Vivek Ahuja, VP-IT at rSTAR, spearheading business and IT transformation with a focus on manufacturing, energy/utilities and construction.gettyEvery enterprise technology leader I talk to wants to deploy AI agents. Autonomous systems that handle customer service, orchestrate field dispatches or manage workflows without human intervention. The pitch is irresistible: lower costs, faster resolution, happier customers.But here is what I have learned after deploying AI-powered solutions across utility and manufacturing enterprises: The reason most agentic AI pilots stall has nothing to do with the model. It has everything to do with the data underneath it.The numbers back this up. Gartner predicted that through 2025, at least 30% of generative AI projects would be abandoned after the proof-of-concept stage, citing poor data quality, inadequate risk controls and escalating costs as the primary reasons. A 2024 RAND Corporation study examining dozens of AI projects across industries arrived at the same conclusion: The top cause of failure was not model limitations but data-related issues, including a lack of clean, well-organized, integrated data.We have an industrywide blind spot. The conversation around agentic AI is dominated by model selection, prompt engineering and orchestration frameworks. These matter, but they are not the bottleneck. The real question most enterprises cannot answer: "Can your systems provide an AI agent with clean, connected, real-time data to act autonomously?"Why Agents Need More From Data Than Co-Pilots DoA co-pilot assists a human. It can tolerate ambiguity because a person is there to catch errors, fill gaps and apply judgment. An agent acts on its own. That distinction changes everything about what your data architecture needs to deliver.When we deployed AI-assisted knowledge bases for utility contact centers, a co-pilot could surface a relevant article and let the representative decide how to apply it. That worked even with some inconsistencies in the knowledge base. But when you want an AI agent to autonomously resolve a billing dispute or initiate a service transfer, it needs to pull from your CIS, CRM, outage management system and AMI data, all in real time, with confidence that the data is current.An agent operating on bad data does not just give a wrong suggestion. It takes a wrong action. In regulated industries like utilities, that wrong action can trigger compliance violations. McKinsey's 2023 research on data quality reinforces this: Organizations with mature data management practices are 2.5 times more likely to see meaningful returns from their AI investments compared to those without.The Three Data Blockers I See In Every EngagementAfter years of working across utility and manufacturing environments, agentic AI readiness comes down to three persistent data challenges.The first is master data fragmentation. Most enterprises have customer, asset and product data scattered across multiple systems with no single source of truth. In utilities, I routinely see customer records split across a CIS for billing, a CRM for interactions, an OMS for outage history and an AMI platform for usage data. Each system has its own version of "the customer." Harvard Business Review has reported that only about 3% of companies' data meets basic quality standards. That is manageable when a human reviews the output. It becomes existential when an AI agent is making autonomous decisions on that data.The second is integration latency. Agentic AI requires near-real-time data access, but many enterprises still rely on batch ETL processes that update systems overnight or weekly. An agent handling a service transfer needs the current account status now, not as of last night's data load. The third is data quality debt. Duplicate records, missing fields, inconsistent formats, outdated entries. For traditional reporting, this is an annoyance. For an autonomous agent making decisions without human review, it is a liability. Now imagine agents making hundreds of autonomous decisions per hour on top of that flawed foundation.Reframe The InvestmentI am not arguing against agentic AI. I am arguing for a more honest allocation of effort. In my experience, a successful agentic AI deployment is roughly 30% model and orchestration work and 70% data architecture, integration and governance work. But most enterprise budgets and project plans invert that ratio. A 2024 McKinsey survey found that companies achieving the highest returns from AI spend nearly half their AI budgets on data-related activities (cleaning, integrating, governing and monitoring data pipelines), while underperformers allocate the majority to model development and tooling.The organizations that will lead in agentic AI are not the ones with the most sophisticated models. They are the ones with the cleanest data, the most connected systems and the governance discipline to keep it that way.Where To StartIf you are evaluating agentic AI, here is what I would do before writing another line of agent code:• Audit your top three agent use cases and map every data dependency back to the system of record. If you find more than two hops or a batch process in the chain, that use case is not agent-ready.• Run a data quality baseline on the domains your agents will touch. Measure completeness, accuracy, duplication and timeliness. Publish the results internally.• Stand up an MDM workstream in parallel with your AI initiative, not after it. Treat master data as a first-class deliverable with its own milestones, owner and budget line.• Rebalance your AI budget. If more than 50% of your spend is going to model development and orchestration tooling, you are underinvesting in the foundation that determines whether those models work in production.Agentic AI will transform enterprise operations. I believe that. But the transformation will not start with the agent. It will start with the data the agent depends on. The leaders who recognize this and invest accordingly will be the ones who move from pilot to production while everyone else is still troubleshooting why their agents keep getting it wrong.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?