AI agents that scale

01

Voucher processing

Industry / client · Employee benefits / corporate payment solutions · [Confidential]

The challenge

Employees upload invoices, receipts, and other required documents to the client’s portal to claim reimbursement under a government programme. Every claim has to be checked against more than 20 hard rules, across two voucher types and four subtypes, on a volume that swings between 300 and 700 vouchers a day under a 72-hour SLA, including weekends.

Our solution

Agentic AI (Claude Sonnet 4.5) integrated with UiPath processes submitted document bundles. It extracts and understands invoices and receipts, compares them against ground-truth data from the application, records the validation outcome and routes each case for approval or rejection automatically. The agent does this processing in a single call, making sure we don’t overcharge the client if not required.

Tech stack

UiPath RPA · Claude Sonnet 4.5 · Client claims portal / ground-truth data

Results

• ~71% faster processing per request (3.5 min → ~1 min).
• 100% rule consistency — zero deviations in validation logic.
• Request-to-resolution cycle down from 72 hours to 18 hours.
• ~3,500 transactions a month processed automatically.
• Analysts freed from false-positive reviews to focus on exceptions and true positives.

02

MIA - Agentic Personal Assistant

Industry / client · Internal platform / ERP · Wittenberger Industries

The challenge

Romanian law requires us to keep timesheets of what we work on and when. Rather than pay for a stack we would only half-use, we built our own ERP to cover timesheets and the rest of our back-office admin.

Our solution

We made the website optional. Mia is an agent we talk to in Microsoft Teams: it fills the timesheet from chat, pulls legal documents on request, and keeps strict isolation between chat sessions so nothing leaks between users. It is built on the NOOA (object-oriented agents) architecture, so new capabilities plug in without reworking the core.

Tech stack

NOOA object-oriented agent architecture · Microsoft Teams as the interface · In-house ERP backend

Results

• Replaced our Workday and Jira usage in a week.
• Timesheets filled from a chat message instead of a web form.
• Session isolation by design, so one conversation cannot leak into another.
• New capabilities plug into the NOOA core without reworking it.

Architecture reference: github.com/NVIDIA-NeMo/labs-OO-Agents. This is the build we most enjoy walking clients through in detail — happy to do that on a call.

03

Outlook sales agent assistant

Industry / client · Finance · [Confidential]

The challenge

The request was to create a conversational agentic assistant embedded directly in Outlook that uses internal historical precedents, policies, and best practices to give employees advice on how to handle different situations, look up relevant data, and draft relevant replies directly in Outlook.

Our solution

We created a UiPath conversational agent, a custom Outlook add-in with a chat UI connected to the agent, and a context creation pipeline made with Python scripts. It takes emails / inboxes from existing team members, classifies email threads into request types, and generates knowledge units (the request content, the solution, whether escalation was needed) which are then loaded inside the context grounding index used by the agent.

Tech stack

UiPath conversational agent · Custom Outlook add-in · Python context pipeline · Context Grounding

04

User identity validation process

Industry / client · Finance · [Confidential]

The challenge

The request was to create an automation that takes documents uploaded by users in order to prove their identity and successfully extract and validate the necessary information from them (name, address, issue date). The documents came in different types (utility bills, bank statements, IDs) and from countries all over the world, in different languages and writing systems (Latin, Cyrillic, Arabic, Chinese).

Our solution

We created a UiPath Document Understanding process triggered automatically through API by the client’s internal back office and trained several robust machine-learning models to handle the data extraction from each document type. Inside the process we also implemented validation and post-processing: transliterating extracted data to Latin, and validating that the extracted address is a real location and not a public or invalid one (an office, a PO box, an address from a different country than the issuing entity).

Tech stack

UiPath Document Understanding · Machine learning models · API trigger · Transliteration · Address validation