The next shift in AI isn't from humans to machines. It's from AI completing individual tasks to AI taking on parts of a workflow. And that changes how leaders should think about the work itself.
You’ve probably heard a lot about AI agents recently.
You may also be wondering what agents are, what they do, and more importantly, whether opportunities exist within your organization for them to unlock and/or drive meaningful value.
Today, most of us currently use AI by giving it a discrete task. Ask an AI tool to summarize a report, for example, and it summarizes the report.
An agent is different. Instead of completing a single task, an agent is designed to take responsibility for a series of tasks in pursuit of an outcome.
If you give an agent a research objective, it can determine what it needs to investigate, search sources, compare what it finds, identify gaps, and pass conclusions onto another agent – or person – or the next step.
This distinction matters because the bigger opportunity from AI is moving beyond speeding up individual tasks to re-imagining how work gets done, and which parts of that work are best performed by people, machines, or some combination of the two.
We’ve experienced this firsthand.
At Edelman, we build our AI systems around the way our teams actually get things done. Rather than taking a complex job like developing a strategy and asking one AI model to complete the work, we deconstruct the work into its components.
Consider the development of a communications plan.
Understanding the brief is one job. Researching relevant company, category, consumer and cultural insights requires several others. Finding tensions and white space in that evidence is yet another step in the process, followed by developing recommendations
And throughout the process, human judgment and accountability are essential to determining whether the thinking is any good.
Once you understand those components, you can build specialized agents around different parts of the process and connect them into a larger workflow.
Importantly, that doesn't remove the expert from the process, rather it changes where their expertise is most valuable.
People can more deeply interrogate the research, challenge conclusions, and add the context and client nuance that AI doesn't have. They redirect the work, and ultimately ensure the output is strategically relevant. Instead of spending expertise equally across every step of a process, we concentrate on where human judgment creates the most value.
Increasingly, we're applying the same principles when we build AI systems for clients: start with the work they are trying to improve, understand the expertise and information that make that work good, and only then decide how and where to deploy AI.
The first question shouldn't be “What can we automate?” It should be “What are the repeatable workflows and should we accelerate, augment or delegate?”
Some work is a natural fit for an agent: gathering information, monitoring changes, comparing sources, summarizing evidence, routing work or producing a first recommendation.
Other work depends heavily on context, relationships, judgment, ethics, or reputation. In those areas, human involvement may be imperative.
And sometimes the best answer isn't an agent at all.
