Enterprise automation becomes considerably more challenging when the information required to complete a process is scattered across contracts, forms, supporting documents and operational emails. In a recent Trade Finance project, TechTalent developed an AI-assisted solution to support the preparation of bank guarantee documentation. The existing process relied heavily on specialists to review incoming documents, locate relevant information and manually transfer it into the systems used to prepare each draft.
TechTalent focused on one of the most demanding parts of this workflow: transforming information from multiple documents into structured, traceable data that specialists could review and validate. The solution combines document processing, AI-assisted extraction and deterministic validation while keeping human oversight at the centre of the process.
The Challenge - Extracting the Right Information from Multiple Sources
Preparing a bank guarantee draft involves considerably more than reading a single form. A request can arrive together with a commercial contract and additional supporting documentation such as proforma invoices, amendments or association agreements. In some cases, information can also originate from an operational email.
This meant specialists had to:
- review information across several documents and formats;
- identify and manually transfer the fields required for the draft;
- compare values appearing in different sources;
- identify missing or conflicting information;
- request additional information where necessary;
- prepare the information required for the appropriate draft.
The process was also time-sensitive. Requests were generally urgent, making processing speed important without reducing the level of control expected in a financial environment.
Designing the Solution Around the Existing Workflow
A key architectural decision was to avoid replacing the application already used by specialists. Instead, the new capability was designed as an extraction service integrated with the existing platform.
The workflow follows a clear sequence:
- Documents are submitted through the existing enterprise application.
- The processing service analyses them, identifying document types and extracting relevant information.
- Values from different sources are compared and reconciled according to predefined rules.
- Structured results are returned together with source, confidence and review information.
- The specialist validates or corrects the fields before the existing platform continues with draft generation.
This separation of responsibilities kept the scope of the new system clear. The service provides the structured information needed for the draft, while the surrounding platform continues to manage the user interface, template selection and generation of the final document.
From Unstructured Documents to Structured Data
Behind that relatively simple interaction sits a multi-stage document-processing pipeline.
The solution was designed to:
- identify and classify the documents included in a request;
- extract text and document structure from PDFs and scanned files;
- process digitally available text from Word documents and operational emails;
- identify the fields required from each source;
- compare information appearing across different documents;
- flag missing or conflicting values;
- return structured information together with its source and confidence level.
The variety of input formats made this particularly important. The system needs to work with electronically generated PDFs, scanned documents, Word files and text from operational emails.
OCR and document-layout analysis are used where necessary, while digitally available text can be processed directly. Once the documents have been interpreted, the extraction layer converts the relevant information into a consistent structured format that can be consumed by the existing enterprise application.
Using AI Where It Adds Value
AI has a clearly defined role within the architecture. LLMs support document classification and the extraction of structured fields from variable document content. This is particularly useful for documents such as commercial contracts, where the same piece of information may appear in different locations and be expressed in different ways.
However, the architecture does not rely on AI alone, combining several approaches according to the task:
- AI-assisted extraction identifies relevant information within variable documents.
- OCR and layout analysis process scanned content and preserve information about where values appear.
- Deterministic reconciliation compares values found across different sources.
- Source verification checks that AI-proposed values can be found in the underlying document text.
- Human validation remains mandatory before the information moves forward in the process.
If a proposed value cannot be confirmed in the source, it can be marked for manual verification rather than silently passed through the workflow.
This creates an important distinction between using AI to interpret information and allowing AI to make the underlying business decision.
Reconciling Information Across Documents
Extraction was only one part of the problem, as the same information can appear in both a request and its supporting contract, and those values do not necessarily match.
The solution therefore includes a deterministic reconciliation layer that makes these differences explicit.
A field can be returned as:
- No difference – the relevant sources contain matching information.
- Conflict – different values have been identified and require human confirmation.
- Not found – the required information could not be identified in the available sources.
- Needs manual review – confidence is low or the extracted value cannot be sufficiently confirmed against the source.
Where conflicting values are identified, the proposed value and alternative value can both be preserved. Instead of hiding inconsistencies, the system gives specialists the information needed to investigate them.
Making Every Extracted Value Traceable
For an enterprise financial workflow, returning a value is only part of the requirement. Specialists and technical teams also need to understand where that value came from and how it was processed.
The solution thus incorporates traceability at field level, including information such as:
- the source document;
- the relevant page and document region where available;
- the value identified in each source;
- confidence and manual-review indicators;
- the method used to obtain the value;
- the reconciliation rule applied.
This creates a clearer path for specialists to investigate unexpected results and gives technical teams more context when an extraction needs to be corrected or debugged.
Traceability also extends across the processing workflow, with structured logging providing visibility into the stages completed during each request.
Keeping Human Validation at the Centre
The project deliberately stops before autonomous decision-making. The extraction service does not generate the final guarantee, choose the final template, approve legal clauses or make credit, risk or eligibility decisions. Instead, it supplies structured information to the existing application, where specialists validate or correct the proposed values before the document-generation process continues.
The same principle applies when the system encounters uncertainty:
- handwritten information can be flagged for additional verification;
- poor-quality scans can result in lower-confidence fields;
- missing information remains visibly incomplete;
- conflicting values are presented for review;
- values that cannot be confirmed against the source are marked for manual attention;
- unsupported or ineligible cases can remain on the established manual workflow.
This allows AI to reduce some of the repetitive work involved in understanding documents without removing the controls required around the final output.
Integrating AI into an Enterprise Architecture
The solution also had to operate within an established technology environment rather than as a standalone AI application.
The architecture combines:
- a .NET-based API and processing layer;
- Azure Durable Functions for asynchronous orchestration;
- Azure Document Intelligence for OCR and layout extraction;
- Azure AI services for LLM-based processing;
- Azure Blob Storage and SQL Server for the required storage and processing components;
- Application Insights and structured logging for observability.
Processing takes place within the organization’s existing Azure environment, with the new capability integrated into the surrounding application through dedicated APIs.
Security, auditability and observability were therefore treated as part of the architecture rather than additions around the AI functionality.
What This Project Shows About Enterprise AI
This project illustrates an important pattern for enterprise AI adoption. Some of the most useful applications of AI focus on a specific part of an established workflow where people currently spend significant time interpreting, comparing and transferring information.
In this case, different technologies have distinct responsibilities: AI supports document understanding and extraction, deterministic logic handles reconciliation and verification, and existing enterprise software manages the surrounding workflow. Specialists remain responsible for validating the information before it is used.
The result is an architecture in which AI becomes one component of a broader engineering solution.
For organizations exploring similar opportunities, this distinction matters. Effective enterprise AI depends on more than the underlying model. Integration, traceability, exception handling, deterministic controls and human oversight all influence whether an AI capability can work effectively within an existing business process.
Building Enterprise AI Solutions with TechTalent
TechTalent helps organizations bring together the software engineering, cloud, data and AI expertise required to build technology around complex business processes.
From document intelligence and AI-assisted workflows to enterprise integrations and custom software development, we build teams around the technical requirements of each initiative.
If you are exploring how AI could support an existing enterprise process, contact us to discuss your project.
