The most senior people in the client organisation regularly want information.

That information is very sensitive.

They want the information immediately or quickly.

They want to analyse the information, compare it against historical data, competitors, and public information. It often leads to streams of Q&A.

Requests from the Board Chair, CEO or owner consume huge amounts of company resources and stress for staff. For three months, our client measured >2,000 company hrs.

Providing this info to the Chair does not directly add value to the employees’ role. Most times they are simply finding and passing on info.

Ideally the senior person could directly access the source data from anywhere in the company.

Using our Consulting approach to fully understand the problem, goals, systems and relevant data.

Define the purpose of the AI app, data sources (internal and external), and the boundaries such as privileged legal materials, HR-sensitive data, information that could affect disclosure obligations.

Develop PoC for stakeholders including Chair and Company Secretary. Use PoC feedback to iterate.

Develop RAG chatbot using a custom private model given sensitive nature of inference.

Document ingestion pipeline to include last 10 years of board packs, financials, risk registers, policies, org charts, all company emails between certain groups, competitor public reports, global macro economic and logistics data from various sources. Use APIs where possible.

Backend uses Python, LangChain, PostgreSQL, Pinecone, Azure, and Okta.

Search and summarisation layer converts all files into embeddings. Assign confidentiality levels.

Access and permissions layer restricts confidential segments to authorised users.

Front end included a custom web app also with a Teams plugin and with the ability to plugin to Diligent Board Portal.

Governance checks re user permissions, data classification, document recency, hallucination guardrails. Model to obtain and summarise information, not add opinion.

Ensure full reasoning with source data references in answers. E.g. “Source: Oct 2024 Audit and Risk Committee Pack – Finance Report, p. 114.”

Using private LLM via Nvidia GPUs on on-prem servers.