CIOs Race To Make AI Outputs Defensible Before Agentic Systems Scale

Boards are demanding proof that AI investments can pay off, and CIOs are absorbing the pressure. Some have two and a half years to show measurable returns; others have six months. Yet many enterprises are already pushing agentic AI into finance, accounting, marketing, software testing, and internal workflows before they have the governance, data architecture, or human validation systems needed to trust what those systems produce. The result is a fast-growing control gap inside the enterprise, where AI can execute more than the organization can explain, verify, or defend.

Ajay Paul is Chief Sales Officer at SIMARN Solutions, an IT services and strategic consulting firm that partners with Fortune 500 companies on workforce management, data analytics, and digital transformation. He brings more than 35 years of experience in enterprise technology, with prior leadership roles at HCLTech, Gartner, Persistent Systems, and ISG spanning sourcing strategy, CRS modernization, and CIO advisory engagements across financial services, energy, travel, and CPG. He said the current wave of AI is fundamentally different from any prior technology shift because it no longer requires human expertise to design the process being automated. "AI is upending all of that. It's saying you don't even need to know what the process is. 'I will define the process, I will modify the process, I will implement the process behind the scenes,'" Paul said.

Previous generations of enterprise platforms, from CRS and EHR systems to Salesforce, still depended on human process design. AI removed that dependency. It can ingest data, define the process, and produce results without human direction. But Paul warned that outputs could be deeply unreliable if the underlying data is flawed. "If your data is horrible, your results are horrible," he said. "But if you don't know that, you're going by the fantastic results that AI produces, and they could be completely faulty."

• The internal governance gap: Paul pointed to a structural blind spot in how enterprises approach AI governance ownership. Most organizations concentrate guardrails on customer-facing outputs while internal teams deploy agents across finance, accounting, and software testing with little oversight. "They figure governance is for outside the organization," he said. "Internally, they haven't put a lot of thought into governing what AI does for them." He expects the consequences to surface soon. The Air Canada chatbot case, where an ungoverned AI offered a fare that the airline was forced to honor, previewed what could happen at a much larger scale.

• Data architecture as precondition: Paul described a conversation with a CIO whose organization had data architects at the business unit level but none at the enterprise level. Siloed data that looked clean in isolation produced suboptimal results when fed to large language models and agentic systems operating across the full organization. "If there isn't a single cohesive data model, your large language models and agentic AI tools will be suboptimized," he said. "The companies that can govern their data are the ones who will do well. The ones who don't trust their own data are going to struggle."

• Culture resists controls: Internal product teams pushed back on enterprise governance because they viewed it as slowing them down. Paul witnessed workshops where product engineers directly challenged governance teams on what they should be allowed to do. "Product teams want to work at the speed of the customer. They think governance is rules and regulations that hamper their ability to move. But the risk is increasing underneath." The tension between scaling discipline and speed is not new, but agentic AI compresses the timeline for consequences.

The accountability question cuts deeper than governance. Paul described sending a routine email through Copilot for rewriting. The output was polished but unrecognizable. The recipient wrote back: 'This doesn't sound like you.' "I couldn't stand behind it," Paul said. "The words Copilot used were not my words." He saw the same dynamic playing out across enterprises where people were adopting AI tools to produce reports and recommendations without verifying or owning the output. Eventually, he said, someone would ask whether the results had been validated. And the answer would be no.

Paul offered three directives for CIOs preparing to scale AI beyond experimentation. Establish governance structures before expanding deployment. Invest in enterprise-level data leadership, starting with a chief data architect and chief data officer. And redesign work so that AI augments the people who must remain accountable for outcomes, rather than replacing them.

"Are you going to put a machine accountable for your financial results in the marketplace? I don't think CIOs and CEOs are going to do that," Paul said. "You need those people in the loop. Use AI to augment human skill sets rather than replace human skill sets. That is what I believe is going to happen, once people start having to defend the results."