Small budgets, hard obligations, real consequences
AI for charities, associations and foundations, built on small models and cached retrieval so the running cost survives the grant.
Economics
The design constraint is what one interaction costs twelve months after launch, when the pilot funding is gone. Small or open-weight models for most traffic, caching for the questions that recur, short retrieval instead of long prompts.
Services
Membership and eligibility questions, bookings, deadlines and service directories across WhatsApp, web, SMS and the phone line: a two-person office with the reach of a call centre.
Cost
Open weights
Self-hosted small models remove per-token pricing once volume justifies it.
Reach
Beneficiaries without smartphones, data or confident literacy are reached by phone and text. Primary channels here, not accessibility afterthoughts.
Safeguarding
Any public service eventually gets a disclosure it was not designed for. The system hands over to a trained human at once: it never assesses risk, reassures or closes the contact.
Data
Consent is often not the lawful basis available, so minimisation does the work: the assistant sees required fields, never the case record, and volunteer access assumes turnover.
Models
Cost control
Channels
Systems
Compliance
2–3 weeks
Which demand is genuinely automatable, what it costs per interaction and per year, and where the answer is not to build.
Cost forecast, safeguarding and data protection constraints, board-readable recommendation
6–12 weeks
One service built end to end on the low-cost architecture, with safeguarding escalation tested adversarially.
Working service, multilingual evaluation sets, twelve-month running cost forecast
Ongoing
Operation at a rate set against your budget, with periodic safeguarding review and progressive handover to your staff.
Supported operation, safeguarding and content review, trustee and funder reporting
Cost is the primary design constraint, not an afterthought. Caching handles the repetitive traffic, a small or open-weight model handles most of the rest, and retrieval returns short context rather than long prompts. Self-hosting removes per-token pricing where volume justifies it, and you get a figure up front.
It hands over to a trained human immediately and does nothing else: no risk assessment, no reassurance, no closing the conversation. The path is designed with your safeguarding lead, written into the guardrails and tested adversarially before every release as a gate. Staff are told what it does not detect.
When the information is not written down. An assistant answers from your content, so if eligibility criteria live in three people's heads and an outdated PDF, the first output will be confidently wrong in public. Writing the answers down costs a fraction and often resolves most of the demand on its own.
Regulator, budget, deadline, legacy core, a board that has been burned once already. Tell us what you are working around.
