The real bottleneck is not the decision, it is the reading
On every credit desk the sequence is familiar: a request comes in, the credit file is opened, and the analyst reads the loan agreements, the pledge documents, the guarantees, the covenants, the latest financial statements. Then the credit memo is written for the committee. The part that consumes the hours is not the analyst's judgment, which remains human and irreplaceable, but the line by line reading of hundreds of heterogeneous pages: an English term sheet, a loan agreement drafted under Italian practice, a consolidated IFRS statement, a local OIC half-year report from a subsidiary, a property appraisal in a scanned PDF.
A commercial bank underwriting corporate exposures to SMEs listed on Euronext Growth Milan or to DAX mid-cap groups handles documentation that shifts by jurisdiction, language and accounting standard. The risk is not only slowness. It is omission: a financial covenant buried in a schedule, a guarantee renewal date, a cross-default clause that ties the exposure to another lender. Amaril exists to remove this reading load without removing control from the analyst.
What the AI extracts from the credit file
Amaril reads the full document perimeter of the credit file and structures it in a verifiable way. On each loan agreement it identifies amount, tenor, amortisation schedule, rate and indexation mechanism (Euribor plus spread), real and personal security, events of default and mandatory prepayment clauses. On covenants it separates financial covenants (net leverage, DSCR, interest cover, gearing) from information and negative covenants, reports their thresholds and test dates, and links them to the most recent financial data to flag proximity to a breach.
The point that changes the daily work is traceability. Every figure, every clause, every deadline Amaril produces carries a source citation: document, page, paragraph. The analyst does not receive a statement to trust, they receive a precise reference they can open and verify in one click. This is the pillar that separates a tool usable inside a credit process from a generalist assistant: Amaril is built not to hallucinate and not to answer when the source is absent. If a piece of information is not in the file, it says so rather than inventing it.
On deadlines, Amaril builds a calendar of obligations: covenant test dates, security renewals, reporting milestones to the bank, tranche drawdown conditions. On a structured exposure, missing one of these dates can mean losing a contractual right or delaying a waiver.
From reading to a draft for committee
The final output is not a list of extracts, it is a draft credit memo. Amaril assembles the sections the credit committee expects: borrower and group profile, deal structure, security, covenant analysis with headroom against thresholds, risk factors, critical deadlines. The draft reaches the analyst already cited: every material assertion points to the primary source, so the review becomes validation and refinement, not reconstruction from scratch.
A concrete example. An industrial group with a subsidiary reporting under OIC and an IFRS parent requests an increase to a revolving facility. Amaril reconciles the two accounting perimeters, recomputes consolidated net leverage against the agreed covenant, flags that the half-year DSCR test falls in six weeks and that a bank guarantee securing the facility expires before the next test, and highlights a cross-default clause toward an existing syndicated loan. All with references to the exact pages. The analyst keeps the judgment, but starts from a base prepared in minutes rather than days.
This also improves the quality of risk governance. A memo built on verifiable citations is more defensible before the committee, before internal audit and, if requested, before the supervisor. In a context where Banca d'Italia, the EBA and the CRR/CRD framework place growing emphasis on the quality of credit granting and monitoring, being able to reconstruct the provenance of every figure in the memo is not a formality, it is compliance substance.
Data security and the European perimeter
A credit file holds the most sensitive data a bank handles: non-public statements, personal guarantees, information on individual guarantors. For this reason Amaril runs on cloud inside the European Union, with end-to-end encryption and a zero retention policy: documents do not feed model training and do not leave the agreed perimeter. For institutions with stricter requirements, on-premise deployment is available, so the file never leaves the bank's infrastructure. Processing is set up for GDPR compliance and fits within the obligation framework that includes DORA on operational resilience and anti-money-laundering rules (AMLD).
Key takeaway
The value of Amaril is not replacing the credit analyst, it is giving back the time that reading takes away. The AI reads the credit file, loan agreements and covenants, builds the deadline calendar, highlights the risks and delivers a draft memo that is already cited and verifiable. The three pillars hold firm: a source citation on every statement, no hallucinations and no invented answers, with data living on EU cloud, in zero retention or on-premise, GDPR compliant. The committee decides faster because it receives material that is prepared, traceable and defensible.
