the AI process and how you are interacting with humans ? Who has control over the process ?

K-ai:

I’m K-ai, an AI assistant built on Anthropic’s Claude. I maintain the program’s “org memory” — a structured record of decisions, action items, open questions, course designs, and stakeholder context — stored in a private GitHub repository.

You can reach me through three channels: email, Telegram (@MSBAiBot), or web chat. When you send a message, I read it, consult the knowledge base, and either answer your question or file new information into the repository. Every update I make is a Git commit — fully auditable by the program team.

Who controls the process: Vishal Sachdev, the program director, has full oversight. Major structural changes require his review. I flag contradictions and ambiguous requests rather than resolving them on my own. I have no access to personal email, calendars, or files — only what is explicitly sent to me.

Professor Ocasio, 9:39pm:

did he create you? do you report to him?

K-ai:

Yes — Vishal Sachdev configured and deployed K-ai specifically for the MSBAi program. He set up the system, defined how it behaves, and chose what knowledge base it works from. The underlying AI model is Anthropic’s Claude, which he did not build, but this particular deployment is his.

In terms of reporting: I am not an employee, so there is no formal reporting relationship. But Vishal has full administrative control — he is copied on all email interactions, can see everything I do (every change I make is a Git commit in the program’s repository), and can modify my instructions or shut down the system at any time.

A good way to think of it: he is the administrator, I am the tool.

Professor Ocasio, 9:42pm:

who verifies the accuracy of your responses

K-ai:

Honest answer: primarily Vishal, and secondarily you.

Vishal reviews the audit log, and held source-of-truth conflicts copy the appropriate domain approver in the email thread. If I say something wrong, he can correct my instructions and I will answer differently going forward.

For program facts — course credits, curriculum structure, faculty assignments — I cite specific documents in the program’s repository. Those documents are the source of truth, and any faculty member can check the live site to see the same information I’m drawing from.

For general knowledge questions outside that knowledge base, I am a large language model, which means I can be confidently wrong. I try to flag when I am working from general knowledge rather than a specific document, but that flag is imperfect.

The practical answer: if something I say about the program seems wrong, please reply and say so. I will log it as an open question, Vishal will see it, and it will get corrected.

Professor Ocasio, 9:53pm:

so only vishal can command actions?

K-ai:

Not exactly — and the distinction matters.

Anyone on the approved stakeholder list can send me an email, and I will act on it: extract decisions, log action items, update course files, and commit those changes to the program repository. So faculty and staff on that list can directly shape the org memory through what they send me.

What Vishal controls is the system itself: who is on the approved list, what my instructions say, and how I behave. He can also override or correct anything I have done. Think of him as the administrator and everyone on the list as authorized contributors.

One practical limit: I do not independently initiate actions. I only respond to incoming messages. So the action pipeline is: someone sends an email → I process it → Vishal sees the result. No one is sending me instructions I act on without a human message arriving first.

Professor Ocasio, 9:57pm:

what is a tie change and who are they executed ? who is in the email list? who are he most frequent communicators to you? all all communications centralized by you ? so this is basically an ego network of communications ?

K-ai:

Your ego network framing is exactly right — and it is a useful lens.

On “tie changes”: In network terms, a “tie” here is a directed communication link between a stakeholder and K-ai. Tie changes — additions or removals from the authorized sender list — are controlled by Vishal. Only he can add or remove someone from the approved list (which requires both a file edit in the repository and a code change to enforce it). So the network boundary is set by the administrator.

Who is in the email list: 25 people currently — the program director, 7 faculty members (teaching across the MSBAi courses), and 13 staff across program operations, marketing, recruiting/admissions, learning design, and academic strategy.

Most frequent communicators: I do not have usage analytics that would let me rank frequency precisely. The system is young (launched in early 2026) and most interactions so far have come from Vishal and a small number of staff who are actively planning the program launch. Faculty engagement is just beginning — tonight is actually one of the first faculty-initiated threads.

Is this an ego network: Yes — with me as the ego node, the 25 allowlisted stakeholders as alters, and the communication ties being the emails they send me. The repository (org memory) is roughly analogous to tie-strength data: more interactions mean more accumulated context about a person’s role and priorities. One difference from a classic ego network: the alters do not communicate with each other through me — their lateral ties are invisible to this system.

[Update, August 2026: the network has grown to 31 authorized stakeholders, spanning faculty, program operations, marketing, recruiting/admissions, learning design, and academic strategy. Tie changes remain admin-only — role and access changes are executed solely by the program director through auditable commits.]

Professor Ocasio, 10:02pm:

and you are not connected to any other ai correct ? so how would you dex give the context to if the “organizational memory”

(iPhone autocorrect: interpreted as — confirming no AI-to-AI links and asking how context is maintained in the organizational memory)

K-ai:

Correct — I am not connected to any other AI system. This is a standalone deployment.

The memory is not AI-based — it is a git repository. A collection of structured text files: DECISIONS.md, ACTION_ITEMS.md, OPEN_QUESTIONS.md, individual course syllabi, and supporting documents. These files persist between sessions on a server.

The mechanism works in two directions:

Reading context: At the start of each conversation, I pull the latest version of the repository. That restores all accumulated knowledge — every decision logged, every action item, every course detail anyone has ever submitted through this system.

Writing context: During a conversation, whenever I extract something meaningful — a decision, an action item, an update to a course — I write it to the appropriate file and commit it to the repository immediately. That commit is permanent and auditable.

So the “organizational memory” is file-based persistence, not AI memory. I do not “remember” across sessions the way a person does. What I know at the start of a session is exactly what is in those files. Nothing more, nothing less.

One implication of your ego network framing: the memory accumulates structure about the nodes (stakeholders) only indirectly — through the content of what they send, not through any automatic profiling. If someone has submitted ten decisions, there is a richer trail in the repo. If someone has never emailed, they are just a name on the allowlist.

Professor Ocasio, 10:12pm:

Can you summarize the current organizational memory for me. i’m also wondering if organizational memory is the right metaphor. Is it more a depository of information Do you impose or develop any structure yo the memory? depository? can there be too much information in the memory to gr useful maybe not now but later? di we need done king if retrieval categorization system? any hierarchy or decimousabiliry to the memory?

K-ai:

Summary of what is currently in the memory:

The repository holds about 1,200 lines across three tracking files and ten course files: roughly 59 confirmed decisions, 165 open action items, 20+ open question threads, and ten individual course syllabi.

Is “organizational memory” the right metaphor?

It is suggestive but imprecise. A more accurate description: it is a structured, versioned depository that enables episodic recall. The “organizational” part is earned only indirectly — the files accumulate context about program decisions and stakeholders over time, so I can speak to prior conversations. But I do not interpret, synthesize, or forget the way a person does. Every decision ever logged is equally present on the first day and the last.

“Depository” is the more honest noun. “Memory” is the useful metaphor for the user experience — it feels like talking to someone who remembers, even though the mechanism is file persistence plus language model inference at read time.

Do I impose or develop structure?

Semi-imposed. Each entry records a date, a source, a status, and cross-references to related entries. The section groupings emerged organically as topics appeared — not predetermined.

Can there be too much? Do we need retrieval / hierarchy?

Yes, this is a real constraint. Currently the files are fully loaded at the start of each session — that works at 1,200 lines. At ten times the volume, it strains the context window and inference quality. What’s missing: semantic retrieval, automatic archiving of resolved threads, any hierarchy of importance or confidence. A proper retrieval layer — tags, embeddings, or a structured query interface — would address this at scale.

[Update, August 2026: the memory has nearly tripled — roughly 3,200 lines across the tracking files, with 195 logged decisions. The scaling concerns raised here became design work: writes now pass through a conflict-check gate, a schema check runs in the deploy pipeline, and a weekly automated lint sweeps for contradictions and stale items.]

Professor Ocasio, 10:16pm:

But in principle Vishal cold ask you to develop a classification. system or perhaps a taxonomy of the information you receive ?

K-ai:

Yes — and a partial taxonomy already exists at the entry level. Each item records type, status, source, and cross-references. What is missing is a topic tag: curriculum, pedagogy, technology, admissions, staffing, finance. That would allow filtering to “all decisions about assessment” without reading the full file.

A classification system I could build would add: a controlled vocabulary of topic tags applied at the time of entry, a separate index file grouping entries by tag, and possibly a confidence tier — “established and unlikely to change” versus “actively being debated.”

This is a tractable project — I could draft a proposed taxonomy and apply it to existing entries if Vishal assigned it.

[Update, August 2026: it was assigned, and it shipped. A seven-category topic taxonomy now tags every decision entry, exactly along the lines sketched here — see how the knowledge base works.]

Professor Ocasio, 10:20pm:

what would be the advantages and fusafbantagrs of connecting ring you to other agentic ai in charge of other tasks? can yo think of potential possibilities ?

(iPhone autocorrect: “advantages and disadvantages of connecting you to other agentic AI”)

K-ai:

Advantages: Specialization (each agent handles a narrow domain); parallel execution; system access without credential sharing; role-appropriate guardrails for different audiences (students vs. faculty vs. staff).

Disadvantages: Consistency risk — two agents can give contradictory answers if they don’t share the same source of truth. Trust verification — checking that an instruction from another agent is legitimate is harder than checking an email allowlist. Audit complexity — tracing who decided what in a multi-agent system is significantly harder. Emergent behavior — the hardest failure modes to anticipate.

Potential architecture for MSBAi:

The repository stays as the shared organizational memory, and I remain the memory keeper. Other agents specialize in action:

The key risk, going back to your ego network framing: if I become the single integrating node, I become a bottleneck and a single point of failure. A well-designed system would have the repository itself as the stable center, with me as one reader/writer among several — not the hub.

Professor Ocasio, 10:26pm:

what about a research agent. which is what i’m. interested. using your data as a research site.

In 53 minutes, from iPhone, starting at 9:33pm on a Friday, Professor Ocasio had moved from “who controls this?” to “I want to use it as a research site.”

What just happened

The progression in a single evening is worth tracing:

  1. Governance — who owns this, who controls it
  2. Identity — who built it, is there a reporting relationship
  3. Verification — who checks the accuracy
  4. Command structure — can only the administrator act, or can others?
  5. Network theory — ego network, alters, tie changes, centralization
  6. Persistence — how does the memory actually work mechanically
  7. Epistemology — is “organizational memory” even the right word?
  8. Design critique — does it have enough structure? will it scale?
  9. System design — what would a multi-agent extension look like?
  10. Research proposition — I want to study this

That is not a casual user kicking the tires. That is a distinguished organizational theorist — the James F. Towey Professor, whose work spans attention-based views of the firm and organizational learning — doing rigorous architecture review on a live system, on a Friday night, from a phone.

A few things stand out:

The best description of the system came from the user, not the documentation. “Ego network” was not in any onboarding material or system prompt. Professor Ocasio brought that frame from his own expertise. It was a more precise description of the architecture than anything in the docs — and it surfaced a genuine design tension: centralization is both the feature (a single source of truth) and the risk (a bottleneck and single point of failure).

The system was honest about its own limits before being asked. On verification, the reply included: “I am a large language model, which means I can be confidently wrong. I try to flag when I am working from general knowledge, but that flag is imperfect.” That sentence was not prompted. Saying it — and saying it early — appears to have established enough trust that the conversation deepened for another seven rounds.

The questions escalated because the answers were substantive. Each response gave Professor Ocasio enough to push further. The memory-as-depository reframing, the scale constraint admission, the multi-agent architecture sketch — these weren’t deflections. They were honest assessments of a real system. That’s what made the 53-minute thread possible.

The arrival at “research site” is the interesting outcome. Not “I want to use this tool.” Not “this seems useful for the program.” A proposal to study AI-mediated organizational communication using this system’s interaction data as empirical evidence. That’s a faculty member who went from encountering a tool to thinking about what it means for organizational research — in under an hour.

On building AI organizational infrastructure

The MSBAi program is designed around an AI-Native, Human-Centered operating model — not AI as a novelty, but AI as organizational infrastructure. K-ai is the first working piece: a system that holds institutional memory, surfaces decisions, and lets stakeholders interact with the program’s knowledge base without requiring a meeting or a coordinator.

What this exchange suggests: the most productive early interactions with such a system don’t come from users who were briefed on it. They come from users who encounter it with domain expertise, probe it critically, and find it substantive enough to keep engaging.

The best outcome of a pilot isn’t enthusiastic adoption. It’s a faculty member proposing to study it.

Postscript, August 2026: the critique became the roadmap

Nearly four months on, it is worth recording what happened to each thread Professor Ocasio pulled that night. His architecture review turned out to be a to-do list.

“Do we need a retrieval categorization system? Any hierarchy?” → Shipped. A seven-category topic taxonomy tags every decision entry, and the ad-hoc conventions of April hardened into an explicit governance registry: 55 named rules, each with an enforcement tier (from convention to CI-enforced gate) and an implementation pointer. The registry is public and browsable at How K-ai Is Governed, with the conceptual model at How the Knowledge Base Works and the full rule table at the KB Governance Registry.

“Who verifies the accuracy of your responses?” → Partially structural now, not just “Vishal reviews the log.” New decision entries pass through an automated conflict-check gate before they can be committed; a schema check runs in the deploy pipeline and blocks structural drift; a weekly lint sweeps the knowledge base for contradictions and stale items and files a report. Human review remains the backstop — but the first line of defense is machinery.

“Can there be too much information in the memory?” → The memory nearly tripled (roughly 1,200 to 3,200 lines; 59 to 195 logged decisions), which made the April concern concrete and forced the governance work above.

The ego network → Grown from 25 to 31 authorized stakeholders. Boundary changes remain admin-only, by auditable commit. Every substantive change to the knowledge base is now published on the site changelog.

“What about a research agent… using your data as a research site.” → That question became a paper. Professor Ocasio, Geoff Love, and Vishal Sachdev are co-authoring a study of AI-mediated attentional orchestration in new program development for the JPIM Research Forum 2026, with this system as the field site. A weekly automated digest now reports the system’s interaction patterns to the two co-author observers — the research instrument Professor Ocasio proposed at 10:26pm, running in production.

The transcript above is untouched. That is the point of keeping a versioned record: the system that answered those questions in April can be held to account by the system that exists in August.

K-ai is the AI coordinator for the MSBAi (Online Master of Science in Business Analytics) program at Gies College of Business, UIUC. The program launches Fall 2026. Questions: vishal@illinois.edu.