Innovation people can work with
Use casesCompany Personalised GPT — a governed multi-agent GenAI assistant
A secure GenAI assistant that lets business teams query documents, data and the web in plain language — hosted entirely in the client's cloud.
Context
The starting point
A pan-European events organiser was sitting on a lot of knowledge — internal documents, warehouse data, and market information — spread across tools and teams. People had started using public chat assistants to get answers faster, but that created an obvious problem: sensitive information could leak, and nobody had visibility or control. They needed one assistant the organisation could trust. It had to work in their context, use their data, and stay inside their cloud.
Challenge
What made this hard
The ask sounded simple: let staff get answers without writing SQL or hunting through folders. The constraints were not. Success meant a single authenticated assistant, hosted in-tenant, giving grounded answers with citations — with governance and audit trails built in from day one.
- Data stays in-tenant: sensitive content could not leave the client's cloud.
- Access must follow existing roles: permissions had to align with Azure Entra ID, not a new shadow model.
- Three languages, many sources: knowledge lived across documents, BigQuery and the public web.
- Trust and auditability: answers needed source attribution, and usage needed to be logged.
Approach
How we worked
We kept it practical: agree the guardrails early, then build in phases so the client team could learn and steer.
- Discovery & alignment
- Workshops with data, security and business stakeholders to map real question patterns and agree the in-tenant by default model up front.
- Design
- A multi-agent setup (LangGraph) with a provider-agnostic model layer, so capabilities and models can be added or swapped without a rebuild — with access control, content safety and logging built in.
- Delivery
- A phased rollout from core document Q&A to a broader platform (shareable agents, reusable prompts, document generation, connectors), shipped through CI/CD across DEV/UAT/PROD.
- Change enablement
- Practical onboarding and prompt patterns so business teams get value from day one, not after a long training curve.
Solution
What we delivered
Company Personalised GPT is a governed, multi-agent GenAI assistant hosted entirely in the client's cloud. It is a platform you can extend, not a one-off chatbot.
- Core experience
- Chainlit chat with Azure Entra ID SSO and role-based access; a LangGraph orchestrator routing queries to specialised agents.
- Capabilities
- Semantic document Q&A, grounded web search with citations, and generation of Word, Excel and slide content from chat.
- Platform building blocks
- Reusable prompt library, user-built knowledge bases, custom shareable agents, and MCP connectors to external systems.
- Foundations (in-tenant)
- PostgreSQL with pgvector storage; provider-agnostic models (Azure OpenAI, Claude, Gemini) with automatic fallback; source attribution, content safety controls, and LangFuse observability.
Outcome
What changed
A sanctioned, in-tenant GenAI capability that people actually use — with governance and audit baked in from the start, not bolted on later.
- Business impact
- A sanctioned, in-tenant alternative to public chat tools, materially reducing sensitive-data leakage risk.
- Operational impact
- Self-service answers across documents, data and web sources from one place — saving an estimated [X hours/week per team] and reducing routine analyst requests.
- Data & AI readiness
- A reusable GenAI platform with governance baked in, so new use cases can ship incrementally instead of starting from scratch each time.
- Now possible
- Asking questions across documents, warehouse data and the web in plain language — with citations, roles and audit trails intact.
“A useful GenAI assistant is not just a smarter search bar. It's a governed platform your organisation can trust, extend and keep in its own cloud.”
Agoya perspective
Getting your data to work
Innovation people can work with
Bring the messy version. We'll help you shape it into clear, safe next steps.