DHL Supply Chain Is Keeping Its AI Agent Team Small on Purpose

An early procurement-requisition agent was built in an afternoon by a functional consultant who had never used the tool before. “He’s never used Agent Studio before ever,” Walker said. That speed makes development more accessible, while giving the team another reason to control who builds and deploys agents.

Invoice processing is further along. Game said Oracle’s Document IO had been running in the environment since around May or June, replacing the previous invoice-recognition engine.

“We did absolutely nothing,” Game said of the update, reporting recognition rates on supplier invoice data improving from roughly 60% to about 95% to 96%. The improvement concerns data recognition; it does not establish how many invoices are processed without human involvement.

Oracle’s documentation describes Document IO as using large language models (LLMs) to extract invoice information and map it to supplier invoice fields.

At the time of the interview, Walker expected Ledger Agent to enter the early-adopter environment on November 5, 2026, with a first production deployment targeted shortly before Christmas. The team was also exploring payables agents for account mapping and tax codes, alongside collections workflows.

The rollout therefore combines tools already processing information with agents still being evaluated for broader responsibilities.

DHL’s Agent Push Extends Beyond Finance

DHL Supply Chain’s HappyRobot partnership, announced in November last year, shows where agents are already working elsewhere in the division.

DHL said deployments across several regions handled appointment scheduling and driver follow-ups, alongside urgent warehouse coordination. The agents conduct phone and email interactions, with deployments targeting hundreds of thousands of emails and millions of voice minutes annually.

More recently, DHL’s Logistics Trend Radar 8.0, released in September 2024, identified agentic AI as a major development in logistics. It also emphasized safeguards and human oversight as systems take on more responsibility.

The finance project brings that shift into accounting workflows. Updating supplier information or acting on financial transactions requires clear rules about access and approval, even when building the agent itself takes little time.

Where FedEx and UPS Are Applying Agents

FedEx is connecting logistics intelligence to procurement workflows. In an expanded ServiceNow collaboration announced in May, the companies outlined plans to embed FedEx Dataworks intelligence into source-to-pay processes.

The proposed applications include using shipment-delay signals to trigger workflows that help resolve disruptions. Supplier assessments and logistics insights would appear within the procurement systems customers already use.

In June, UPS said that it had deployed agentic control-tower capabilities at customer sites. These combine data and predictive models to flag and help resolve disruptions across multi-carrier networks.

These applications focus on coordinating the movement of goods and responding to exceptions. DHL’s finance team is applying agents to the records and controls behind the business, starting with bounded tasks such as supplier-data updates and invoice ingestion.

An Agent Can Change Data, So Who Checks It?

Game said agents that enter or modify transactional data need permissions tied to the appropriate Oracle roles. The team also needs to control which employees can access those agents.

Information-security colleagues are involved in development discussions, including work around preapproved invoice processing. Game said “everything has to stop with a human for now until we can prove that it works.”

Game cautioned against “marking our own homework by building an agent that says that the agent did a good job.”

The executives discussed confidence-based processing as a possible next step after proving consistent performance. They did not describe the example thresholds raised in the conversation as an established policy for autonomous transactions.

The existing Oracle enterprise resource planning (ERP) environment provides a foundation for this work. Game said introducing agents had not required changes to the underlying data models, allowing the team to focus on how agents operate within established processes.

Better Controls Before Headcount Targets

Walker warned that leading with efficiency can narrow the board’s expectations: “all they’re going to think is headcount reduction, right?” He wants stronger financial controls and better-quality information to lead the case for agents, with efficiency gains following.

Faster development still leaves operating costs to account for, including model usage and the work needed to test and supervise an agent.

The control process must also produce evidence an auditor can assess. If an agent applies a policy or performs a financial check, the team needs to establish how its actions are recorded and reviewed.

For DHL Supply Chain’s finance operation, scaling agents means deciding which tasks they can perform and who remains responsible. Walker described the team’s focus on deployment and organizational change: “We do not want to be developers, right? We want to deploy change management.”