Building AI into products, with the costs and risks kept in check
Two of our own products in development: help for schools planning support for children with special needs, and an AI review panel for novelists' manuscripts.
- ~10,000automated tests on the schools platform
- 6AI reviewers on the authors' panel
The client
More Input's own products, in development.
The challenge
AI can draft in seconds what takes people hours, but it makes mistakes, its costs can climb without warning, and the documents it reads are often sensitive or untrustworthy. Our rule is that AI does the drafting, a person makes the decision, and the system keeps an eye on the bill.
For schools: planning support for children with special needs (in progress)
A platform for special educational needs coordinators, teachers and support staff. It began in 2025 as a prototype that drafted support plans from a child's care plan, and is now a full system:
- A profile for every pupil with additional needs, and a map of the support each one receives and what it costs
- Plans with clear targets, with AI-suggested wording that staff review and edit
- Incoming documents read and filed against the right pupil, with staff checking the result
- Annual reviews, views from parents and outside professionals, and evidence for leadership and inspections
- Pupil details kept in step with the school's management system, and a portal for parents
- Built to current accessibility standards
For authors: a virtual review panel (in progress)
An author uploads a manuscript. The system breaks it into scenes, builds a record of the characters, places, plot and timeline, then a panel of AI reviewers reports on it: a developmental editor, a beta reader, a copy editor, a continuity checker, a sensitivity reader and a market analyst. When the author revises, only the changed parts are looked at again.
Keeping AI sensible
- The cost of a job is estimated before it runs, and spending limits stop work automatically
- If one AI provider is unavailable, another takes over
- Text inside an uploaded document can't give the AI instructions
- People approve anything that goes into a child's plan
- Clear rules for how long data is kept, and for deleting it
Technical details
Both products are .NET 10 with MongoDB. The schools platform is a modular Blazor Server application (platform, SEND and quiz modules, each with its own domain, application, infrastructure and web layers), multi-tenant, with MFA, an audit log and a GDPR record of processing. About 10,000 automated tests, plus Playwright, axe-core (WCAG 2.2 AA) and Stryker mutation testing.
AI calls go through a provider-agnostic service: Claude via Azure AI Foundry first, Azure OpenAI as failover, with per-tenant usage and cost caps. The manuscript engine batches extraction in parallel, uses prompt caching with a shared reviewer prefix, anchors quotes back to the text by fuzzy matching, and has an evaluation harness for prompt quality and a persisted spend fuse.
Other case studies
Catching mistakes before a person has to
How we've automated every stage of building software, from requirements to the live system, so fewer mistakes reach a developer and development costs less.
The toolkit behind our systems
Our own building blocks, refined over twenty years: data handling, reporting, courier labels, marketplace links and box packing, so each new system starts ahead.