What your AI Credit Controller learnt this week
Every change that affects how the controller drafts, decides, escalates, or asks for your approval lands here. New capabilities ship behind workspace-level flags first; nothing changes without your explicit opt-in.
- Transparency
Persona consistency tests visible to super-admins
Every assertion that guards Ruby (firm), Mason (friendly) and Alex (balanced) tone drift is now visible in Admin → Persona Tests. Read-only view of the regression suite.
- Guardrail
Statutory interest correctness pinned with tests
UK Late Payment Act 1998 maths (BoE base + 8%, 365-day daily interest, s5A fixed compensation tiers) extracted to a single library and locked behind a unit-test suite.
- Guardrailflag · sector_flag · default off
Sector flag scaffolding (healthcare, education, vulnerable)
Workspaces can now record a sector hint. The self-critique skill blocks any draft that is harsher than 'balanced' for sensitive sectors. Default off; UI picker pending.
- Transparency
Dry-run preview before any action approval
Every pending agent action exposes a Preview button that renders the full payload exactly as the AI proposed it — no execution until you click Approve.
- Transparency
Per-debtor agent timeline
Admin → Agent Timeline merges tool calls, reasoning traces, and pending actions into a single chronological feed per workspace.
- Skillflag · feature_flag_skill_self_critique · default off
Self-critique skill (default off)
When enabled, every drafted communication is reviewed against banned phrases, tone vs. stage rules, and SMS/WhatsApp length caps before it can become a pending action.
- Skillflag · feature_flag_skill_outcome_summary · default off
Outcome summary skill (default off)
Optional end-of-cycle summariser that produces a short headline and 'what changed' narrative when a recovery completes.
- Capabilityflag · feature_flag_inbound_email_parse · default off
Inbound email triage (advisory only)
Replies to chase emails can be classified as promise / dispute / partial / OOO / unsubscribe / abuse and routed to the approval inbox. Never auto-executed.
- Guardrail
Approval inbox + reversible writes
All AI write actions (activate/pause recovery, create plan/settlement/penalty, settings updates) flow through pending_agent_actions with one-click approve, reject, or reverse.
- Guardrail
Autonomy slider (Observe → Suggest → Act-with-confirm → Auto)
Per-workspace autonomy tier with hard caps (£ per action, debtors per hour, no LBA without admin). 'Auto' is hidden behind a super-admin flag; default remains Act-with-confirm.