When you're moving too fast, you don't get the governance put in at the right point, or at all. Suddenly you're all in on things, making fast decisions that sound very good with high confidence, but fundamentally inaccurate.
Trent LezerCDAO, Senior Director, Data & Analytics @ Daikin Applied AmericasEnterprise AI programs are reaching production faster than the data foundations they depend on. A team that layers an agent over its existing sources of truth inherits every ambiguity inside of them, and the agent reports its conclusions with the same assurance, regardless of whether or not they hold. Verifying the data behind these sources is the work that often gets deferred when adoption moves fastest, and governance then has to be fitted retroactively around earlier decisions about which systems the agent can reach. For a data leader, the question now is whether the environment around the model has defined limits, who sets them, and who answers for what comes out.
Trent Lezer is CDAO and Senior Director of Data & Analytics at Daikin Applied Americas, which builds custom HVAC systems and the control software that runs them. His 25 years in technology include building a governance program from scratch at Greystar and setting up a governance and analytics office at Trimble, where he trained stewards and governed master data. He has also run large global teams and multi-million dollar budgets, and governance has been part of his work for longer than agentic AI has existed.
"When you're moving too fast, you don't get the governance put in at the right point, or at all. Suddenly you're all in on things making fast decisions that sound very good with high confidence, but fundamentally inaccurate," says Lezer. For him, governance runs wider than data quality. It covers which technologies get adopted, where they run, and how they get monitored once they're live. These choices are all about scope, and they get harder to revisit after a system is in production.
Governance before the first project
Applying AI the same way across a company this size takes a structure that sits above any single project. Ashish Srivastava, Chief Digital Officer at Daikin Applied Americas, describes the program built to hold that structure. "That's why we built an enterprise framework called DaikinIQ that lets us apply AI consistently, responsibly and at scale by aligning our people, processes, data, technology and governance," says Srivastava.
Two bodies do the work inside that framework. An AI Council at Daikin Applied sets direction and settles disputes, and Lezer points to an AI Center of Excellence that owns execution. Anyone who wants to do AI inside the company goes through it, and the choice of technology, the selection of the project, and the implementation all fall under that review. He treats that routing as a filter on what gets built at all. "We're not just adopting AI for the sake of adopting AI," says Lezer. "We're adopting it for the right problems that are part of our strategic goals."
Lezer sorts a good share of what arrives into automation, conventional business intelligence, and statistical analysis, and he's clear the underlying problems are still worth solving. Daikin Applied had made the foundational investments and did deep groundwork years before AI reached the agenda, so the data lake and the curated layers were already running when the first agent arrived. "A lot of the things that we field in the Center of Excellence are actually projects that don't need an AI solution," he notes. "They look amazing. If you solve that problem, it would probably be valuable. But it's not an AI-centric problem."
Scope and accountability
The word "shipment" means one thing to a site manager and another to someone in finance. For the site manager, it describes goods physically leaving the property. For finance, it's a revenue event, timed to the customer's receipt of the order. Lezer treats the shipment case as what happens when one shared term carries two meanings, and as his reason for resisting a single assistant that serves every function. "You've got the same term creating confusion, and asking a generic top-level agent a question about shipments can lead to misleading answers," he explains. "Either the agent knows to ask clarifying questions, or you've got separate agents that these two audiences would go to, so that they always get the one that they expect."
Business users ask Lezer for an assistant with access to everything, and his answer is to scope what it can reach. A document store holds early drafts alongside the version that was finally approved, and an assistant deployed without onboarding reads them all with equal authority. In his view, what should be excluded varies by case. A contract search shouldn't reach superseded drafts, while the history of an engineering document is often the reason for the search. "If you don't control that access, then you give up the agency and the ability to make sure that the answer is correct," says Lezer. "You've got to make sure that the access gets to the final versions of your data, rather than just saying the more data, the better for the agent."
Access controls settle what an agent can reach and leave open who answers when it gets something wrong. "Every AI system we run has a named owner, and a person stays responsible for its outcomes. Our AI Center of Excellence tracks accuracy and keeps tuning the models, and the AI Council sets how far any agent is allowed to go on its own," Srivastava explains. "For most of our work, the AI advises and a human decides. Where it does more, it stays inside limits our AI Council has approved, and someone is always watching and able to step in."
Storage, access, and boundaries
Lezer treats where data sits and who can reach it as one question. He describes the organizing layer, the one recording what exists and how it's arranged, as valuable on its own. The access controls over it are what makes the arrangement safe to open to an agent. On that reading, storage sits inside the foundation a team builds for AI. "Storage isn't just a place where you put something. It's also the rules about who can access it," says Lezer. "It's that combination of storage, access control, and security that becomes an entire holistic story."
Inside Lezer's own team, agentic AI now helps maintain the curated layers the business runs on. The group is taking on contractors and new hires who don't carry the institutional history, and a model arrives already holding context that would otherwise take hours to collect from colleagues. What the team gets out of it depends on what they put to it, and the judgment about what to build stays with them. "It accelerates us because it brings that base of knowledge and puts it at our fingertips so that if we're asking the right questions, the entire iteration of deciding what's valuable and what needs to be done becomes accelerated," says Lezer.
Asked for advice for a data leader at a less mature company, he returns to how the environment gets built. "Our goal is not to try to teach people how to prompt better. Our goal is to create ecosystems and frameworks so that the agent knows the guardrails and can't get outside of them," Lezer concludes. "You're building it in such a way that it can't go outside those boundaries."
