In one sentence
The useful boundary is not “human or AI.” It is routine synthesis for software, institutional judgment and relationship for people.
What Project Ardi showed
Ardi tested a conversational planning experience during a 44-day CU Boulder pilot. Students used the conversation to add constraints, question suggestions, and revise plans. The pilot also exposed a hard boundary: the product had no functioning advisor-referral mechanism, and distress-related conversations did not produce handoffs. Any institutional version must treat escalation as an operating workflow, not reassuring copy.
Appropriate work for an advising assistant
High-volume preparation work is a strong fit: explaining terminology, gathering goals and constraints, showing how requirements connect, comparing scenarios, generating questions for an appointment, and reminding a student which facts still need official confirmation. These tasks can happen before an appointment and can make scarce meeting time more useful.
The system should reveal its sources and reasoning. A student should be able to see which requirement, policy, or preference produced an answer. When records conflict or the source is stale, the correct response is not a smoother paragraph. It is a visible uncertainty and a path to resolution.
Work that should remain human
Exceptions, appeals, financial-aid implications, mental-health concerns, disability accommodations, immigration status, and major life choices require accountable human judgment. So do conversations where the student's stated question is not the real problem. An advisor can notice hesitation, history, and institutional context that a data model will not fully capture.
A handoff must be specific. “Talk to your advisor” is not a workflow. The system should identify the right destination, explain why, show urgency, help the student prepare, and—where policy allows—transfer the context with consent.
- Name the owner of the decision.
- State what the system does not know.
- Preserve the student's words, not only a machine summary.
- Confirm whether the student reached the next step.
The operating model matters more than the chat interface
An institution needs content owners, data stewards, escalation queues, review cycles, incident response, and a way for advisors to correct recurring errors. Without those systems, a conversational interface simply makes fragmented information sound confident.
Advisors should be involved before a pilot begins: in defining safe use cases, writing escalation rules, reviewing sample conversations, and deciding which metrics would actually indicate improvement. Adoption is not a training problem after launch; it is a design input before launch.
Decision framework
Five questions to take into the room.
- 01
List routine preparation tasks that currently consume advisor time.
- 02
Define decisions the system may never make on its own.
- 03
Attach every answer to an owned source and review date.
- 04
Build named, testable escalation paths before student access.
- 05
Measure advisor readiness and student understanding, not deflection alone.
Questions this note answers
- Will AI replace academic advisors?
- Which advising tasks are safe to automate?
- How should an AI advisor handle uncertainty?
- What makes a human handoff operational rather than performative?
About the evidence
Project Ardi was one 44-day product pilot at CU Boulder with no comparison group. It measured conversations and product behavior, not academic outcomes. “2,000+” is Ardvarq's company-reported estimate of student users; 903 is the tracked conversation count. Read the research notes and study limits.