Case study / AI credit decision support / 2025–2026
Fairline AI-assisted SME lending
I designed a connected underwriting workspace and mobile applicant portal. The implemented workflow reduced the cost of a correctly reviewed credit application by 12% and delivered $1.2M in net cost savings over 12 months.
The product name and branding have been changed to protect confidentiality. The case presents original product solutions and the results of their implementation.
01
Overview
Project info
Fairline supports underwriters reviewing SME credit applications and applicants correcting information on mobile. I designed the review queue, a unified dossier that separates verified data, lending policy and AI conclusions, and the mobile correction flow. These changes reduced repeated checks and document requests while keeping final credit decisions with an authorised underwriter.
02
The challenge
Underwriters had to search across systems and recheck the same information to resolve incomplete SME credit applications. Applicants needed a clear way to understand a request, correct a value and submit current evidence. The design challenge was to reduce this repeated work without weakening source checks or human authority.
- Human authority
- AI could recommend and simulate terms. Approval, rejection and the final credit record required an authorised underwriter's confirmation.
- Source integrity
- Data, policy and model claims needed a source and timestamp. The final decision event had to remain immutable.
- Fairness and drift
- A threshold breach had to pause automation, route affected applications to manual review and require signed recovery checks.
03
Why AI
SME underwriting requires comparison across cash flow, policy, security and application evidence. AI can prepare that comparison and suggest an eligible alternative when the product keeps authority and recourse explicit.
Evidence synthesis
The model combines consented financial data and submitted documents in one review.
Policy-aware guidance
It identifies conflicts between the application, lending rules and model recommendation.
Counterfactual terms
It can prepare a policy-eligible alternative while preserving the original request.
04
AI system model
The workflow connects consented data to a sourced recommendation, human decision and applicant explanation.
- 01
Collect
Read consented banking data, application fields and submitted documents.
- 02
Assess
Prepare claims about cash flow, affordability and security with sources and freshness.
- 03
Check
Compare the recommendation with lending policy and fairness controls.
- 04
Decide
Require an authorised underwriter to confirm the outcome and rationale.
- 05
Explain
Show the applicant the review reasons and available correction or review path.
05
Key decisions
-
Separate recommendation from authority
I gave model output, policy checks and the human decision distinct states. The interface never presents a recommendation as an approved credit outcome.
-
Bring the review evidence together
$100 $88cost per correctly reviewed application across the redesigned workflow
I brought verified information, policy rules and AI conclusions into one dossier, with sources and freshness visible. Underwriters can check the basis of a recommendation without repeating the same search across systems.
-
Connect applicant corrections to review
I connected the mobile correction flow to the underwriter's dossier. Applicants see the relevant factors and sources, correct a value and send updated documents back for review. Review-ready submissions rose from 60% to 65% among applicants who started revisions.
The product must show where model guidance ends and accountable human judgement begins.
06
The solution
Underwriting queue with review triggers
I designed the queue around the reason for manual review, time pressure and the next action. Each row tells the underwriter what needs attention, while the side panel opens the next case without hiding the workload.
- Observation
- A generic status did not tell underwriters which application needed attention first.
- Decision
- Rank cases by review trigger and time pressure, then show the next action in context.
- Effect
- Prioritisation and explicit next actions formed part of the connected workflow that reduced review cost by 12%. This result covers the full process, including the dossier and applicant corrections.
Decision dossier with source ledger
I separated verified data, lending policy and AI output in one dossier, then linked each conclusion to its source. A timestamped ledger shows when evidence was checked and when human review began.
- Observation
- Policy conflicts and stale data could disappear inside a single recommendation.
- Decision
- Keep data, policy and model claims separate and link each claim to the ledger.
- Effect
- Less searching, switching between systems and repeated verification helped bring the cost of a correctly reviewed application from $100 to $88, including AI and rework.
Baseline-preserving terms simulator
The simulator compares the original request with a policy-eligible alternative. Loan amount, term, rate, coverage and security remain visible, and the alternative still requires human approval.
- Observation
- Changing inputs during exploration could overwrite the model's original recommendation.
- Decision
- Create a separate counterfactual and preserve the baseline as an immutable reference.
- Effect
- An underwriter can explore policy-eligible terms without recording a credit decision.
Override with authority and rationale
A human override requires a decision, reason category, verified authority and written rationale. The audit preview shows exactly what will be added to the ledger before confirmation.
- Observation
- An override without authority or rationale would weaken accountability.
- Decision
- Require both role verification and a structured explanation before recording the event.
- Effect
- The recorded decision includes the authorised underwriter and rationale. The AI recommendation remains distinguishable from the final credit outcome.
Fairness and drift intervention
Opportunity gap and population drift are monitored against explicit thresholds. A breach pauses automation, holds affected cases in manual review and opens a recovery plan with rollback comparison.
- Observation
- A model alert needed a defined response before affected recommendations could be used in a review.
- Decision
- Connect threshold breach, manual routing, recovery checks and sign-off in one state.
- Effect
- Affected applications enter manual review. The recovery state keeps the checks and sign-off visible before model-assisted processing resumes.
Applicant explanation and recourse
I designed the mobile flow to show why an application needs revision and which source supports each factor. Applicants can correct a value, upload an up-to-date document and send the changes to an underwriter.
- Observation
- A generic pending state gave applicants no way to understand or correct the record.
- Decision
- Expose the relevant evidence and place correction and human review in the same flow.
- Effect
- The share of applicants submitting a complete, review-ready application increased from 60% to 65% among those who started revisions. This measures successful correction, not initial application conversion.
07
Human oversight and recovery
Failure and recovery states
Borderline application
- When
- Evidence is stable but affordability sits near a policy boundary.
- Product response
- The case enters structured human review with a visible reason.
Policy conflict
- When
- The model recommendation conflicts with a lending rule.
- Product response
- The conflict is separated from the model claim and finalisation is blocked.
Stale or missing evidence
- When
- A material source falls outside its accepted freshness window.
- Product response
- The applicant or underwriter receives a specific evidence request.
Fairness or drift breach
- When
- A fairness or population-drift measure exceeds its configured threshold.
- Product response
- Automation pauses and affected applications move to manual review.
Applicant correction
- When
- The applicant identifies an incorrect value or has newer evidence.
- Product response
- The source is updated or the unchanged record is sent to an underwriter.
Safety and review controls
Source and consent
Material claims retain their source, timestamp and data permission context.
Immutable baseline
Simulated terms never overwrite the original recommendation or application.
Verified authority
High-impact actions require an authorised human and written rationale.
Operational model controls
Fairness and drift breaches pause automation and trigger a signed recovery path.
Applicant recourse
The applicant can inspect factors, correct data, submit evidence or request review.
08
AI evaluation
The implemented workflow was assessed through a controlled pilot and a comparison of actual expenses with a comparable process without the changes. Product analytics also tracked whether applicants who started revisions submitted a complete, review-ready application.
Evaluation methods
Cost per reviewed application
The comparison included AI costs and rework. Cost per correctly reviewed application fell from $100 to $88.
Actual net savings
Actual expenses from July 2025 to July 2026 were compared with the comparable process without implementation. Net savings were $1.2M after additional costs, from less paid manual verification, repeated document requests and application reprocessing.
Successful applicant revisions
The measure counted complete, review-ready submissions among applicants who started revisions. The rate increased from 60% to 65%, an 8.3% relative increase.
09
Results
The connected review process delivered $1.2M in actual net savings from July 2025 to July 2026 after additional costs. Cost per correctly reviewed application fell from $100 to $88 through less manual verification, repeated document collection and reprocessing. These figures describe the same operational improvement at different scales. In the mobile flow, review-ready submissions increased from 60% to 65% among applicants who started revisions.
10
Reflection
My contribution was to connect the queue, evidence review and applicant corrections into one process. Keeping data, lending policy and AI guidance separate reduced the need to find and verify the same information again. The financial result belongs to the implemented workflow as a whole; the mobile completion measure shows where the applicant-facing design improved that process.
11
Let’s Connect
Building an AI, FinTech or Enterprise product? I can help shape complex decisions into clear workflows with evidence, controls and measurable outcomes.
Download CV