In one sentence
The interface can be conversational. The institution underneath it must become legible, owned, and composable.
What Project Ardi showed
Ardi tested a narrow slice of this idea: bring academic history, interests, constraints, and course information into one planning conversation. The pilot suggested that students value synthesis more than another source to search. It also showed the limits of a narrow product: data gaps, broken tools, and missing human handoffs immediately become part of the student experience.
Five layers beneath the conversation
The first layer is authoritative content: policies, services, requirements, and instructions with owners and review dates. The second is operational data: availability, eligibility, appointments, records, and status. The third is identity and consent: who the student is, which context may be used, and which information may cross a boundary. The fourth is orchestration: how a question becomes a workflow. The fifth is experience: how the system explains, asks, remembers, and hands off.
Most institutional AI efforts begin at layer five because it is visible. Reliability is determined by the four layers underneath. A polished conversation cannot compensate for an unowned policy, contradictory data, or a referral that leads to a dead end.
Capabilities worth building first
Begin where the institution already has authority and the student faces fragmentation: resource discovery, appointment preparation, requirement explanation, form and policy navigation, and status-aware next steps. Keep the system read-oriented before granting it power to submit, enroll, approve, or change a record.
Every response should preserve provenance. The student should be able to open the source, see its owner, know when it was reviewed, and understand whether the answer is general guidance or an official determination. That visible chain of authority is a feature, not clutter.
- A shared resource registry with owners and review dates.
- A policy and requirement layer with version history.
- Consent-aware access to student context.
- Workflow APIs that return status, not just prose.
- Institution-wide escalation and correction channels.
A strategy for the next three decisions
First, choose one cross-silo student journey and map every source, owner, delay, and handoff. Second, establish the content and data contracts that make the journey machine-usable without replacing the systems of record. Third, pilot the conversational layer with inspection, correction, and closure built in.
The long-term advantage is not owning a particular model. Models will change. The durable asset is an institution that can express what it knows, who owns it, how current it is, and what action follows—while giving students meaningful control over how their context is used.
Decision framework
Five questions to take into the room.
- 01
Map a complete student journey across organizational boundaries.
- 02
Assign owners and freshness contracts to the underlying knowledge.
- 03
Separate systems of record from the conversational orchestration layer.
- 04
Make provenance, permissions, and handoffs visible in the experience.
- 05
Treat model choice as replaceable and institutional legibility as durable.
Questions this note answers
- What does AI-native mean for a university?
- Should a university build one chatbot or many assistants?
- Which data foundations are required before conversational AI?
- What is the durable institutional advantage if models keep changing?
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.