Knowledge Management with AI: 3 Real Cases for 2025

In June 2015 we published how Berlitz was breaking down its regional silos with IBM WebSphere Portal and Lotus Connections. Ten years later, the problem has not gone away: it has mutated. Educational and training organizations are still drowning in information. Now the volume is measured in terabytes per day. Moreover, teams are hybrid by default and the expectation of an immediate response is non-negotiable. This is exactly the ground where BAOSS IT consulting works every day with educational institutions and corporate universities.

At BAOSS we have spent 2024-2025 deploying knowledge architectures. They are based on RAG (Retrieval-Augmented Generation), autonomous agents and local models on your own infrastructure. We don’t sell «chatbots». We deliver systems that reason over your documentation, connect invisible dots and return actionable answers in milliseconds.

The real problem in 2025-2026: cognitive fragmentation at scale

The talent analytics consultancies (Degreed, LinkedIn Learning, Fosway) agree on one figure. The knowledge worker spends 3.2 hours a day searching for information that already exists in their organization. (Source: State of Knowledge Work 2024, n=4,200.) In the education sector —universities, business schools, corporate universities— the cost skyrockets. Curricula are outdated and content is duplicated. Moreover, onboarding adjunct faculty takes weeks. There is also zero traceability of which methodologies actually work.

1. Technical silos: Moodle, Canvas, Teams, SharePoint, Google Drive, Notion, Confluence… every department picks its own tool.
2. Semantic silos: A «leadership module» in HR. Has nothing to do with the «leadership module» in the internal business school.
3. Human silos: The expert who knew how to solve the edge case left six months ago. Their knowledge lives in Slack threads, Zoom recordings and untagged PDFs.

The 2015 solution was a centralized portal. The 2025 solution is a smart semantic layer that orchestrates heterogeneous sources without migrating terabytes. And that learns with every query.

BAOSS IT consulting’s reference architecture: multimodal RAG + specialized agents

Our standard stack for education/training clients (validated across 7 production deployments in 2024-2025):

• Ingestion: Native connectors (Moodle REST API, Microsoft 365 Graph API, Google Workspace, S3/MinIO, WebDAV). Plus, LlamaIndex for multimodal parsing (PDF, SCORM, video with Whisper-large-v3, audio, H5P).
• Semantic chunking: A hybrid strategy —recursive character for text, proposition-based for pedagogical content (objectives, competencies, evidence). Plus, sentence-transformers/paraphrase-multilingual-mpnet-base-v2 for multilingual embeddings (es/en/pt/fr/de).
• Vector store: Qdrant on a Kubernetes cluster (on-prem or sovereign cloud) with metadata filtering: tenant, program, competency, version, author, last_updated.
• RAG pipeline: LangGraph orchestrating retrieve → rerank (Cohere Rerank 3.5 / bge-reranker-v2-m3) → generate. 128k-token context window (GPT-4o / Claude 3.5 Sonnet / Llama-3.1-70B-Instruct via vLLM or Ollama depending on data policy).
• Specialized agents (CrewAI / AutoGen): CuratorAgent (detects outdated/duplicate content), GapAgent (identifies competencies with no documentation coverage). Plus, OnboardingAgent (generates personalized paths for new faculty). Finally, QualityAgent (evaluates pedagogical alignment against institutional rubrics).
• Observability: LangSmith + OpenTelemetry → Grafana. We measure P95 latency, token usage and hallucination rate (weekly eval set). We also measure user satisfaction (thumbs up/down + quarterly NPS).
• Governance: MCP (Model Context Protocol) to expose secure tools (search_moodle, create_resource, notify_owner) with RBAC and full auditing.

Typical deployment: 6-8 weeks to time-to-value (first pilot in one program/master’s), 12-16 weeks for a full organizational rollout. Infrastructure cost: €0.12-0.18/user/month (shared GPU) vs €2.5-4/user/month for the equivalent SaaS.

Case 1: International business school — Adjunct faculty onboarding in 48h (previously 3 weeks)

Context: 1,200 adjunct professors/year, 18 campuses, 5 languages. Each new instructor received an 80-page PDF, access to 4 platforms and 2 welcome Zoom sessions. Average time to «first autonomous class»: 21 days. Recurring complaints: «I can’t find the grading rubric», «the Madrid campus syllabus differs from Mexico’s».

BAOSS intervention (Q1 2025):

• Ingestion of 47,000 resources (syllabi, rubrics, recordings, policies, FAQs) in 3 weeks.
• Deployment of OnboardingAgent + CuratorAgent with a knowledge base versioned by campus/language/program.
• Interface: widget embedded in Moodle + Teams bot + offline PWA for campuses with limited connectivity.
• Automated evaluation of teaching competencies (institutional framework) via QualityAgent against uploaded evidence.

Metrics at 6 months (July 2025):

• Time to first autonomous class: 1.8 days (a 91% reduction).
• Queries resolved without escalating to the coordinator: 87% (previously 34%).
• Detection of outdated/duplicate content: 2,340 resources (12% of the corpus) flagged by CuratorAgent; 1,890 updated/removed.
• Faculty satisfaction (NPS): +42 (previously -18).
• Infrastructure cost: €0.14/user/month (shared A10G GPU, 3 regional clusters).
• Estimated ROI: 4.7x (coordinator hours saved + fewer grading errors + faculty retention).

«The agent told us in seconds which rubric to use for the Sustainable Finance module in Bogotá. Before, it took three emails and two days.» — Academic Director, LatAm campus.

Case 2: Public university — Cross-curricular alignment in 72h (previously 6 months)

Context: 32,000 students, 2,100 faculty, 85 degree programs. The ANECA accreditation process required demonstrating alignment: degree competencies → learning outcomes → activities → evidence. Documentation scattered across 85 course guides (Word/PDF), 120 Moodle courses, 400+ historical syllabi. Quality team: 3 people. Deadline: 4 months.

BAOSS intervention (Q3 2024):

• Bulk ingestion pipeline: OCR + layout analysis (Nougat) for scanned PDFs + Moodle API + SharePoint Graph.
• GapAgent trained on the MECD + ANECA competency framework: extracts competency, learning_outcome, activity, evidence, assessment_criterion from each document.
• Knowledge graph (Neo4j) connecting entities: Degree → Competency → Course → LO → Activity → Evidence.
• Interactive dashboard for the vice-rectorate: coverage heatmaps, gaps, redundancies, conflicting versions.
• Automatic export to ANECA format (XML) + traceable evidence.

Metrics (December 2024):

• Full analysis time: 72 hours (vs 6 months manually).
• Gaps detected: 237 (competencies without coverage in 14 degree programs).
• Redundancies removed: 89 (identical activities in different courses with no coordination).
• Quality team hours: 120h (vs an estimated 2,400h).
• Accreditation granted without conditions (a first in 12 years).
• Project cost: €48k (vs an estimated €180k for traditional consulting).

«For the first time we see the curriculum as a living system, not as 85 airtight documents.» — Vice-Rector for Quality.

Case 3: Multinational corporate university — Preserving critical expert knowledge

Context: 45,000 employees, energy sector. 127 experts over 55 (retiring 2025-2028). Critical knowledge: predictive maintenance procedures, nuclear incident management, regulator negotiations. Documented in: 3,400 technical reports, 890 meeting recordings, 12,000 Teams/Slack threads, 450 outdated Confluence wikis. Zero semantic structure.

BAOSS intervention (Q4 2024 – ongoing Q2 2025):

• Ingestion prioritized by criticality (risk × expert × available documentation matrix).
• CuratorAgent + QualityAgent specialized in sector terminology (Llama-3.1-8B-Instruct fine-tuned on 50k domain pairs).
• Tacit knowledge extraction agents: AI-guided structured interviews (voice → text → graph). Then, expert validation in the loop.
• Decision digital twins: DecisionAgent reproduces expert reasoning in simulated scenarios (A/B validated against real experts).
• Air-gapped deployment (classified): Ollama + vLLM on own infrastructure, with no dependency on external APIs.

The lesson from these three cases is the same as in 2015. The difference: the tools are now radically more powerful. In short, knowledge only creates value when it flows. RAG, autonomous agents and local models now make it possible to turn scattered documentation into actionable answers. All of it with traceability and without sacrificing confidentiality.

Want your organization to stop searching and start knowing? At BAOSS we design and deploy AI knowledge management architectures on your own infrastructure. Let’s talk.

If you want to see the framework underpinning these deployments, start with what BAOSS is. And remember: disruption gives no warning. This analysis of AI agents and disruption explains why the training sector is no exception. In addition, this guide explains how to control agents already deployed.

Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.