The best AI platform for education use cases

A school or training organisation rarely needs just one bot. Admissions questions, student IT support, and staff HR questions are three different audiences with three different knowledge bases — and Botmanor’s bucket-based organisation exists precisely for running several of them side by side under one governed workspace.

The blueprint per audience is the familiar one: an app as the front door, a domain specialist behind it, an escalation agent for handoff to a staff member, and a knowledge base grounded in your own material.

The App → Agent mechanic, applied to a campus

One app as the front door, specialist agents behind it — a general-questions agent alongside a financial-aid or enrolment-specific agent — with an escalation agent for handoff to staff, and routing configuration wiring it together.

The knowledge base is where the content is genuinely yours: course catalogues, admissions policies, financial-aid guides, or student handbooks. Documents you upload are the shipped ingestion path: each is chunked with the strategy you pick, embedded into that knowledge base’s own vector collection, and tracked through its own processing status, so a failed document is visible instead of quietly missing. Site crawling and pull-based connectors are on the roadmap. Each corpus is chunked with the strategy you choose, embedded into its own vector collection, and validated with semantic search before anything goes live.

One workspace, several bots, organised properly

Bucket-based organisation groups related apps and agents hierarchically, agent versioning keeps configuration changes traceable, and workspace RBAC lets multiple staff collaborate on the same bots under role-based permissions — so an admissions office and an IT department can each own their bot without stepping on each other.

Each app can deploy across different channels for different audiences: channel connectors and routes cover Slack and Microsoft Teams for staff-facing bots, plus Discord, WhatsApp, and inbound webhooks. The embeddable web chat widget for a public prospectus site is modelled but not yet shipped.

Grounded in your own materials, tunable per corpus

Retrieval settings are configurable per knowledge base — mode, chunk size and overlap, top-K, and score thresholds — and saved as a reusable retrieval profile you can attach to other agents or distribute through the store, so “the retrieval configuration that works for our admissions FAQs” becomes something you build once and reuse across programmes or campuses. Vector search is the implemented path today; hybrid and BM25 modes are stored configuration ahead of the runtime that consumes them.

One boundary, stated plainly: routing configuration, retrieval settings, MCP tool entries, and policy definitions are managed objects you author and version today — the agent runtime that dispatches on them inside a live conversation, grounds answers in retrieved chunks, calls tools, and hands off to a human is the next milestone. You build the governed layer first, on purpose.

Data handling you can explain to a board

Model selection is per agent and per provider across 11 provider types, including Azure OpenAI and a custom OpenAI-compatible endpoint pointed at a model server you host yourself, so an institution with data-residency requirements chooses where inference runs. Your knowledge bases and conversations are used to serve your agents, not to train foundation models — a distinction that matters when the content touches student records or enrolment data.

Education FAQ