NEWMindMap Digital has acquired Bluetide.co— deepening our data & agentic-AI stack.Read more →
Home · Customer Stories · Gulf Bank
BFSI · Middle East

SME Credit Underwriting at a Gulf Bank — From 11 Days to 36 Hours, Approval Rate Up 28%

DocGenie + Coding Assistant analogues applied to SME credit — automated financial-spreading, covenant-extraction and policy-aligned recommendation drafting.

36 hrs
Underwriting cycle (was 11 days)
20w
Delivery duration
Private Cloud
Deployment
5
Accelerators used
Private CloudGulf Bank — 36 hrs Underwriting cycle (was 11 days)
36 hrs
Avg underwriting cycle
28%
Approval-rate uplift
67%
Analyst time per case freed
94%
Memo accuracy vs policy
In this storyCredit UnderwritingBFSIDocGenieArabic NLPPrivate Cloud
01
The challenge

The challenge

The bank — a Tier-1 Gulf commercial bank with an SME book of approximately fifteen thousand active borrowers — was losing SME originations to competitors whose underwriting cycle was structurally faster. The bank's average underwriting cycle from complete application to credit decision was 11 working days, against a market median of 5 to 7 days and a digital-first competitor advertising 48 hours.

The bottleneck was not the credit decision itself but the work that preceded it: spreading the applicant's financial statements into the bank's standard template, extracting the relevant covenants from existing facility documents, pulling the applicant's transaction-banking history from the bank's own systems, looking up the applicant's exposures across the group, drafting the credit memo, routing it for committee review, and assembling the supporting documentation pack. A typical case absorbed 14 to 18 hours of credit-analyst time before the credit committee saw the memo.

The bank's CRO had set a target of reducing the cycle to under 48 hours without weakening underwriting standards. The credit team's leadership was sceptical — they had seen previous automation attempts that had degraded the quality of credit decisions, and the CRO had been explicit that any approach that increased portfolio risk was a non-starter.

02
The approach

The approach

MindMap deployed DocGenie as the document-intelligence layer, with Compliance Engine (Ce) for policy enforcement, Workflow Automator (Wa) for the credit-committee routing, and Summarization Wizard (Sw) for the credit-memo drafting. The accelerators were composed into an SME-credit-specific workflow — branded internally as the bank's own — that credit analysts used as their primary case workspace.

Phase one was the financial-spreading rebuild. DocGenie was trained on the bank's standard financial-spreading template and on five years of historical applicant financial statements (audited and unaudited, in Arabic and English, in the various accounting standards the bank's SME applicants use). The model produces the spread financial statement in the bank's template format with line-item-level confidence scores; the credit analyst confirms or corrects each line in a single-screen UI rather than re-entering the data from scratch.

Phase two was the credit-memo drafting layer. The memo template — the document the credit committee actually reads — was decomposed into structured sections (applicant overview, financial analysis, repayment-capacity assessment, security and covenants, risk factors, recommendation). For each section, the system generates a draft using the spread financials, the applicant's transaction-banking history, the bank's policy library and the credit analyst's case notes. The analyst edits and approves the draft section by section, with the final memo carrying full provenance for each generated paragraph.

Phase three was the committee-routing layer. The credit-committee threshold matrix was encoded as Compliance Engine rules; based on the recommended facility amount, applicant risk band, sector exposure and any policy exceptions, the case is auto-routed to the appropriate committee level (relationship manager, credit hub, regional committee, head-office committee) with the required supporting documentation pack auto-assembled.

Accelerators in this engagement

The pre-built building blocks

Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.

Dg

Data Governance

Policy library catalog and document lineage

Dm

DocuMage

Financial-statement spreading and covenant extraction

Ce

Compliance Engine

Credit-policy enforcement and committee-routing rules

Sw

Summarization Wizard

Credit-memo drafting with structured-section provenance

Wa

Workflow Automator

Credit-committee routing and documentation pack assembly

03
The architecture

The architecture

The platform runs on the bank's private cloud tenant inside Azure UAE, with full data residency in-region. The credit data — applicant financials, facility documents, the bank's policy library — is processed and stored entirely within the bank's tenant.

DocGenie's financial-spreading model is a fine-tuned Llama 3.1 70B variant, trained on the bank's historical financial-spread cases. The training corpus is approximately 120,000 spread cases, with the spread template as the supervision signal. Inference is served on the bank's tenant H100 cluster via vLLM, with the model output constrained to a JSON schema that matches the bank's spread-template structure exactly. The line-item-level confidence is computed from the model's token-level log-probabilities aggregated to the field level.

The covenant-extraction layer uses a separate model — a smaller fine-tuned Mistral 7B — trained on the bank's facility-agreement corpus. This model extracts the structured covenants (financial covenants, negative covenants, reporting covenants, events of default) into the bank's facility-management system, where they are tracked through the life of the facility.

Compliance Engine encodes the bank's credit policy as a deterministic rule set — facility-size limits, sector caps, single-name exposure limits, country-risk caps, related-party rules — and applies them to every case as a hard gate before the credit-committee routing. Any policy exception requires explicit waiver from the appropriate committee level.

Integration with the bank's existing systems is via the bank's standard ESB: customer data from the CRM, transaction history from the core banking platform, facility data from the loan-management system, group exposures from the credit-risk warehouse. All integrations are read-only from the underwriting platform's perspective; the final credit decision is written back to the loan-origination system for downstream booking.

The outcomes

The numbers behind the story

36 hrs
Avg underwriting cycle
28%
Approval-rate uplift
67%
Analyst time per case freed
94%
Memo accuracy vs policy

Average underwriting cycle from complete application to credit decision has dropped from 11 working days to 36 hours, well inside the CRO's 48-hour target. The improvement is roughly evenly split between time saved on financial spreading, time saved on memo drafting and time saved on committee scheduling and document pack assembly.

Approval rate has risen 28% — counterintuitive at first, but driven by two factors. First, the previous cycle's length meant that meaningful numbers of applicants withdrew before decision (typically to accept a faster competitor's offer); the faster cycle has reduced this attrition. Second, the structured memo format has caught categories of declinable risk that the previous narrative-heavy memos had not surfaced consistently, allowing the credit committee to approve borderline cases with appropriate conditions rather than declining them outright.

Portfolio quality has not degraded. Twelve months post-go-live, the new-origination cohort's 90-day default rate is statistically indistinguishable from the historical cohort's rate. The CRO's audit team has reviewed a sample of 240 approved cases and confirmed that the model-generated memo sections accurately reflected the underlying data in every case sampled.

Credit-analyst capacity has been redirected. Headcount has not been reduced; instead, the freed time per case (analysts now spend roughly five hours on cases that previously took fifteen) has been redirected to portfolio monitoring, early-warning case work, and relationship-banking work on the bank's higher-value SME clients.

Eleven days to underwrite an SME case in 2025 was a structural problem we knew we had to solve. MindMap delivered thirty-six hours in twenty weeks, without weakening a single underwriting standard. Our credit analysts are now doing the genuinely judgement-intensive work they were trained for instead of re-typing financial statements.
Chief Risk Officer· Gulf Tier-1 Bank
04
Why MindMap was chosen

Why MindMap was chosen

The bank had previously evaluated three credit-automation specialists. Two were US-headquartered and built on US-credit-bureau-style data assumptions that did not transfer cleanly to the Gulf SME context. The third was a regional vendor whose approach was rules-based rather than model-based and did not handle the Arabic-language financial statements meaningfully.

MindMap won on three counts. First, the pre-built DocGenie accelerator was already handling Arabic-language financial documents at three other regional banks, and we could demonstrate field-level accuracy on the bank's own sample documents during the bid. Second, the willingness to deploy entirely inside the bank's Azure UAE tenant — including the model fine-tuning, not just the inference — was unique among the bidders. Third, the embedded credit-domain expertise on our delivery team (two former credit officers from peer Gulf banks) meant the bank's credit-policy committee felt the platform was being designed by people who understood credit, not just by ML engineers.

Pricing was structured around milestone delivery with a meaningful component tied to the post-go-live underwriting-cycle reduction. The CRO valued this commercial alignment significantly.

Want an outcome like this?

Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.

Book a walkthrough →Explore BFSI
Talk to the product team