Meet Penny: AI that drafts and checks while the adviser decides

01The challenge

Penny House had already put AI to work. Claude skills read client documents, checked case files and drafted case paperwork; ChatGPT scheduled tasks watched the inbox, sent reminders and produced reports; a lead-control app had been built in Lovable. But the adviser was still the integration layer. Information moved between email, case files, the CRM, sourcing tools, lender portals, spreadsheets and AI projects by re-keying, and no tool owned the case: its stage, owner, next action, deadline or approval history. Case progression, the stage that takes the most admin time, had no tool at all, and every change had to respect the controls of the principal firm the brokerage advises under.

02What we did

1. 01Handover — the brokerage's Claude skills, ChatGPT scheduled tasks, lead-control app and compliance checklists, mapped against a single case lifecycle from first enquiry to renewal
2. 02Phase 0, discovery and controls — demonstrate every existing skill, schedule and app, and agree with the adviser and the principal firm what the system may read, draft, write, send, move, delete and submit
3. 03Phase 1, control layer — one lead, case and task model with unique IDs, an owner, next action and deadline on every case, an audit log and an exception queue
4. 04Phase 2, intake and file control — match email to cases, create case folders from the file model, and run the existing document intake and file checks as orchestrated modules
5. 05Phase 3, routine-case pilots — product transfers and straightforward remortgages run by the administrator, reaching the adviser only at defined gates with complete evidence

03What we built

The design rests on one rule. If a job needs durable state, timing, permissions, integration, retry, audit or a guaranteed outcome, it belongs in software. If it interprets unstructured evidence or drafts language, AI can assist. If it is regulated judgement or final accountability, it stays human.

So the workflow engine, not the model, decides when AI runs, what it may touch and where its output goes. Each run gets a bounded task, an approved snapshot of the facts and an output schema, and returns its evidence, confidence and exceptions. The adviser makes every recommendation and signs off every suitability letter, and a person always clicks Submit on a lender portal.

The brokerage's method moves out of chat prompts into a version-controlled library, so Claude and ChatGPT become replaceable workers, shadow-tested against each other before either becomes the default for a task.

04The result

Meet Penny is in build, control first. The first goal is that every new enquiry and pilot case is visible, owned and auditable. Each workflow then starts in shadow mode, with AI proposing while a person acts, and moves through testing, the adviser's approval, limited production and monitored scale, with a rollback and a manual fallback. Success is measured as revenue per adviser hour, alongside time saved, re-keying removed and approval exceptions.