AI Strategy & Innovation
AI Strategy & Innovation · Bank of Georgia · 2026
A strategic service design and AI strategy project exploring how Bank of Georgia could reduce mortgage approval time by integrating an AI-assisted property valuation model into the existing process. The project spanned discovery research, stakeholder interviews, journey mapping, and service blueprinting — resulting in an 8-month implementation roadmap that targets a reduction in approval time from 9 business days to 5.
The core insight was that a single step — manual property valuation — accounts for more than half the total approval time. By replacing it with an Automated Valuation Model (AVM) for standard residential properties, the bank can dramatically accelerate processing while freeing licensed appraisers for complex, high-value cases where human judgement is genuinely needed.
Led the end-to-end service design process — from stakeholder interviews and journey mapping through to service blueprinting and roadmap development. Responsible for facilitating workshops, synthesizing insights, and translating findings into actionable proposals for a cross-functional team.
Conducted interviews with mortgage specialists, branch managers, compliance officers, and licensed property appraisers to map the current-state process and surface operational bottlenecks. Key questions explored:
Interviews were conducted with 12 participants across 5 roles. Sessions were recorded and synthesised into an affinity map highlighting themes around valuation delays, document chasing, and inconsistent communication with applicants.
Facilitated a half-day cross-functional workshop with mortgage specialists, branch managers, a compliance officer, and a systems lead. The goal was to map the current mortgage application journey across three lenses simultaneously:
The workshop surfaced 14 distinct pain points. The property valuation step appeared across all three lenses as the most consistent blocker — slow for customers, expensive for the bank, and highly dependent on appraiser availability.
Plotted all 14 pain points on an effort/impact matrix to identify where a focused intervention would deliver the greatest return. Five issues emerged as high-impact and low-to-medium effort:
The property valuation step was selected as the primary focus: it was the largest single contributor to total approval time, it affected every application, and a viable technical solution (automated valuation models) was already proven in other markets.
The current property valuation process is entirely manual. A licensed appraiser physically visits each property, prepares a written report, and submits it to the bank. This step takes 3–5 business days and is required for every mortgage application — regardless of property type, value, or complexity.
This creates three compounding problems:
Introduce an Automated Valuation Model (AVM) that cross-references property registry data, recent comparable sales, and market trend signals to generate an instant initial valuation — eliminating the 3–5 day manual appraisal step for the majority of applications.
The model uses a tiered routing logic:
This approach targets 70–80% AVM coverage for standard residential mortgages while preserving the human review layer for cases where model confidence is insufficient.
Created a detailed TO-BE service blueprint mapping the redesigned mortgage process across all layers of the service. The blueprint served as the primary alignment artefact for cross-functional discussions with technology, compliance, and operations teams.
Key blueprint layers:
Full-time core team:
As-needed contributors:
| Period | Activities |
|---|---|
| Month 1 | Discovery & AS-IS journey mapping · TO-BE journey & business case · data audit & feasibility |
| Month 2 | Regulatory consultation · data partnership agreements · core team formation |
| Month 3 | Data infrastructure setup · registry API integration · comparable sales database ingestion |
| Month 4 | AVM model development · initial training on historical valuation data |
| Month 5 | Internal testing against historical data · accuracy benchmarking · confidence threshold calibration |
| Month 6 | Limited pilot — 50 applications · AVM run in parallel with manual appraisal for comparison |
| Month 7 | Pilot analysis · model refinement · compliance sign-off · staff briefings |
| Month 8 | Phased rollout to all standard residential mortgage applications · monitoring dashboard live |
The pilot (Month 6) runs across 5 steps:
After pilot sign-off, rollout is phased: starting with standard residential properties under a defined value threshold, expanding as the model's accuracy track record builds. Appraiser resource is redeployed to complex and commercial valuations — the work that genuinely requires physical inspection and professional judgement.
The proposed changes target a 44% reduction in average approval time — from 9 business days to 5 — by eliminating the manual appraisal step for 70–80% of standard residential applications. Appraiser capacity freed from routine valuations is redeployed to complex and commercial cases where specialist judgement genuinely adds value. The applicant experience improves through faster decisions and reduced uncertainty during the process.
This project reinforced how service design and AI strategy complement each other when applied to operational transformation. The technical feasibility of an AVM was never the hard part — models like this have been running in other markets for years. The harder problem was designing the governance layer: when to trust the model, how to log decisions for regulatory compliance, and how to bring mortgage specialists along without creating resistance to a process that would change their role.
The service blueprint became the key alignment tool. It made the abstract concrete — showing every team member exactly where the AI touched the process, where a human remained in the loop, and what happened at every exception. Without it, the conversation stayed at the level of whether AI should be used at all; with it, the conversation shifted to how.
If I were to extend this work, the next step would be building a more granular data quality framework for the registry feed — the AVM is only as good as the data it's trained on, and the Georgian property registry has known gaps in recent transaction history for some regions.
Skills applied: Service Design · Journey Mapping · Stakeholder Interviews · Service Blueprinting · AI Strategy · Workshop Facilitation · Roadmap Development · Process Optimisation · Risk Analysis · Change Management