Strengthening Families, Strengthening Communities moved online in 2020 and kept running.
Years into that operating model, the programme faced a different kind of pressure. The work had generated a deep institutional memory: programme descriptions, evaluations, policy research, funding history, delivery records. That memory lived across hundreds of documents and a CRM, and the people who needed it most had the least time to search for it.
A staff member writing a grant application needed the official programme description, in the authorised wording, with a source they could cite. A local authority funding delivery needed to see how its own cohorts were progressing. Both needs were being served by hand: people assembling material from systems that did not talk to each other, week after week.
AI made a different answer possible. It also raised the question that matters more than any demonstration: could the organisation rely on it?
What had to be trusted
Two constraints shaped everything.
The first was authority. Organisational knowledge is situated. A document was written by someone, for a purpose, at a moment in time, and it carries a level of authority that a semantically similar paragraph from somewhere else does not. A grant writer quoting the programme is asking for the canonical text. A strategist exploring patterns across evaluations is asking a different kind of question entirely. Retrieval had to respect the difference.
The second was safety. The programme works with families who often arrive through councils, schools, and referral pathways. Reporting on the programme could not be allowed to become a new place where personal information accumulates. Every additional copy of a family’s data is an additional way to fail them.
What was built
Two systems, one principle: the intelligence layer inherits the organisation’s standards. The first is in use. The second is designed and not yet built.
Scroll sideways to see the whole drawing.
The first is Abena, an assistant for the team’s own knowledge. Documents are indexed from the organisation’s shared drive, chunked with their structure and metadata intact, and made searchable two ways at once: full-text search for the moments when exact phrasing matters, and semantic search for the moments when meaning matters more than wording. Answers carry source attribution, so staff can verify where a claim came from before it travels into a grant application or a policy paper. Access is restricted to the organisation’s own accounts, conversations are persistent and searchable, and staff can bring their own documents into a thread when a question needs specific context.
The second is a reporting layer designed to sit on top of the programme’s CRM as a read-only window. A connector will pull identifiers and scores on a schedule, never names, addresses, or contact details, and cache them for fast presentation. Role-based access will let a local authority see its own boroughs and nothing else, with attrition funnels, demographics, and downloadable reports. The design makes the safety property structural: personal data never leaves the CRM, so the reporting layer will hold nothing that could be lost.
What changed
The route to institutional memory shortened. Questions that once meant searching folders, asking colleagues, or reconstructing history from old applications became conversations with sources attached.
Reporting is the next step. Once the reporting layer is built, funders and local authorities will reach their own figures, scoped to what they are entitled to see, without staff assembling them by hand.
The quieter change is organisational. The programme’s infrastructure now includes an AI assistant that is maintained, governed, and owned like the rest of it. It has access rules, update processes, and known limits. It was built to be relied on, and reliance is the standard it is held to.
What it taught
An AI system holds inside an organisation when it inherits the organisation’s standards for authority, safety, and evidence.