How Universidad Panamericana (UP) Consults Student NPS with a Conversational Assistant

Cliente: Universidad Panamericana (Aguascalientes and Guadalajara campuses)

Socio: INMEGA Market Research

Tecnología: Claude (Anthropic)

The student NPS study that INMEGA conducts for Panamericana University has accumulated 5,153 responses across two campuses, two levels (undergraduate and graduate), 15 schools, and two comparable cycles in 2024 and 2025.

Results were delivered in a Power BI dashboard. It worked for viewing metrics, but had the usual limitation of any dashboard: it answers only the questions anticipated when building it. When a school director wanted a different cross-tab or an explanation of why a metric moved, the request went through an INMEGA analyst and the answer arrived days later. In practice, this meant the study was thoroughly reviewed at the results presentation and not much after.

INMEGA rebuilt the dashboard as its own web application, outside Power BI, with UP's visual identity: six sections, five cross-filterable filters across all of them (campus, year, level, school, and program), and full mobile functionality. On top is the "Crystal Ball (Bola de Kristal)", a conversational assistant built on Claude and embedded in the dashboard itself, where users ask natural language questions about the results.

The relevant design decision is the scope of context. The Crystal Ball doesn't query the complete database: it receives only the slice corresponding to the active filters on screen—that is, the KPIs, satisfaction by aspect, segments, and a sample of real comments from that subset. If the user filtered graduate Engineering in Aguascalientes, the response refers to those 185 students. Before asking, users can expand the exact summary of the data the assistant is receiving to verify where each response comes from.

The dashboard reports a global NPS of 68.7, with 66.5 in 2024 and 71.0 in 2025. Aguascalientes improved by 6.4 points and Guadalajara by 3.4. The distribution is 74.8% promoters, 19.0% passives, and 6.2% detractors, with a reinstatement probability of 8.4 out of 10.

On that basis, the redesign made three things queryable without intermediaries that previously required an analyst. First is priority by impact rather than rating: the findings module orders schools by combining their gap against the global average with their weight in the sample, so Engineering, with an NPS of 58.1 and 731 responses, moves the institutional indicator more than other areas with fewer responses. Second is the gap by aspect between detractors and promoters: average satisfaction places the student environment as the best-rated aspect (4.52 out of 5) and tuition as the lowest (3.45), but the actionable insight is the range—in administrative processes, detractors rate 2.85 versus 4.23 for promoters, a gap comparable to tuition (2.32 versus 3.66), with the difference that processes can be improved without changing the pricing model. Third are specific profiles rather than general alerts: the cross of campus, level, and school with a minimum of 10 responses identifies eight concrete groups.

The place of verbatims also changed. The study recovered 303 open comments and 500 accounts of extraordinary moments, with 57.4% of students reporting having experienced a positive one. Previously these were an appendix; now they are filterable text by student type, and the assistant cites them when explaining a metric.

The choice of Claude had two reasons. First is the handling of open text in Spanish and the ability to reason over structured data and qualitative comments in the same response, the combination that drives an NPS study. Second is that UP needed to be able to answer for the tool before putting it in front of its results, which made the discussion with the institution center on context control and answer traceability rather than whether the model was generally trustworthy. The option to see the context the assistant receives came out of that conversation.

The study didn't change: same database, same 5,153 records, same fieldwork. What was reduced is the time to consult information. For INMEGA it's a shift in the form of the deliverable, from a report presented once to a tool consulted when questions arise.

Process Impact:

Before, from survey to final report took 40 days; now, comprehensive information—general and detailed by campus, program, and graduate level is ready in less than 15 days.

Before, a single general report; now, a focused report for each study area with specific recommendations, one version per undergraduate program and one per graduate program per campus, with actionable recommendations.

Before, 7 general users; now, 60 platform users.

Before, a manual comment classification process taking over 120 hours; now completed sequentially in under 8 hours.

Before, manual verbatim coding consumed 90 seconds per mention; now automated classification takes less than 3 seconds per classified mention.

The Crystal Ball (Bola de Kristal) is available as a conversational layer for quantitative and qualitative studies.