The signal
ZDNET has reported on how enterprise leaders are approaching AI agent rollout, drawing on comments from Scott Likens, global chief AI engineer at PwC, and Lasherelle Morgan, senior vice president of AI innovation and acceleration at NBCUniversal.
The message is not simply “move fast” or “slow down”. It is more useful than that. Teams are being pushed to experiment quickly, but the work only becomes safe when humans remain in charge, the process is understood, the data is ready, and governance is matched to the risk.
That is a good description of where many organisations now are. AI agents are no longer just a research curiosity or a demo feature. They are being pointed at workflows, decisions, calendars, content, analysis, customer service and internal operations. Once that happens, the question changes from “can the model do something impressive?” to “who owns the outcome if it acts inside the business?”
Why the human is still the control layer
One phrase in the ZDNET piece is worth sitting with: the human is not just “in the loop”. The human is the loop.
That matters because agentic systems can make automation feel deceptively tidy. A tool can be given a task, collect information, trigger actions and report back in a neat summary. But the neatness of the interface does not remove the need for judgement. It can hide it.
For leaders, the practical test is whether a person still understands the goal, the allowed actions, the evidence being used, and the point where the agent must stop. If those boundaries are not clear, the organisation has not gained a colleague. It has added a fast-moving ambiguity machine. Delightful. Exactly what every governance meeting was missing.
Start with the process, not the tool
The strongest point in the report is also the least glamorous one. Before introducing an agent, write down the process.
That sounds basic, but it is where many AI projects quietly wobble. If a workflow is already messy, unclear or owned by nobody, adding an agent does not fix the operating model. It exposes it. Morgan is quoted as saying AI is good at “blowing up a bad process”. That is not a reason to avoid AI. It is a reason to do the boring process work first.
Good agent rollout starts with repeatable work, available data, clear ownership and a real user pain. What is someone doing repeatedly that they hate doing? Where is the work slow because people are copying information between systems? Where is judgement required, and where is it not? Which data is the agent allowed to see? Which actions can it take without further approval?
Those questions are not bureaucracy. They are the difference between a useful deployment and a mess with a nicer dashboard.
Experiment quickly, but make the architecture deliberate
The ZDNET article describes PwC running AI experiments in short cycles, including one-day and five-day experiments. That pace is sensible. AI systems are moving too quickly for eighteen-month transformation theatre.
But short cycles do not mean casual foundations. PwC’s example also points to the need for data architecture, access control, context and telemetry. In plain English: people need to know what the agent is doing, what it is using, what it is allowed to touch, and what can be learned from the work afterwards.
That is the tension most organisations now need to manage. They need lightweight experiments because the technology is changing quickly. They also need enough structure that successful experiments can become safe operating capability rather than a drawer full of clever prototypes.
Match governance to the blast radius
Not every agent needs the same level of control. An agent that helps arrange a lunch is not the same as an agent that changes supplier data, sends customer messages, approves refunds, updates a CRM, or drafts advice used in a regulated decision.
The useful governance question is the one Morgan reportedly uses: what is the blast radius?
If an agent gets something wrong, who is affected? Can the mistake be reversed? Does money move? Does a customer see it? Does personal or commercially sensitive data leave the organisation? Does the agent create a record that someone may later trust as fact?
Those questions make governance practical. They also stop the two common mistakes: treating every AI use case as too risky to touch, or treating every workflow as safe because the demo looked convincing.
Why this matters in practice
AI agent rollout is becoming an operating-design problem, not just a software-adoption problem. The winning organisations will not be the ones that either ban the tools or throw them everywhere at once. They will be the ones that know where speed is useful, where control is non-negotiable, and where a messy process needs to be fixed before an agent is allowed anywhere near it.
The practical starting point is simple. Pick one repeatable process, name the owner, map the data, define the allowed actions, set the stop points, and decide what evidence would prove the agent is helping. Then move quickly inside that boundary.
Before the agent gets the keys
If AI agents are starting to touch real work inside your organisation, map one workflow before you automate it. Name the owner, the data, the decisions, the permissions, and the point where the agent must hand back to a person.
That is where useful adoption starts: not with another pilot, but with a process clear enough to trust.