From Pilots to Institutional Capability: Generative AI Agents

The project aimed to turn isolated generative AI initiatives at Uniube, a non-profit university in Brazil with over 60,000 students and a university hospital, into a single, governed institutional capability.

Its goals were:

1. Reduce manual workload and processing time in academic, administrative, financial and HR processes, such as essay and open-question grading, student service triage, résumé screening and document handling.

2. Integrate the AI agents into the institution's core systems (virtual learning environment, academic management, student service and digital registry). 

3. Establish a common governance model, with safeguards against prompt injection, interaction auditing and regulatory compliance. 

4. Extend the approach to the university hospital. 

5. Measure results against baselines to support investment decisions.

Cumulative results since each agent's launch. Some agents started before October 2025, and the expansion phase ran through October 2026. 

• Student service: 124,900 requests classified by AI, with 96% accuracy on the first attempt (baseline: 20,488 requests in the period before AI). 

• Essay grading: 79,280 essays graded, saving BRL 396,000. Open-ended questions: 30,656 graded; analysis time fell from about 40 to about 4 minutes, saving 1,660 working days. 

• HR: résumé submissions rose 68.9% (from 190 to 321) after the selection process was automated. 

• Accessibility: an in-house agent replaced paid third-party tools, saving BRL 49,177.52 per year. 

• Hospital: a first-level support agent is being rolled out.

Software: generative AI agents based on large language models (LLMs) accessed through cloud services, each specialized by domain (academic, student services, HR, finance, accessibility and health); agents developed in-house by the university's IT team and orchestrated through the cloud provider's managed services; guardrails against prompt injection and manipulation, interaction audit trails and human approval for higher-risk decisions. Integrations (via API): virtual learning environment (AVA), academic management system (SIGA), student service system (SAE), digital registry, corporate university platform and institutional website; WhatsApp-based service platform at the university hospital, where a first-level support agent is being rolled out. Infrastructure: cloud-based environment and existing institutional servers and network, with no dedicated hardware purchased for the agents.

The complexity came from scale, integration and governance, not from any single technology. Breadth: agents for academic, student service, HR, finance, accessibility and health processes, each with its own rules, users and data. Integration: the agents connect through APIs to several core systems (virtual learning environment, academic management, student service, digital registry, corporate university and institutional website). Each system has its own data model and owners. Governance: a common model had to be defined for all agents, covering protection against prompt injection and manipulation, audit trails, human approval for higher-risk decisions and compliance with data protection law. Health sector: the hospital required review of specific medical regulation on AI and stricter regulatory requirements. In-house delivery: the agents were designed, built and operated by the university's own IT team, which carries the full responsibility for quality and continuity.

INITIAL SITUATION: Uniube is a non-profit university with over 60,000 students, in-person and distance, and a university hospital. Before this phase, generative AI existed as isolated tools, each built for a single task such as essay grading. Many academic, administrative, financial and HR processes were still manual and labor-intensive: grading essays and open-ended questions, classifying student service requests, screening résumés, handling documents and monitoring funding calls. The result was high operating cost, slow response times and uneven quality across teams. There was no common governance model for AI.

IMPLEMENTED SOLUTION: Starting in October 2025, the IT team consolidated these tools into a single portfolio of domain-specific generative AI agents, built in-house and run on cloud services. The agents were integrated through APIs into the core systems: virtual learning environment, academic management, student service, digital registry, corporate university and the institutional website. New agents were added for areas such as HR, document screening and accessibility. The approach was extended to the university hospital, where a first-level support agent is being rolled out on its WhatsApp service platform. 

IMPLEMENTATION SPECIFICS: The main decision was to treat governance as a requirement from the start. All agents share the same safeguards: protection against prompt injection and manipulation, audit trails of interactions and human approval for higher-risk decisions. The institution's first AI policy was drafted and approved by the executive committee, and health-sector regulation was reviewed for the hospital. Building in-house kept each agent close to the real process and replaced paid third-party tools in some cases. The most demanding part was the hospital, where clinical and regulatory requirements are stricter. 

CHANGES AFTER IMPLEMENTATION: Routine tasks that used to take staff hours, such as grading and request triage, are now handled by agents, with people reviewing results and handling exceptions. AI is now managed as an institutional capability, and IT was elevated to executive directorate level.

The project was implemented in Brazil, in the state of Minas Gerais, with the university and its university hospital based in the city of Uberaba. The agents serve students, faculty and staff of the institution, which has over 60,000 students in in-person and distance learning programs. Through distance learning, the services reach students located across Brazil. The hospital agent is being rolled out for patients and staff of the university hospital.