AI advises. Humans decide.
The MCC programme uses AI throughout the applicant journey — and governs it under an adopted, published standard: MCC-850 AI Governance. The rules below are not policy aspirations; each one names the mechanism in the platform that enforces it. If a capability is not on this page, the programme does not claim it.
Advisory analysis, at every step an applicant takes
Methodology guidance
In the application wizard, the model explains which credit family fits a described project. Register facts — which methodologies are actually adopted — are stated by the server from the register, never by the model.
PDD structuring
The PDD builder structures the applicant’s own notes into the template. Where information is missing it writes an explicit [INFORMATION NEEDED] marker instead of inventing content.
Draft precheck
The moment a draft PDD is attached, a preliminary completeness analysis runs automatically and appears on the applicant’s project page. Informal and advisory; the formal check runs fresh at submission.
Submission completeness review
Every submission gets a section-by-section completeness analysis with risk flags. The applicant receives the score and the weakest sections by email, immediately, so they can act without waiting for anyone.
Escalation and prioritisation
Confidence tiers are governed, not self-reported: each assessment carries an escalation level and a human review deadline. High-scoring, unflagged submissions are marked as fast-track candidates — for prioritised human review, never for automatic approval.
Enforced in architecture, not promised in prose
Every project decision goes through one server route reserved for human decision roles, behind multi-factor step-up authentication. The AI service account holds no decision permission at the database level.
MCC-850 Art. 27: the applicant-facing surface carries the completeness analysis only. The recommendation field is never rendered to applicants — so no one is anchored by the machine before a human has judged.
The governance review panel shows the application and its record — not an AI score or recommendation. Board members judge the material.
MCC-850 Art. 24: when a human decision contradicts the AI assessment, it is recorded as an override with the decision-maker’s rationale, in the permanent decision record.
The review engine only analyses. Illegal state transitions are additionally rejected by a database-level guard — defence in depth, not policy on paper.
Every assessment leaves a permanent record
Each AI assessment is stored with the fields below, and every public project page carries an AI-assistance disclosure stating what ran and when.
| Recorded | What it pins down |
|---|---|
| Model identity | Model ID and version for every assessment, including when a fallback ran |
| Prompt template | Versioned template ID, separated by trigger (precheck, submission, staff request) |
| Input hash | SHA-256 over the project data and document set the assessment read — any change to the input changes the hash |
| Confidence tier | Governed tier with provenance: what the model reported and what the governance rules assigned |
| Escalation | Escalation level, path and the human review deadline the tier requires |
| Timing | Assessment timestamp and processing time |
Degrades honestly, never silently
If the AI provider cannot be reached, the completeness check runs a limited heuristic instead — and says so, on the applicant’s page, in as many words. Applicant tools that cannot answer reliably answer with an honest error rather than a guess. Human review proceeds unaffected either way: the programme’s decisions never depended on the model being up.
Read the rules the machine runs under
MCC-850 AI Governance v1.2 is the served, in-force text of the AI Governance standard, published in full on the document register with its amendment history. No Board-approval instrument for this version designation exists in the record (Article 43 requires Board approval to take effect).