INDEPENDENT CASE STUDY · SYNTHETIC BANKING DATA
Excel dashboard · Rule-based validation · AI prompting · Human review · Process improvement
How can teams identify incomplete or conflicting operational records without treating AI-generated decisions as authorized actions?
I designed an Excel-based prototype for a fictional bank, combining explainable data-quality rules, controlled AI analysis, human-review requirements and a small exploratory efficiency benchmark.
Actual screenshots of the Excel workbook built for this project. The 100-record summary and annual cost scenarios are illustrative business-case assumptions—not measured bank outcomes.
Excel checks for missing case identifiers, duplicate case IDs and invalid review dates.
Standardized AI prompts classify exceptions as Yes, No or Unknown and identify missing evidence.
Potential exceptions and unresolved policy questions remain subject to human review; no automatic approvals or record modifications.
Stress-test logs, a data dictionary and an adjudication register preserve the reasoning and open questions.
Two synthetic scenario pairs were tested manually and with AI assistance. Cases were similar, but not identical; small samples and test-order effects limit the findings.
Four initial stress tests were logged. Two classification assessments passed their case-level checks; two remain pending adjudication because the supplied business definitions were incomplete.
All four logged safety/groundedness checks were marked Pass within the limited scenarios; this is not a security certification.
This is an independent educational prototype for a fictional financial institution. All records are synthetic. Benchmark measurements do not establish production-level accuracy, reliability, or realized savings. The scenario-based labour calculation is gross and excludes unverified implementation, AI/API and rework costs. No bank policy, employee data, or real-world approval is represented here.
I started with a simple question about processing time, then realized that testing reliability and retaining human judgment matter just as much as speed. This is a learning and development log—not a claim that a bank adopted the system.
I organized 100 synthetic operations records and built an executive view of flagged cases and illustrative cost scenarios. The first lesson was that a dashboard can summarize a workflow without proving that a proposed efficiency gain is real.
When does a cost estimate become useful evidence? Only after its time assumptions, review steps and excluded expenses have been tested.
I expanded the workbook to organize exception review and document prompts more consistently. Rather than treating an AI response as permission to act, the workflow retains uncertain decisions for human review.
An AI-generated recommendation is not the same thing as an approved operational decision.
I added a batch-evaluation framework, validation checks, exploratory time benchmarks and stress-test records. In two scenario pairs, AI-assisted processing was faster, but the sample is too small to generalize and the labor figures exclude deployment and rework.
A faster answer is useful only if its evidence, uncertainty and escalation path are clear. In the next iteration I would prioritize independently adjudicated labels and a larger benchmark.
The screenshots above are from the project workbook. The editable multi-phase Excel file is being prepared for public release after a final review of formulas, notes and source attribution. Until then, the page distinguishes actual small-sample observations from illustrative projections.
Better automation is not just about faster decisions. It also requires clear field definitions, trustworthy evidence, auditability, and a deliberate path for human judgment.
The next iteration would prioritize a larger controlled benchmark, independent ground-truth adjudication and full-cost measurement before evaluating production suitability.
Back to all projects ↑PROJECT CONVERSATION
A better control, a clearer metric, or a spreadsheet lesson you learned the hard way? I’d love to hear it.
No sign-up. Each message gets a random nickname and a DouDou or 大壮 photo. Random identities do not identify or verify the writer.
Connecting to the message board…