Much of the conversation around agentic AI focuses on autonomy.
How much work can an agent complete on its own? How many decisions can it make? How far can organizations remove people from the process?
Those questions make sense, but they can also obscure where businesses are finding value today.
Across marketing, health care, financial services and retail, speakers at SAS Innovate described a more practical division of labor. AI agents handle structured execution, repetitive analysis and time-consuming workflow steps. People remain responsible for goals, strategy, ambiguity, exceptions and decisions with meaningful consequences.
That does not make human involvement a temporary limitation that organizations will eventually engineer away. It makes judgment an operating control – and potentially one of the most valuable parts of the process.
Agents excel when the work is defined
The clearest boundary appeared in a session exploring whether AI agents will take over marketing departments.
The answer was no.
Cameron Fisher, a software developer at SAS and Mac Carlton, a solutions architect at Vercel, demonstrated a different future – one where agents take on structured, repeatable work while marketers retain the judgment behind it.
During the session, an agent interpreted a marketer's written instructions and whiteboard sketch, located the relevant objects in the system, and began automatically constructing a customer journey.
The demonstration also highlighted where human judgment still matters.
When the agent encountered a potential issue in the marketer's instructions, it didn't blindly proceed. Instead, it proposed an alternative and requested approval before proceeding. The agent handled the search, assembly and execution. The marketer provided the context and made the final decision.
That same division of labor extended beyond the demonstration. Fisher and Carlton described agents supporting work such as audience creation, customer journey orchestration, data analysis and quality assurance, while marketers remained responsible for business priorities, campaign strategy, budgets and creative direction.
The technology accelerated execution, but it did not replace the human role in making consequential decisions.
Human work moves closer to risk and value
Health care shows what that division of labor can look like when the stakes are higher.
Tamas Bosznay of Katalyze Data discussed agentic workflows that could support diagnostics, patient communication and clinical documentation. One example used several specialized agents to transcribe an appointment, summarize the conversation, extract medical concepts, create clinical notes and identify follow-up actions.
The purpose was not to replace the practitioner. It was to prepare better information before the practitioner had to review it or act.
“If it removes the low-value-added tasks from the actual workforce so that they can focus on the more risky and more value-added tasks, that definitely can help,” Bosznay said.
That may be one of the more useful ways to understand what agentic AI changes.
AI does not simply remove work. It changes where people spend their time.
Tasks that involve collecting information, summarizing records, checking documentation or following repeatable steps can be delegated to agents. Tasks that involve clinical judgment, sensitive communication, uncertain evidence or patient-specific risk remain with qualified people.
The human role becomes more concentrated around the moments where experience, accountability and context matter most.
Regulated industries are starting with assistance
In financial services, that division of labor becomes a question of how much autonomy organizations are willing to allow.
Adolfo Lopez of JPMorgan Chase described most current agentic AI work as “assistive or delegative.” Agents help employees improve processes and support decision-making, but they do not operate with complete autonomy.
Andrea Pholman of Allianz pointed to high-volume, lower-risk claims as one place where that approach can work. Faster processing can improve the customer experience without handing complex cases entirely to an AI system.
That starting point reflects the realities of regulated industries.
Banks and insurers manage decisions with financial, legal and reputational consequences. A system that mistakenly flags fraud, denies credit or handles a complex claim incorrectly can cause real harm. Even if the technology can perform more of the process, the organization still needs to decide where to grant autonomy.
Lopez summarized the immediate opportunity as automating routine and time-consuming processes while keeping people in control of the decision. Know-your-customer workflows provide a useful example.
Agents can gather information, identify missing documentation and reduce the administrative work required to develop a customer profile. Employees can then focus on the judgment and actions that follow.
The goal, at least for now, is not maximum autonomy. It is finding the level of autonomy an organization can deploy, verify and govern with confidence.
Human oversight also needs better design
Keeping a person involved does not automatically make an agentic system trustworthy.
The question, then, is not only where humans stay in the loop. It’s how that oversight should work.
Many workflows rely on a rigid yes-or-no checkpoint. A person receives a request with limited context, approves or rejects it and then receives another nearly identical request later.
As the volume grows, the review process becomes noisy and repetitive. People can develop approval fatigue, turning oversight into another bottleneck instead of a meaningful control.
SAS' Josh Beck, Diana Maris and David Weik explored a more contextual approach to human oversight.
Their framework compares a new request with previous requests, decisions and surrounding information. If the system has seen and approved a genuinely similar action under similar conditions, it may proceed. If the circumstances differ, the request moves to human review. Organizations can also establish firm limits – for example, requiring human approval for any purchase above a defined amount, regardless of prior decisions.
This creates a more selective role for people.
Humans do not have to approve every routine action indefinitely. They can focus on unfamiliar, higher-risk or materially different cases while the system learns from previous decisions.
That is an important evolution. Human oversight cannot simply mean putting an approval button in front of every agent action. Effective oversight will require context, memory, risk thresholds and clear escalation rules.
Agents can scale execution, but operating models still matter
Coppel’s experience with personalization at scale reinforces another part of this division of labor: technology does not resolve organizational choices.
The Mexican retailer uses analytics to support more precise, continuous customer engagement across its retail and financial services businesses. But David Benabib, who leads CRM at Coppel, emphasized that scale depends on more than having the right platform.
Teams still need clear business goals. Creative and analytical employees need a shared understanding of the customer. The organization must decide whether it will optimize for individual product sales or long-term customer value. Those are operating-model decisions, not tasks an agent can settle independently.
Agents may help analysts understand content, execute more campaigns and measure performance. They can increase the volume and speed of personalization. People still decide what the organization is optimizing for and whether the resulting customer experience supports that goal.
As execution becomes easier, those choices become more visible – not less important.
The new human role is not passive
It would be easy to interpret this pattern as agents doing the work while people watch.
That is not what the sessions described.
The emerging human role requires active ownership. People define the objective, supply context and determine acceptable risk. They review exceptions, resolve ambiguity and remain accountable when a decision affects a customer, patient or business.
In some cases, people will also design and govern the systems themselves. They will decide:
- Which tasks are structured enough to automate?
- Which outputs can be independently verified?
- When an agent must escalate.
- What information a reviewer needs.
- Which actions always require human approval?
- How the organization monitors outcomes over time.
These responsibilities demand more than basic familiarity with AI. They require domain expertise, critical thinking and an understanding of how automated actions connect to business consequences.
Autonomy is not the only measure of progress
Organizations should not evaluate an agentic system solely by how much human involvement it removes.
A more useful measure is whether the system assigns each part of the work to the participant best equipped to handle it.
Agents can process information continuously, execute repeatable steps and coordinate activity across systems. People can interpret uncertain situations, make value-based tradeoffs and take responsibility for consequential outcomes.
As agentic systems mature, the boundary will continue to move. Some tasks that require review today may become safe to automate tomorrow. New capabilities may also create new risks and exceptions that demand human attention.
The sessions at SAS Innovate do not settle the long-term debate about AI and employment. But they do reveal how organizations are introducing agents now.
That does not make human work irrelevant. It moves people toward the decisions where judgment, context and accountability matter most.
