Introduction
There is currently a race across organisations to implement artificial intelligence. Many companies are eager to demonstrate that they are “AI-enabled”, often by hiring AI specialists or launching AI initiatives.
However, in many cases the underlying data environment is deeply flawed.
In one organisation, a new role was created and an “AI Specialist” was hired into a department where critical processes were supported by broken spreadsheets, manual data entry, and inconsistent reporting. The organisation believed it was ready for AI. In reality, the foundations required to support it did not exist.
While the hype around AI continues to grow, many organisations would benefit far more from focusing on something much simpler: fixing their data processes.
This article uses a real-world style example to illustrate how relatively low-tech improvements to processes and data management can dramatically improve operations — and create the foundations necessary for meaningful AI adoption.
The Current Process
Consider a typical loan portfolio management process inside a financial institution.
The first source of data comes from legal contracts. These contracts are drafted in Microsoft Word, reviewed internally, sent to the client for signature, and then returned once signed.
Once the contract has been executed, lenders must manually extract the relevant data and enter it into a spreadsheet. These spreadsheets typically have little or no data validation, no structured controls, and no standardised format.
Within the department, a “cottage industry” of spreadsheets develops.
Each lender builds their own spreadsheets to capture borrower financial information. These spreadsheets are used to track financial performance, calculate ratios, and support internal credit reviews.
To populate them, lenders manually enter data from two primary sources:
- The borrower’s published financial statements
- Public filings retrieved from the SEC website or similar regulatory databases
Every lender performs this process independently.
The individuals responsible for this work are typically highly skilled finance professionals, but they are not trained in data management, process design, or automation. As a result, the environment becomes highly fragmented.
Common characteristics include:
- Manual data entry
- No standardised data definitions
- No validation controls
- No consistent reporting structures
Financial data stored across dozens or hundreds of spreadsheets
Because each lender builds their own tools, reporting varies widely between individuals. Portfolio-level reporting becomes extremely difficult to produce.
Even routine processes such as annual credit reviews become unnecessarily cumbersome. Lenders must manually copy financial data into review templates and then write narrative commentary explaining borrower performance.
Risk monitoring exists almost entirely at the individual borrower level, often implemented as simple spreadsheet checks.
Despite this fragmented and manual environment, the organisation may still decide to introduce AI.
In practice, AI cannot function effectively in an environment like this.
Why AI Cannot Work in This Environment
Artificial intelligence depends on one fundamental prerequisite: structured, reliable data.
When data is scattered across spreadsheets, inconsistently formatted, manually entered, and poorly validated, AI models cannot operate effectively. Any output produced by such systems will be unreliable.
Trying to apply AI on top of poor data quality is not innovation. It is a waste of time and money.
The real issue in many organisations is not a lack of advanced technology. It is the separation between technology specialists and subject matter experts. Finance professionals understand the business but not data architecture, while technology teams may not fully understand the operational processes.
The result is a system where both sides operate in isolation and the underlying data environment never matures.
Before AI can deliver value, organisations must first address the fundamentals.
The Improved Process
Principle One: Single Point of Data Entry
A key principle in modern system design is simple:
Data should be entered once and never again.
The first step is to remove the process of drafting contracts manually in Word and replace it with web-based contract templates.
Users would enter the required information through an online form with built-in data validation. Once the data is entered, the contract could be automatically generated as a formatted document.
The process might work as follows:
- A lender completes a web form containing the contract inputs.
- The form enforces validation rules and required fields.
- A second reviewer approves the inputs rather than reviewing the entire document.
- Once approved, the contract is rendered as a PDF.
- The PDF is sent to the client through a digital signing platform such as DocuSign.
- Once signed, the contract is marked as completed and stored.
Rather than storing only the final document, the structured contract inputs can be stored as a JSON object in a database. This allows the data to be retrieved and reused across systems.
Standardising Financial Data
The most transformational improvement an organisation can make is the standardisation of financial data.
Currently, each lender records financial information in their own format. This makes it extremely difficult to aggregate data across the portfolio.
By defining a standard set of financial data points and input templates, organisations can ensure that borrower financial information is captured consistently.
Once financial data is standardised, it becomes possible to store it in structured tables rather than isolated spreadsheets.
This single change unlocks significant capabilities:
- Portfolio reporting
- Automated analysis
- Consistent financial ratios
- Improved risk monitoring
Using Tables and APIs
Once data is stored in structured tables, organisations gain the ability to automate large parts of the data collection process.
Rather than requiring lenders to manually input financial data, organisations can retrieve it directly from external sources.
Using tools such as Python, financial data can be pulled from:
- regulatory filings
- public financial databases
- third-party data providers
However, even when external data sources exist, challenges remain. While regulators such as the SEC often standardise financial line items, individual companies may still label items differently.
To address this, organisations may need to maintain mapping tables that translate external financial labels into the organisation’s standardised financial structure.
Designing these tables properly is critical, particularly in regulated environments such as banking where systems must scale and support strict governance requirements.
Excel vs Front-End Systems
Even in a modern data architecture, Excel can still play a useful role.
In a lower-cost solution, Excel templates can be used as structured input forms that feed data directly into database tables. Approval processes and basic workflows can also be implemented within Excel.
However, this approach only works if a single standardised template is enforced.
Designing this template is not trivial. It requires someone with significant experience in:
- Excel data validation
- spreadsheet controls
- user interface design
- data architecture
The template must allow users to customise views so they only see inputs relevant to their role.
Excel has advantages, particularly for financial users who prefer performing calculations within spreadsheets. However, front-end web interfaces offer benefits as well.
For example, web interfaces make validation rules more transparent and allow customised views for different departments, borrowers, or portfolios.
In hybrid environments where both Excel and front-end systems are used, governance becomes essential.
One useful approach is to maintain a definitions sheet inside the Excel template. This sheet can function as a schema describing all fields and validation rules. The same schema can then be used to generate forms within the web application.
This ensures that both systems remain synchronised.
Portfolio Reporting and Automation
Once financial data and contract information are stored in structured tables, organisations can begin building powerful portfolio reporting.
Tools such as Power BI or Tableau can be used to analyse the entire loan portfolio using standardised data.
This allows organisations to monitor:
- industry concentration
- single borrower exposure
- credit rating distribution
- non-performing loans
- IFRS 9 stage distribution
- loan maturity profiles
- collateral coverage
- RAROC
- early warning indicators
Because the underlying data is structured, these reports can update automatically.
Annual review templates no longer require manual copying of financial data. Instead, the latest data can be automatically inserted into review documents or dashboards.
Risk monitoring can also be run daily at the portfolio level. Organisations can implement automated checks not only based on internal policy but also on industry-specific risk indicators.
Where AI Finally Adds Value
Only once these foundations are in place does AI begin to deliver real value.
One immediate application is the generation of narrative commentary for reports.
Using a large language model running locally, organisations can feed structured JSON data into the model and automatically generate portfolio commentary or borrower performance summaries.
Tasks that previously required hours of manual writing can be completed in seconds.
Automated commentary also introduces several benefits:
- consistent reporting language
- clearer explanations of key financial indicators
- standardised analytical frameworks
- easier reviews by senior management
Industry-specific commentary can also be generated automatically, reflecting the key metrics most relevant to each sector.
Human input still remains important. Systems should always allow lenders to add their own commentary and contextual insights.
In regulated industries such as banking, organisations must also ensure that systems comply with frameworks such as BCBS 239, which emphasise strong governance, data accuracy, and traceability.
Conclusion
Artificial intelligence is often presented as a technological revolution that will transform organisations overnight.
In reality, the biggest challenge is rarely the AI itself.
The real challenge is building the data foundations required to support it.
Organisations that attempt to layer AI on top of fragmented spreadsheets and manual processes will see limited results. AI does not fix broken systems — it amplifies them.
The organisations that will benefit most from AI are those that first focus on the fundamentals:
- structured data
- standardised processes
- controlled inputs
- strong governance
Once these foundations exist, AI becomes a powerful tool rather than an expensive experiment.
Before implementing AI, many organisations should ask a much simpler question:
Have we fixed our spreadsheets yet?