How Aeris and AltexSoft Cut Data Preparation Time by 30

Background

Our US-based client Aeris helps investors and lenders evaluate organizations that channel capital into affordable housing, small businesses, healthcare, and other community development projects. Its ratings rely on hundreds of data points pulled from audits, loan schedules, and portfolio reports, excluding borrowers’ PII, which Aeris doesn’t collect.

Aeris associates used to manually find the relevant information across all these documents, enter it into spreadsheet templates, and then check everything again for accuracy.

To automate this repetitive work, AltexSoft built an AI assistant that searches Aeris’s document library, fills in the required data while linking every figure to its source, validates the completed workbook, and helps assemble the final report. Importantly, it never makes up a number: If the required figure cannot be found in an approved document, the assistant flags it instead of guessing. The solution, which uses Claude by Anthropic as its primary reasoning model, has reduced the time needed for data preparation by 30–40 percent

Challenges

By incorporating AI into the Excel and Word templates people already use, AltexSoft avoided disrupting operations or requiring staff to learn a new tool, while addressing several business challenges.

Protect employees’ time for higher-value work

Scale assessment capacity without adding proportional staff workload

Improve data accuracy and traceability to reduce rework and correction risk

Lower the cost of skilled staff time spent on manual preparation

Value Delivered

Turning documents into usable data

We started by building a data extraction layer that can process PDFs, spreadsheets, Word documents, and scanned files. Native digital documents are parsed directly, while scanned pages first go through optical character recognition (OCR) before the extracted content is structured and processed. The system also detects complex elements such as rotated tables and breaks large documents into searchable chunks.

This transforms Aeris’s source materials into consistent, machine-readable content that can be searched and used by the AI assistants.

Grounding AI in approved sources

Next, we added a retrieval-augmented generation, or RAG, layer that indexes Aeris’s document library and uses semantic search to identify the most useful passages for each task.

Instead of relying on the model’s general knowledge, the assistants work with Aeris’s own approved documents. This reduces the risk of hallucinations and makes it possible to trace every suggested figure back to its source.

Creating three specialized AI workflows for different tasks

Rather than relying on a single general-purpose chatbot, AltexSoft created three specialized AI workflows as custom Claude skills, each tailored to a specific task:

The data-entry workflow finds required figures in source documents and fills the existing Excel templates, attaching a source to every value.
The verification workflow checks completed workbooks against the original documents and the related spreadsheet tabs, flagging discrepancies or missing information.
The report-generation workflow transfers approved data into the final Word report, follows the required formatting, and highlights items that still need human input.

The three workflows run directly inside Excel and Word, so employees can use them without leaving familiar tools. A separate browser-based assistant supports general document Q&A, with secure login, multiple user sessions, and built-in tracking of usage and costs.

Adding security, controls, and traceability

We integrated the AI workflows with Aeris’s documents and supporting services through a secure, standards-based Model Context Protocol (MCP) connector. Authentication and controlled permissions ensure that each assistant can interact only with approved resources and data.

The workflows also follow a structured methodology built from Aeris’s existing guides and staff knowledge. It defines which sources to use, how to calculate values, and what to do when documents disagree.

An audit trail records what changed, why, and which source was used, making every figure traceable. If information is missing or conflicting, the assistant flags the issue and leaves the final decision to the reviewer.

Approach and Technical Info

In total, the solution was designed and delivered in approximately six months.

The AltexSoft team consisted of one project manager and two data science/AI engineers, with only one engineer working on the solution at a time.

The technology stack included the Claude API from Anthropic, Python, ChromaDB, Streamlit, the Model Context Protocol (MCP), and AWS EC2.

PythonChromaDBClaude API by AnthropicMCPAWS EC2