Computing is approaching the era of "nearly human-level agents," as noted by a major AI executive. However, businesses face significant hurdles in redesigning workflows and determining access rights for agentic AI programs. A recent report from a database company found that only 19% of organizations have deployed AI agents, and most of those deployments are limited in scope.
Chief financial officers often express three core concerns: Can the system be controlled? Does the output provide real value? And what are the costs? To address these, enterprises should adopt three best practices before implementing agents: governance (control), thorough evaluation, and starting with small, focused projects.
Governance: Controlling data access and identity
The first practice is governance, which starts with controlling what data an agent can access. An AI agent is a program that goes beyond simple prompting; it can connect to corporate databases, execute code, invoke external systems, and chain together multiple actions. The golden rule is to do no harm with data.
For example, the women's health app Flo serves 75 million users with personalized assessments. The company must ensure that no user ever sees another user's sensitive information. A proper governance system can selectively enforce which data is available to which users or agents. Asset manager Franklin Templeton faces a similar challenge when sending portfolio reports to clients. The system must deterministically enforce that responses are tailored to the individual, not just suggested in a prompt.
Governance also involves connecting questions to the right data sources. Instead of making the agent a transactional chatbot, design it to automatically gather multiple relevant pieces of data. For instance, the online car-buying platform Edmunds built an internal agent called Edmunds Mind. It merges listings, traffic data, and demographics so users can ask, "Which dealerships are underserved?" The agent takes many steps to ensure high-quality responses, relieving the user of having to supply all context.
To implement governance, organizations use a data catalog. This provides a single pane of glass for IT admins to see all data sources, external tools, and invoked functions. It also enforces identities for both the agent and the user, tracking them throughout the process to keep data segmented.
Evaluation: Ensuring correctness at every step
The second practice is rigorous evaluation of the agent's output. In Flo's case, the evaluators were physicians, not programmers. While software developers built the orchestration system, medical experts judged whether responses were accurate and appropriately nuanced. Evaluation must be ongoing and happen at multiple levels — not just the final answer, but every intermediate reasoning step. If something is off, the agent should be rolled back, refined, and redeployed in a continuous loop. This approach has helped Flo bring a differentiated, high-quality application to market. Companies that can effectively evaluate agent outputs are six times more likely to get into production.
Start small to manage costs and complexities
The third concern is cost, which often works itself out once governance and evaluation are in place. But cost must be considered from the start. The key is to start small and build at a manageable pace, testing each component individually before scaling up. Trying to replace an entire enterprise system (like an ERP or general ledger) with a single agent is ill-advised. Instead, break the workflow into atomic pieces.
For example, the convenience store chain 7-Eleven faced a problem: service technicians often lacked the correct manuals when repairing equipment on-site. The company built an agentic "super assistant" that gave technicians access to all historical issues, manuals, and specs. This eliminated the need to call a colleague for help. As a result, first-time fix rates improved by 25% and time to repair dropped by 40%.
Another example is Baylor University, which uses agents to review recordings of every call with prospective students. The agent analyzes decision factors and conversation details, freeing human advisors from taking comprehensive notes. This enables the university to learn more about its organization from customer interactions than ever before.
Real-world payoff and implementation timelines
While it is still early for industry-wide ROI figures from agentic AI, promising anecdotes exist. Franklin Templeton automated portfolio analysis and identified over $15 million in new product opportunities. Key performance indicators like first-time fix rates and time to repair show measurable improvements. The most critical factor for speed is data readiness. If data is clean and well-organized, an agentic system can be built and deployed quickly — even within an afternoon. If data is messy, the project's bottleneck becomes data cleanup. Organizing data upfront increases project velocity and enables developers to run fast.
As one expert put it, this era of agentic AI is still in its early days, analogous to the early 2000s web — companies invest but don't yet fully understand the ultimate value. Those that prioritize deliberate design in governance, evaluation, and incremental growth will be the ones that successfully bring agents into production.
Source: ZDNET News