KPMG explains how agentic AI could reshape software development

KPMG’s 2025 paper on agentic AI in software development says coding-assist tools boost productivity by 10% to 30%, and the firm is treating agentic AI as a change in how software gets built, not just a smarter autocomplete. The shift could compress planning, coding, testing and documentation inside client delivery teams while forcing tighter governance over who approves what and when.

Why this matters inside KPMG

For consultants and auditors, the practical issue is not whether AI agents can write code. It is whether they can shorten the cycle between a client problem, a working prototype and a controlled deployment without breaking the evidence trail that regulated work depends on. Software delivery now sits inside finance, tax, operations and industry-specific consulting, not only in technology engagements.

KPMG’s 2025 paper on agentic AI in software development calls coding agents the most significant innovation in software development history. It is not the same as handing routine delivery over to machines. KPMG says the near-term gain comes from removing bottlenecks, not eliminating human oversight.

What agentic AI changes in the delivery stack

Agentic AI systems can perceive, reason, plan and act with minimal human intervention. In practice, that shifts attention from one-off prompts to multi-step workflows. Instead of helping a developer draft a function, an agent can help plan tasks, interact with repositories, coordinate tests and support deployment steps with less manual back-and-forth.

That shifts attention to coding, testing, documentation and governance. If an agent can draft code, trigger tests, summarize changes and update documentation, the work inside a delivery pod becomes less linear. The tradeoff is that more of the process must be designed up front, because autonomy without clear checkpoints is a control problem, not a productivity gain.

KPMG calls agentic AI a step-change and a leap beyond generative AI. That distinction matters for managers who have spent the last two years pushing copilots into development teams. The next stage is not just better content generation. It is software that can coordinate actions across multiple tools and adapt in real time.

The governance question is now part of the build decision

KPMG’s July 2025 paper on AI governance for the agentic AI era is blunt about the stakes: enterprise deployment only works if trust is built into the system. That is especially relevant in software development, where a small change can alter access rights, audit logs, controls testing or downstream reporting.

For KPMG auditors, the implication is immediate. Software changes affect internal controls, change management and evidence trails, which means agentic workflows cannot sit outside governance simply because they are faster. For consulting teams, the same issue shows up in operating-model redesign, where clients want speed but still need review steps that satisfy risk, legal and compliance teams.

The line between “AI-enabled delivery” and real deployment becomes visible when a tool can move faster than the approval process, creating pressure to redesign sign-off, exception handling and traceability.

How the firm is already positioning itself

In April 2025, KPMG Canada launched an Agentic AI Engine, moving from concept to internal and client-facing implementation. KPMG is building a commercial story around agents across the firm, not just in a lab.

That broader push became even more visible in KPMG’s global alliance with Anthropic: Claude would be integrated across the business and workforce of more than 276,000 people in 138 countries and territories.

KPMG has also used Microsoft and Azure in customer-story materials to position agentic AI for audit work, where the phrase “redefine audit” carries obvious implications for workflow, evidence capture and review.

What leaders need to decide before deployment

KPMG’s January 2026 paper asks whether enterprises should build, buy or borrow AI agents. Software development teams rarely need one generic agent. They need a stack of choices about how much to own, which tasks to automate and where the control points live.

A practical deployment plan inside KPMG would need to answer several questions:

• Which steps in the development lifecycle can be delegated to an agent, and which must remain human-approved?

• How will agents interact with code repositories, testing environments and deployment pipelines?

• What evidence will be preserved for audit, quality review and change management?

• Which teams need deeper software literacy so they can supervise agents, not just use them?

• Where does the firm need to build capability, and where should it adopt vendor tools instead?

Agentic AI changes team design as much as output. Early-career professionals who once assumed software literacy was optional will need to understand how agents work across the delivery chain. Senior leaders will need to decide whether faster iteration means new staffing models, different review cycles and more blended teams that combine business consulting, architecture, controls and technical oversight.