How SAP Concur Uses Agentic AI to Automate Expense Reporting

Expense automation has long operated on a straightforward principle: machines that can read text can handle the processing. Yet anyone who's submitted a faded, wrinkled receipt knows that optical recognition alone falls short. When critical details like location or date are illegible or absent, automation stalls and employees must manually fill the gaps.

SAP Concur's engineering team identified this limitation as an opportunity. While competitors focused on refining conversational interfaces, SAP Concur anticipated a more fundamental transformation: the shift from text extraction to intelligent inference.

The outcome is an agentic AI enhancement to ExpenseIt that moves beyond simple text recognition to solve complex data puzzles, dramatically cutting manual data entry. Users now photograph receipts, upload digital copies, or forward email receipts, and ExpenseIt automatically generates complete expense entries without requiring manual date input or itemization.

Executing this vision required a partner capable of matching both technical ambition and rapid development velocity. SAP Concur combined its strategic vision with Google Cloud's comprehensive AI infrastructure—the only provider engineering every layer from custom silicon through data platforms to advanced models and agents. The collaboration produced a genuine advancement in expense management: an AI agent that captures receipt data while understanding the business traveler's context.

Engineering intelligence at scale

Traditional expense systems excel at extracting visible receipt data but cannot infer missing information. SAP Concur recognized AI agents as the path to systems capable of reasoning, decision-making, and autonomous action.

Consider uploading a lunch receipt from "The Main St. Café" without an address. Previously, this gap would halt automation entirely, forcing manual data entry.

Agentic capabilities enable analysis of contextual signals—vendor names, expense categories, travel itineraries—to complete incomplete data. SAP Concur aimed to build an AI agent reasoning like a human assistant: "This receipt shows 'Main St. Café.' The transaction timing aligns with a business trip including a Dallas flight and Greenville, Texas hotel booking. This vendor is likely the restaurant near the hotel in Paris, Texas—not Paris, France."

The teams tackled this challenge with startup agility, prioritizing rapid prototyping over extended development cycles.

Leveraging Google's Gemini models, they constructed the Receipt Analysis Agent with a cognitive architecture.

The workflow:

1. Ingestion: Users capture photos via the SAP Concur mobile app, upload digital scans, or forward email receipts.

2. Deterministic core: SAP's foundational technology, refined through decades of global expense processing, extracts visible receipt text with precision.

3. Intelligent routing layer: Clear receipt data bypasses additional processing. Ambiguous data (such as missing location) triggers routing to the Receipt Analysis Agent.

4. Contextual reasoning: Built on Gemini models, the agent infers missing information using tools and grounding data. ExpenseIt provides partial receipt data alongside contextual information like travel itineraries and business calendars.

5. ReAct framework: The Receipt Analysis Agent validates vendors against location history and completes expense entries.

In this example, ExpenseIt detects the missing location and activates the intelligent routing layer to trigger the Receipt Analysis Agent. Using Gemini, the agent identifies gaps, analyzes contextual clues and user-specific data, and makes decisions informed by travel bookings and calendar events.

Core design patterns for AI agents

The Receipt Analysis Agent architecture draws from Agentic Design Patterns, a practical guide by senior Google engineer Antonio Gulli. This framework enabled SAP Concur to transform ExpenseIt into a system that reasons across data both present and absent from receipts to generate accurate expense entries.

The teams implemented the Routing Pattern to optimize cost and intelligence by selectively engaging the AI agent. The routing architecture classifies tasks: high-confidence OCR receipts follow the standard deterministic path, while low-confidence receipts (such as those with missing locations) route to the Receipt Analysis Agent.

The Reflection Pattern addresses challenges like the Paris Paradox, ensuring the agent validates answers rather than generating responses like a basic chatbot. This pattern employs an internal generator-critic loop where the model proposes a hypothesis ("This is Paris, France") then critiques it against known facts ("The itinerary indicates Dallas, Texas. This hypothesis is likely incorrect").

The agent also implements the Tool Use Pattern, providing direct API access to grounding sources like Concur Travel itineraries. This approach enables the agent to retrieve factual information rather than generate it, transforming the system from text generator to factual researcher.

Building for uncertainty: Google Cloud's integrated advantage

This project represents a significant evolution in intelligent system design. By pairing a deterministic core with an agentic reasoning layer, SAP Concur demonstrated that AI's greatest value often lies not in processing available data, but in reasoning to discover missing data. A critical breakthrough was reconceptualizing how the model functions—moving beyond treating Gemini as a generative interface to deploying it as a logic engine.

SAP Concur selected Google Cloud because agent effectiveness depends on world knowledge—and Google's understanding of the digital landscape is unmatched.

While the current release leverages Gemini's reasoning capabilities, the partnership enables future multimodal, full-stack intelligence unique to this ecosystem, including:

• Real-world grounding: Future agents could cross-reference receipts with Google Maps data to verify business locations.

• Frictionless integration: Potential integrations with Google Wallet for instant transaction matching, or Gmail for automatic hotel receipt surfacing.

• Edge intelligence: Mobile advances like Gemini Nano and Android AICore could enable on-device processing, delivering speed and privacy without data transmission.

SAP Concur provides deep domain expertise powering global financial transactions. Google Cloud delivers the complete AI stack from custom-designed TPUs optimized for training to the mobile OS in users' pockets.

Building your reasoning engine

Creating a reasoning engine like ExpenseIt doesn't require starting from scratch. The architectural patterns discussed—Routing, Reflection, and Tool Use—are implemented directly in the Google Agent Development Kit (ADK). The ADK provides frameworks and best practices to transition from prompt engineering to system engineering, offering a blueprint for building reliable, scalable, enterprise-ready agents.