AI Agents for Expense Automation: SAP Concur & AmexGBT

Why AI Agents Are Changing the Rules in Business Travel Expense Management

In 2024, SAP Concur launched a pilot of autonomous agents for processing expense reports in real time. In parallel, American Express Global Business Travel announced the Amex GBT AI platform, which by 2026 should fully automate the cycle from booking to reimbursement for corporate clients. Both solutions share one idea: replace the manual work of travel managers and accountants with agents that make decisions without human involvement.

According to the Deloitte report "AI in Travel and Expense Management 2024," companies with business travel turnover exceeding 5 million dollars per year spend an average of 18 minutes processing one expense report. For an organization with 500 trips per month, this amounts to 150 person-hours monthly. AI agents reduce this time to 2-3 minutes, without employee participation.

How SAP Concur AI Agents Work: Architecture and Functions

SAP Concur built the system on three agent levels. The first level is the data capture agent. It scans receipts, invoices, and emails, extracts amounts, dates, expense categories, and links them to the corporate card or advance report. Optical character recognition (OCR) technology is supplemented by a language model that understands context: for example, it distinguishes a client dinner from a personal dinner by time, location, and amount.

The second level is the policy verification agent. It matches each expense against the corporate travel policy in real time. If an employee books a hotel for 250 euros in Berlin when the limit is 200 euros, the agent automatically rejects the transaction or requests justification via messenger. SAP Concur integrated this function with Microsoft Teams and Slack so employees can respond directly in chat without opening the portal.

The third level is the approval and reimbursement agent. It analyzes the employee's expense history, frequency of violations, expense category, and makes an approval decision. If the agent approves the report, it automatically initiates payment through integration with the accounting system (SAP S/4HANA, Oracle NetSuite, or QuickBooks). The average cycle from report submission to payment dropped from 7-10 days to 24 hours.

AmexGBT Case: Autonomous Booking and Dynamic Expense Optimization

Amex GBT AI went further and combined booking with expense management into a single agent. An employee writes in the corporate chat: "Need a flight to London on March 15, return March 18, budget up to 600 euros." The agent analyzes the corporate policy, airline contracts, employee preferences (window seat, direct flight), and offers three options with justification.

After booking, the agent tracks flight changes, hotel prices, and currency rates. If the hotel price drops 15% two days before check-in, the agent automatically rebooks the room and returns the difference to the corporate card. According to AmexGBT data, such dynamic optimization saves clients an average of 8-12% of accommodation budget.

An important detail: the agent considers not only price but also risks. For example, if a flight has a high probability of delay (based on historical data), the agent will suggest an alternative even if it costs 50 euros more. This reduces operational risks and emergency rebooking costs.

Practical Example: Engineering Company with 200 Employees

Consider a real scenario. An engineering company from Munich sends 40 employees on business trips monthly: project meetings, site audits, conferences. Before implementing AI agents, the travel manager spent 12 hours per week processing reports, checking receipts, and coordinating policy exceptions.

After integrating SAP Concur with AI agents, the company configured rules: automatic approval of expenses up to 100 euros without a receipt (for small taxis and lunches), automatic hotel booking within 2 km of the meeting location with a 180-euro-per-night limit, automatic reimbursement within 48 hours for policy-compliant expenses.

Results after six months: report processing time reduced by 73%, exception requests decreased by 41% (employees see limits before booking), average reimbursement time dropped from 9 to 2 days. Savings on administrative staff salaries amounted to 28 thousand euros per year.

Technical Requirements for Implementing AI Agents

For AI agents to work effectively, clean data infrastructure is needed. The first requirement is a unified base of corporate policies in machine-readable format. If your travel policy exists only in PDF or Word, the agent cannot apply it. SAP Concur and AmexGBT require the policy to be structured: limits by categories, geographic zones, positions, and projects.

The second requirement is integration with corporate cards and banking APIs. Agents receive transaction data in real time through open banking APIs (PSD2 in Europe, equivalents in other regions). Without this integration, the agent cannot automatically match expenses with reports.

The third requirement is a unified identification system. The agent must know exactly who is making the expense: position, project, budget center. This requires integration with HRIS (Workday, SAP SuccessFactors, BambooHR) and single sign-on (SSO). Without this, the agent cannot apply personalized rules.

The fourth requirement is APIs for accounting systems. The agent must automatically create entries, initiate payments, and close advance reports. SAP Concur supports direct integration with SAP S/4HANA, Oracle ERP Cloud, Microsoft Dynamics 365, and dozens of other systems. If your accounting operates in an isolated system without APIs, automation will remain incomplete.

Implementation Economics: When AI Agents Pay Off

SAP Concur publishes subscription costs for AI agents: from 12 to 25 dollars per active user per month, depending on transaction volume and automation level. For a company with 200 trips per month, this is 2,400-5,000 dollars monthly.

Payback depends on three factors: cost of administrative staff working time, frequency of policy violations, and reimbursement speed. If a travel manager with a 4,000-euro monthly salary spends 40% of time processing reports, automation saves 1,600 euros monthly on salary alone. Add reduced policy violations (averaging 7-9% of travel budget) and accelerated reimbursement (which reduces employee dissatisfaction and turnover).

According to AmexGBT calculations, companies with travel budgets from 500 thousand dollars per year recoup AI agent implementation in 8-14 months. For companies with budgets below 200 thousand dollars, payback stretches to 24 months, and in this case it makes sense to start with partial automation: only expense approval or only booking.

What Can Go Wrong: Common Implementation Mistakes

The first mistake is trying to automate a chaotic policy. If your company allows exceptions in 30% of cases, the agent will constantly escalate decisions to humans, and automation will lose meaning. Before implementation, simplify the policy: clear limits, minimal exceptions, transparent criteria.

The second mistake is ignoring employee training. The agent works well when employees understand how to formulate requests and provide data. If an employee photographs a receipt on crumpled paper in poor lighting, OCR will produce errors and the agent will request manual verification. SAP Concur recommends conducting 30-minute onboarding for all users.

The third mistake is lack of agent decision monitoring. AI agents make decisions based on probabilities, and sometimes they err. Regular auditing is needed: which expenses the agent approved automatically, which it rejected, how many escalations to humans. AmexGBT built a dashboard for travel managers where all automatic decisions from the last 30 days are visible.

How to Start: Step-by-Step Plan for Travel Managers

Step one: audit current processes. Collect data from the last three months: how many reports processed, how much time one report takes, how many policy violations, average reimbursement time. This is the baseline for measuring effect.

Step two: structure the corporate policy. Translate all rules into "if-then" format: if position X and city Y, then hotel limit Z euros. Remove vague wording like "reasonable expenses" or "by agreement with supervisor."

Step three: select a pilot group. Start with 20-30 employees who travel frequently and are familiar with digital tools. Launch AI agents for this group for three months and collect feedback.

Step four: integrate systems. Connect corporate cards, HRIS, accounting, and booking systems to the AI agent platform. This is the most labor-intensive stage, requiring help from the IT department or external integrator. For example, integrations with corporate systems allow linking employee data, budgets, and transactions into a unified ecosystem.

Step five: launch automation in stages. First enable automatic approval only for expenses up to 50 euros. After a month, raise the threshold to 100 euros. After three months, add automatic booking. Gradual implementation reduces risks and allows time for rule adjustments.

What Awaits the AI Agent Market in 2026

SAP Concur plans to add predictive analytics: the agent will forecast travel budget a quarter ahead based on meeting calendars, historical data, and seasonality. AmexGBT is working on an agent for supplier negotiations: the system will automatically request discounts from hotels and airlines if booking volume exceeds a threshold value.

Independent players are also emerging. Startup TravelPerk announced an AI agent for small business with simplified setup: policy is set through a chatbot in 10 minutes, accounting integration happens automatically via Zapier. This lowers the entry barrier for companies with travel budgets from 50 thousand dollars per year.

Simultaneously, transparency requirements are growing. The European AI Act regulator requires autonomous systems to explain their decisions. SAP Concur already added an Explain function: if the agent rejects an expense, the employee sees the specific policy rule and can dispute the decision through the interface.

Risks and Limitations of Autonomous Agents

AI agents still handle non-standard situations poorly. For example, if an employee is stranded at an airport due to flight cancellation and must book a hotel for 300 euros instead of the permitted 150 euros, the agent may automatically reject the expense. For such cases, a quick escalation mechanism to a human is needed.

The second risk is dependence on data quality. If the corporate card does not transmit detailed category codes (MCC), the agent cannot distinguish a business lunch from a personal dinner. This leads to false rejections or, conversely, missed violations.

The third risk is employee resistance. Some perceive AI agents as a tool of total control. To reduce this resistance, it is important to emphasize the benefit to the employee: fast reimbursement, less bureaucracy, transparent rules.

Conclusions for Practitioners

AI agents for expense automation are no longer an experiment but a working tool. SAP Concur and AmexGBT show that autonomous systems can process 70-80% of reports without human involvement, reduce the reimbursement cycle threefold, and cut policy violations by a third.

Implementation requires preparation: structured policy, system integration, employee training. Payback comes faster for companies with travel budgets from 500 thousand dollars per year, but even small businesses can start with partial automation.

The main thing is not to try to automate chaos. First simplify processes, then hand them over to agents.

FAQ

How much does implementing AI agents for business travel expense automation cost?

SAP Concur charges from 12 to 25 dollars per active user per month. For a company with 200 trips monthly, this is 2,400-5,000 dollars. Payback occurs in 8-14 months with a travel budget from 500 thousand dollars per year.

What systems need to be integrated for AI agents to work?

Integration is required with corporate cards (via banking APIs), HRIS (Workday, SAP SuccessFactors), accounting (SAP S/4HANA, Oracle, QuickBooks), and single sign-on (SSO). Without these integrations, the agent cannot automatically match expenses and initiate payments.

How do AI agents verify expense compliance with corporate policy?

The agent matches each expense against machine-readable policy in real time: checks limits by categories, geography, position, and project. If an expense exceeds the limit, the agent automatically rejects the transaction or requests justification via messenger (Teams, Slack).

Can AI agents be implemented in a company with a travel budget below 200 thousand dollars?

Yes, but payback will stretch to 24 months. In this case, it makes sense to start with partial automation: only approval of expenses up to 100 euros or only automatic hotel booking within the limit.

How do AI agents handle non-standard situations, such as flight cancellations?

Agents still handle exceptions poorly. If an employee must book an expensive hotel due to force majeure, the agent may automatically reject the expense. For such cases, a quick escalation mechanism to the travel manager via interface or chat is needed.