Ask any enterprise executive to describe one of their AI agents, and you'll likely get a clean answer: "It summarizes documents" or "It handles first-line support tickets." But ask them how that agent is governed, measured or corrected, and you'll often hear silence. That silence is the story of agentic AI right now.
I talk regularly with data and AI leaders from leading global companies about deploying agents, not pilots, in production. What separates the organizations pulling ahead isn't a better model or a bigger budget. Rather, they've stopped treating agents as prompts to be engineered and have started treating them as roles to be filled.
Prompts describe tasks; job descriptions describe roles #
A prompt tells an agent what to do in a given moment. A job description goes further: What does this role require, how does it fit into the organization, and who is accountable for its output?
Many companies have skipped the job description step. One of the top concerns tied to agentic AI is maintaining human oversight and preventing rogue actions. It raises the question: Who is managing the agents, and who trained those managers to do it?
Leaders I’ve spoken with build their agent programs around the following three deliberate steps, each of which maps to a step organizations would take with any new hire.
Hire deliberately
Before writing a line of agent logic, ask what task deserves automation. At a music technology company, that meant auditing workflows to find the most repetitive, highest-volume work. For this organization, a music-detection agent now classifies AI-generated tracks at a scale no human team could sustain.
At a global automotive manufacturer, an AI lead within each business division harvests use cases through structured ideation, then runs them through a prioritization framework before anything gets built. As one executive put it: "It has become so easy to make agents that teams do it first. Do you have a process? Stage gates? Metrics? Most organizations are not there yet."
Design matters as much as prioritization. Agents don't work in isolation any more than employees do, and trust starts at the onset. At a digital marketing agency, a solution architect role now sits between AI teams and business units, ensuring agents launch as governed workflows tied to real objectives, not isolated tools bolted on to a process.
Prepare the handbook
If a job description is the offer letter, data is the employee handbook. An agent without access to accurate, diverse and current enterprise data is an employee who never got onboarded. Yet data readiness remains the top barrier to enterprise AI adoption: In Snowflake's own research, 65% of companies say breaking down AI data silos is challenging or very challenging, and 62% say the same about prepping data to be AI ready in the first place. The fix isn't more data. It's structured data. "People often underestimate the importance of the data structure," says a leader at a digital marketing agency. "They assume the agents can do it all. But models don't work that way. They cherry-pick. Effective agents require very curated onboarding."
An energy solutions company learned that lesson building an agent for repair engineering. "Building the agent only takes about 20 minutes," says one leader. "But the engineering effort to gather data from scattered sources and build the embeddings it needs is a multiweek effort." The agent is only as good as the handbook it's given.
Manage actively
Hiring and onboarding are the easy parts. Managing is where programs can fall apart, because managing an agent looks nothing like managing a person. Yet it demands as much rigor.
The head of global data and AI at a healthcare technology firm describes three emerging roles for the humans overseeing this shift: the designer, who builds the process and the agent that executes it; the controller, who monitors execution and validates outputs; and the auditor, who independently vouches for the agent's work in high-stakes processes. "Doing the task itself may increasingly become optional or even a hobby," she tells me. "But human accountability will not disappear. People must remain responsible for the decisions embedded in their designs, the autonomy granted to agents and the outcomes those agents produce.” That accountability has to be measured, not assumed.
A global automotive manufacturer tests trainee agents on 170 questions across five rounds before deployment and then holds them to a simple bar: outperform a human, not merely approximate one. A music technology company's operations team performs a final human check on every agent output before it ships. And every leader I interviewed gave the same answer when I asked who's ultimately responsible for what an agent does: Humans are.
That answer is the whole point. Building trust between employees and their new agentic colleagues takes time. One leader estimated it takes four weeks to build an agent and eight or nine months to reach real adoption, because trust is earned through iteration. It’s not declared at launch.
The next org chart has AI managers on it #
A new role is emerging inside every function that deploys agents at scale: the AI manager. "The next stage is managing multiple AI agents," a leader from a music technology company tells me. "Now you are a manager, and you are going to manage these AIs to make them more efficient."
The organizations moving fastest with agents are explicit about where an agent belongs, deliberate about the data those agents depend on and rigorous about who is checking their work. The technology is ready faster than most management structures are. The gap between "we have agents" and "we employ agents effectively" isn't a technology gap. It's a management gap. Closing it starts with a simple shift in posture: Stop prompting, start employing.
I go deeper into each of the conversations mentioned here, including how a global information services provider trains its agentic workforce on 175 years of legal and tax expertise, in the full ebook “Employ Agents Effectively: A Data Executive's Blueprint for the Agentic Enterprise.”