PwC narrows AI use in audits with harnesses, measurement, and human checks

PricewaterhouseCoopers has accelerated use of large language models and other AI in audit work but is building operational controls around that use. PwC describes a layered approach: governance before deployment, tooling that structures data and agent behavior, and mandatory human oversight to validate results.

The harness is described as the scaffolding that organizes where the AI "brain" (the LLM) looks for information. PwC says the harness can be programmed with contexts, reach into multiple repositories to pull documents, and shape the user experience that leverages the model.

Human oversight and model governance

PwC emphasizes governance before releasing AI tools. The firm measures model effectiveness and model accuracy as part of that governance. Humans continue to check outputs: if an agent or model produces an incorrect result, auditors annotate the error and feed that annotation back into the agent. That loop of human annotation plus agent self-improvement is intended to make agents better over time.

PwC's assurance transformation leader, Shawn Panson, said the firm uses models for "very discrete and intentional purposes" and that it maintains "very strong governance before we release anything." He also stated plainly, "We're not there yet," in reference to achieving artificial general intelligence.

PwC works with multiple major AI vendors, including OpenAI and Microsoft, and evaluates which model performs best for specific tasks. The firm avoids reliance on a single LLM and gives employees access through a secure environment for administrative and work-related tasks. PwC trains employees on tool use and relies on AI engineers to pair tasks with the most appropriate models.

PwC's annual corporate directors survey of 561 directors found widespread concern about boards' readiness for AI. Seventy-one percent said their boards need stronger AI skills to oversee AI integration. The top AI-related risk identified was cybersecurity, data privacy, or intellectual property—cited by 69% of respondents. Other worries included significant investment without clear returns (44%) and overreliance on AI outputs that could weaken human judgment (35%).

Where PwC positions AI in audit practice

PwC treats AI as an efficiency and capability lever but not a replacement for human judgment. The harness and measurement systems aim to reduce operational mistakes and improve model performance while preserving human review where technical conclusions, audit judgments, and controls are involved. PwC's current approach combines centralized governance, tooling that standardizes and routes data, continuous measurement of model output, and human validation loops that feed corrections back to agents.