1. An agent does the work. A chatbot talks about it.
Agentic AI has become one of the biggest topics in tech this year. In your own words, what separates a true AI agent from a chatbot or a traditional automation workflow?
The difference is what happens after the system understands the request.
A chatbot provides an answer. Traditional automation executes predefined steps. An AI agent can take an objective, reason over context, determine what needs to happen next and act within defined boundaries.
The distinction becomes clearer when something unexpected happens. Traditional automation generally needs an exception path someone has already designed. An agent can evaluate the situation, adapt its approach or recognize when it needs to hand the decision to a person.
I've spent three decades working with enterprise platforms and workflows. For most of that time, software helped people understand what was happening, but people still had to move the work forward.
Agents change that relationship. But intelligence alone doesn't give a system permission to act. Someone still has to define its authority, set the limits and own the outcome.
We're moving beyond systems that know what needs to happen toward systems that can actually do the work.
2. Real value shows up where work waits
Where are you seeing agentic AI deliver real, measurable business value today, and which use cases are still more hype than reality?
The value shows up when AI changes how much work a team can execute, not simply how fast an individual task gets done.
We're seeing that firsthand at KMS. With Evergreen Telehealth, we used an AI-native engineering approach to take a FHIR-first healthcare platform from concept to production in 75 working days. That included three applications, four live integrations and two production agents. AI supported requirements, development and testing, while engineers retained responsibility for architecture, quality and critical decisions.
Another example is a maritime software company modernizing a legacy platform that had been running for more than 20 years. In a six-week agentic development engagement, a small team delivered 35–45% of the MVP scope in just 809 team hours. The original MVP had been estimated at 6,000–8,000 hours using traditional delivery methods, and the engagement demonstrated an estimated 8–10x feature-completion productivity multiplier compared with a traditional team of similar size. Agents also generated more than 5,400 unit tests.
That's what I mean by greater delivery leverage. We're not just generating code faster. We're changing how much work a team can accomplish while keeping people accountable for quality and outcomes.
Beyond engineering, I see strong opportunities in IT service management and document-heavy operations, where agents can gather context, reconcile information and move work forward across systems.
Where the hype gets ahead of reality is the assumption that deploying more agents automatically creates more value. It doesn't. Putting autonomous agents on top of fragmented data and disconnected systems won't fix the underlying problems.
The number of agents deployed isn't a business outcome.
Faster tasks are not faster outcomes. The real measure is whether the entire workflow delivers a better result.
3. A pilot proves the model works. Production proves the organization works.
What does a successful agentic AI deployment look like in practice, from pilot to production? What are the most common reasons projects stall or fail?
A pilot tests whether the technology can perform a task. Production tests whether the organization can operate the entire workflow.
Those are different challenges.
A pilot often works with controlled data, limited integrations and a team that can manually resolve exceptions. In production, the agent encounters real systems, inconsistent information, permissions, failures and consequences.
That's where projects stall. The model works, but the integration doesn't. No one has defined what the agent can approve. The business hasn't assigned ownership of the outcome. Or the organization discovers that the ongoing cost of operating and monitoring the solution was never part of the original plan.
Another mistake is measuring success through adoption rather than business impact. An agent can be used frequently without materially improving the process.
I'd treat the pilot as the first stage of production. Start with one consequential workflow, connect it to the real systems, define the agent's authority and name the person accountable for the outcome.
Then measure what actually matters: time from request to resolution, cost, quality and where the work still waits.
The technology can succeed while the deployment fails. That's why the operating model has to be designed alongside the solution.
4. A human in the right loops, not every loop
As agents take more autonomous actions, how should companies approach trust, security and governance? Where should humans stay in the loop?
I'd start with a basic principle: the ability to perform an action is not permission to perform it.
An agent may have access to customer records, payment systems and operational workflows. That doesn't mean it should have unrestricted authority across them.
The practical approach is to map the decisions inside a workflow. What can the agent recommend? What can it decide independently? What can it execute? And where must a person approve the action?
Risk and reversibility matter. A routine, rules-based decision that's easy to undo is different from one that commits significant spending, changes a critical record or affects a customer relationship.
The technical controls have to support those boundaries: distinct agent identities, least-privilege permissions, audit trails, monitoring, exception handling and the ability to reverse actions where possible.
Every agent also needs a named owner. When something goes wrong, accountability can't become another question that nobody can answer.
But governance shouldn't mean putting someone in every step of every workflow. Start with tighter supervision, build evidence that the agent operates reliably, and expand autonomy within defined limits.
The goal is a human in the right loops, not every loop.
People remain responsible for the authority they delegate and the consequences of those decisions.
5. People move from coordinating work to owning outcomes
How is agentic AI changing the way teams work, and what does it mean for roles, skills and hiring over the next few years?
For years, employees have been the unpaid integration team inside the enterprise.
They find information in one application, chase approvals in another and move the results into a third. Highly skilled people spend an extraordinary amount of time coordinating work rather than doing the work that actually requires their expertise.
Agents create an opportunity to change that.
In software engineering, for example, the bigger opportunity isn't simply helping a developer write code faster. It's coordinating more of the delivery lifecycle, from requirements and development through testing and quality assurance, while engineers retain responsibility for direction and results.
That's greater delivery leverage. Teams can accomplish more complex work while concentrating human expertise where it matters most.
The skills that become more valuable are clear problem definition, domain knowledge, workflow design, evaluation of AI-generated work and the ability to set effective guardrails.
I expect those capabilities to matter increasingly in hiring and leadership development. Organizations will need people who can direct intelligent systems, challenge their outputs and understand whether they're solving the right business problem.
The objective isn't to remove people from the equation. It's to create teams where people spend less time coordinating routine activity and more time exercising judgment and delivering outcomes
6. If the workflow can cross the stack, why should the user have to?
Multi-agent systems and interoperability standards, such as MCP, are getting a lot of attention. How do you see agents working together across tools, platforms and organizations?
The real opportunity is moving from individual agents completing isolated tasks to coordinated systems executing entire business workflows.
Think about the enterprise technology stack. CRM holds customer information. ERP manages operations. Data platforms provide insight. Workflow platforms coordinate processes. People have traditionally connected the work between them.
Now those boundaries are changing. Data platforms are becoming more operational, workflow platforms are becoming more intelligent, and enterprise applications are increasingly exposing headless interfaces.
We're moving from systems that record what happened, through systems that help interpret it, toward systems that can execute the next action.
Standards such as MCP help agents discover and interact with tools and data. That's important, but access isn't the same as reliable execution.
The underlying systems of record still matter. In fact, their data, permissions, business rules and audit trails become more important when agents begin taking action.
And as agents work across organizational boundaries, authority becomes critical. One company's agent might request an action, but another company still has to determine whether that action is permitted and who is accountable for the result.
I expect orchestration to become a core enterprise capability, with agents coordinating work across systems and people supervising decisions, exceptions and outcomes.
The opportunity isn't to connect as many agents as possible. It's to make the whole system work together.
7. Prediction for Agentic AI
Looking ahead 12 to 24 months, what is your boldest prediction for agentic AI, and what one piece of advice would you give leaders starting their journey today?
My prediction is that agent authority will become a board-level issue.
As AI agents begin approving transactions, changing operational plans and taking actions across enterprise systems, boards and executive teams will increasingly need to understand what those agents are authorized to do, how their actions are controlled and who owns the consequences.
That reflects a larger shift I call the Execution Economy.
Intelligence is becoming more accessible. The competitive advantage increasingly comes from an organization's ability to turn that intelligence into coordinated action.
Two companies may have access to the same models and similar information. The difference will be how effectively they can execute, with clear decision rights and measurable results.
My advice is to start with one consequential business workflow.
Identify where the work waits, which systems it depends on and what decisions are required. Determine what an agent can recommend, decide and execute. Name the person who owns the outcome, and measure the end-to-end improvement before expanding.
Don't begin by asking how many AI agents your organization should deploy. Begin by asking what work could be done fundamentally better.
The companies that benefit most from agentic AI won't necessarily be the ones that adopt it fastest. They'll be the ones that change how work gets done.