How agents will rewire enterprise work

KPMG became the first Big Four firm to integrate Anthropic's Claude directly into its global client services platform, Digital Gateway. Anthropic has also named KPMG a preferred consultant for the private equity market, and the two organizations will jointly help PE firms and portfolio companies deploy Claude to cut development life cycles and drive value creation.

Swami Chandra Shekharan identified two classes of enterprise agentic users: technically sophisticated users who want to instruct AI through harnesses at a desktop level, and simpler web application users who need reliable long-running task completion. Core operational requirements for enterprise agentic AI include managing state and memory, retrieving context, correcting errors, escalating to humans when needed, and surfacing results reliably. Token budget management is already a real concern on long agentic jobs, and the cost calculation is complicated by the fact that LLM tokens are priced differently depending on whether they are purchased through a SaaS vendor, a coding tool, or directly from a model provider.

Rafiq argued that Microsoft, Google, and Anthropic have all recently rolled out harness-enabled AI capabilities into the enterprise, and that enterprises will begin meaningfully experiencing agentic AI within roughly a quarter from the time of recording. At that point, agentic AI will completely rewire expectations for how work happens, with security and cost questions accompanying the rollout.

Swami drew a sharp distinction between automating existing processes and genuinely reinventing work. Simply automating invoice processing resembles prior generations of automation, whereas true reinvention would involve cross-system agent actions spanning GPS data, procurement contracts, third-party pricing, and vendor risk monitoring to detect leakage, fraud, and abuse. The biggest value from AI accrues to organizations treating it as a way to create new opportunities rather than purely as an efficiency technology, and that reinvention cannot be designed top-down without input from people doing the work on the ground.

KPMG has set up safe experimentation environments where staff can test agentic capabilities using non-confidential data, addressing the limitation that employees experimenting on personal devices cannot use confidential enterprise data. The firm has also deployed a passive knowledge extraction tool that uses AI to interview subject-matter experts and capture undocumented institutional knowledge, with early results drawing interest from additional contributors inside the organization. Despite these initiatives, AI usage at KPMG and its clients remains predominantly chat or research assistant rather than collaborative work agent, and token maxing as a deliberate practice has not yet been observed at scale outside Silicon Valley.

On measurement, the speakers argued that token consumption is a lazy and wrong metric for AI productivity. Meaningful measurement requires instrumenting for specific business value outcomes such as cycle time reduction, faster vendor onboarding, and reduced vendor turnover. Without a north star metric, organizations will produce isolated one-off projects that do not move the needle, and enterprises should not use token consumption as a year-end performance metric or employee incentive.

Practical first steps recommended for enterprises include selecting one or more harnesses from options such as Claude Code, Codex, and Claude SDKs, then connecting those harnesses to organizational data by establishing connectors, context, and standardized agent skills before expecting meaningful productivity gains. Both bottoms-up employee adoption and hands-on top-down leadership tone-setting were described as necessary, and the next quarter or two was identified as the critical window for large and medium enterprises to lean into adoption.