Advanced Micro Devices (AMD) designs processors and other semiconductor products used in data centers, personal computers, gaming systems and embedded devices. Its products also support the computing demands of AI applications.
Its 2025 annual filing reported $34.6 billion in revenue, including $16.6 billion from its Data Center segment, up 32% from 2024.
AMD’s investment in its AI strategy includes expanding its workforce and acquiring specialist expertise. Research and development spending rose 25% to $8.1 billion in 2025. AMD attributed the increase primarily to higher employee-related costs as it expanded headcount to support that strategy.
The company also reported acquiring teams with expertise in software compilers, machine learning, photonics and AI inference. These investments span the hardware and software needed to develop and run AI applications.
This article examines these two internal AI use cases:
- Automating Pre-Silicon Verification Triage: Engineers must identify actionable design failures among the results of roughly 1 million nightly simulations.
- Accelerating Customer Order Inquiries: Specialists must assemble accurate order and supply information quickly enough to respond to customer questions.
We begin with the engineering work required to distinguish a real chip-design problem from an uninformative test failure.
Automating Pre-Silicon Verification Triage
AMD uses GitHub Copilot to analyze overnight chip-design simulation logs, identify known issues and propose fixes before engineers begin their morning review. The application addresses the manual sorting required to determine which failures need investigation.
Before manufacturing, engineers run virtual tests to check whether a chip behaves as intended. Failures can indicate a design defect, a problem with the test or an issue already understood by the team. Harry Foster, chief scientist for verification at Siemens, explained that verification oversights can require multimillion-dollar manufacturing reruns to correct designs and disrupt product cycles.
Combining multiple chiplets into an integrated system increases that workload. AMD corporate fellow Alex Starr said the complexity had created roughly four verification engineers for each hardware designer. Repeatedly sorting known failures consumes specialist capacity that could otherwise support investigation of new design problems.
AMD consolidated its hardware and software repositories on GitHub Enterprise before introducing Copilot into verification. GitHub’s customer account describes an agent that checks simulation logs against historical issues and proposes fixes. AMD also worked with the provider to adapt the tool to hardware description languages such as Verilog, addressing the context general-purpose AI models initially lacked.
The agent performs the initial triage overnight, allowing engineers to begin with sorted results and proposed fixes. Engineers still need to investigate new failures and validate that changes resolve the underlying design problem.
The customer account reports automated triage of more than 1 million nightly simulations and zero manual morning triage time. It provides no baseline hours, total engineering review time after deployment, triage-accuracy measure or chip-quality outcome, and its broader productivity claims cannot be attributed solely to this workflow. The reported 20,000-plus engineers use GitHub Enterprise; the number using the verification agent is undisclosed.
A separate paper by AMD verification engineers describes human review of AI-generated assertions before formal testing. That process concerns a different application and does not establish how engineers approve the nightly agent’s fixes or escalate unresolved failures.
For chip-development leaders, the reported reduction in sorting is a starting point for evaluation. Measuring the unresolved failures left each morning, whether proposed fixes pass established tests and whether defects recur would help determine whether the agent releases engineering capacity while preserving verification quality.
Accelerating Customer Order Inquiries
AMD introduced a generative AI assistant to help employees investigate changed delivery commitments and answer customer order inquiries faster. The tool analyzes order allocation and supply availability information to identify potential causes and recommend action.
When a customer-facing ship date changes, supply chain specialists must explain why to planning and customer teams. An AMD account published by SAP AppHaus says the investigation could require up to 14 manual verifications. Identifying an allocation problem or supply shortage helps teams decide how to respond to the disruption and what to tell the customer.
The application emerged alongside AMD’s consolidation of its operations and those of the acquired Xilinx business onto SAP S/4HANA, beginning in January 2023. AMD IT worked with SAP’s Lighthouse generative AI program to develop the assistant.
A specialist enters a sales order number through the interface. SAP Business Technology Platform connects the assistant to the ERP environment, where it analyzes allocation timing, constrained quantities, available supply, scheduled delivery and material availability. The output supports investigation and corrective action; the published workflow does not show automatic changes to order commitments.
An architecture presentation identifies an OpenAI GPT deployment and an anonymization component. AMD subsequently expanded access to customer operations and sales teams for self-service and introduced access controls restricting sensitive order information to authorized personnel.
The company estimates that the assistant saves 3,000 staff hours annually and reports shorter inquiry response times. Its target was to reduce inquiry response times from 20 to 30 minutes to under two minutes. A later partner article describes a 90% reduction and 3,120 annual hours saved as achieved, but provides no measurement period or calculation method.
AMD does not publish a measured current response time, active-user counts, answer-accuracy results or customer-service outcomes; the differing savings figures therefore cannot be reconciled or treated as demonstrated financial returns. The architecture materials also leave the model version, anonymization rules and employee validation or escalation process unspecified.
For leaders considering a similar deployment, the evaluation should pair response time with answer accuracy and the frequency of further specialist investigation. Sampling answers against ERP records would test reliability, while tracking first-contact resolution would show whether customers receive sufficient information.
These cases point to three practical tests for enterprise AI investments:
- Evaluate the full verification handoff — Track triage time alongside unresolved failures, accepted fixes and defects found in later tests.
- Test order answers against operational records — Sample allocation and delivery responses against ERP data and record inquiries requiring further investigation.
- Calculate capacity from actual use — Use observed inquiry volume and time saved after review, then identify how employees use the released hours.