Technology

KPMG trains tax staff to build software with vibe coding

Pilot turns non-developers into internal toolmakers, governance burden shifts from engineering to audit trail

Images

A pilot at KPMG had tax pros working with technical staff to vibe code software. 
                              
                                EschCollection/Getty Images A pilot at KPMG had tax pros working with technical staff to vibe code software.  EschCollection/Getty Images businessinsider.com

KPMG has been running a six-week internal pilot to turn tax professionals into hands-on software builders using what staff called “vibe coding” — prompting AI tools to generate working code with guidance from engineers, according to Business Insider. By the end of the program, the firm said the tax workers had produced software that KPMG now uses, and leadership is considering permanent mixed teams that pair domain specialists with engineering staff.

The immediate appeal is arithmetic. Tax advisory work is full of repeatable workflows — data ingestion, rule-based checks, document generation, and client-specific reporting — that historically required a handoff to a scarce pool of developers or expensive external contractors. Generative coding tools collapse that queue: the person who understands the edge cases can now build a prototype directly, and the marginal cost of “one more internal app” falls sharply. For a consulting firm, that can look like a productivity gain and a way to productise know-how that previously lived in slide decks and senior staff’s heads.

But the same shift also changes where risk accumulates. When non-developers can ship software, the organisation tends to get more software than it can govern: small automations become internal products, internal products become dependencies, and dependencies become audit problems. Version control, testing discipline, access management, and change logs are not optional in a tax practice that must defend calculations years later under regulator or client scrutiny. If a model-generated script silently changes how a deduction is calculated, the question is no longer “who wrote this” but “who approved it, when, and under what controls.”

Ownership becomes murky too. In a traditional build, a team can point to a repository, an author list, and a review trail. In AI-assisted coding, the “author” is often a sequence of prompts, edits, and regenerated snippets spread across chat histories and local files. That makes it harder to prove provenance, harder to reproduce results, and easier for sensitive client data to leak into tooling that was never procured for regulated workloads.

KPMG’s pilot suggests the firm is trying to capture the upside while keeping engineers in the loop. The test will be whether those guardrails survive when the novelty fades and hundreds of small tools start competing for the same compliance attention.

The pilot lasted six weeks. KPMG says the software built by tax staff is already in use.