You talk, and a voice on the other end interviews you. You’re asked a question, given room to trail off and double back, and the questions keep adjusting to what you just said.
You might never notice you’re speaking with an agent. Behind it, a second agent is reading along against a rubric, working out what it should ask next.
Improving president Michael McCullough and director of consulting Michael Ell build agents like this for a living. They told a room of tech leaders at CIOCAN’s Peer Forum last month that these agents are moving from handling single tasks to running entire workflows end-to-end across a business, with people watching instead of driving.
The capability is further along than you might think. What hasn’t caught up is everything around it, the governance, the tooling, and the question of who answers for the work once an agent is the one doing it.
From doing tasks to owning workflows
McCullough and Ell sort agents into three kinds. Capture IQ, the voice interviewer, falls into the category they call value stream agents.
There are chatbots like Copilot and Gemini that are good for personal productivity. Then there are horizontal agents, the same tool serving the whole company, like one that lets anyone pull from a shared base of company knowledge. A value stream agent is more specific, aimed at one line of work, like agents that write software or handle the work of a legal or finance team.
Improving runs one of these on its own staff. When a project wraps, Capture IQ interviews the team and turns the conversation into a sales case study, from interview to draft to approval to publication. Work that used to take a person walking through a script now runs through the agent.
“Fundamentally, instead of using a keyboard as input, we created agents that you’d actually talk to,” said Ell.
The three kinds describe what an agent does. A separate question is how much autonomy it has, and McCullough mapped that out too.
At the near end, companies just experiment with AI. Then agents handle one task under close supervision. Next, they run across departments, with people monitoring instead of approving each step.
At the far end is full autonomy, where, as McCullough put it, agents are “initiating and carrying them out end to end” inside guardrails.
Most companies, he said, are nowhere near that end.
The bridge nobody’s standing on
The technology is the easy part, but McCullough warns that everything around it is where companies get stuck.
When an agent takes over a piece of work, it becomes two new jobs instead of replacing one.
Someone has to tell the agent exactly what to produce, and someone has to check what it hands back. McCullough said that’s where things jam up now, and many companies haven’t figured out how to do either well.
The people best placed to define that work, the ones who plan projects and write the business case for them, don’t have good software for it the way developers do. Some companies hand them an integrated development environment (IDE), the workspace developers write code in, and hope it works, which McCullough said it doesn’t.
His own staff have pushed back, too.
When an agent does the job someone used to do, it raises the uncomfortable question about what that person is now for.
The real obstacle sits underneath all of it. Agents need clear direction from people who understand the business, and technical guidance from people who understand how agents work. Those people are often not on the same team.
One is knowing the business well enough to say what a good result looks like. The other is knowing how to get an agent to produce it. McCullough said each side tends to wait for the other to bridge the distance, and while they wait, nothing moves.
Who answers for the output
A lot of companies already know they need their data in order before any of this works, a point worth its own discussion and one I’ve written about before.
Agents raise another question that’s far from settled. When the agent carries the work, who answers for it?
When a person runs a cross-departmental process, you know who to ask when it goes wrong. Run the same process through an agent and that clarity becomes murky. The responsibility is still there, but the owner is harder to identify.
In practice, it probably falls to the CIO as the technology leader, who gets the responsibility without the authority that should come with it. The work spans departments the CIO doesn’t run, so answering for the result means answering for things they can’t directly change.
Losing control of something you’re still accountable for is a problem CIOs already know in another form.
Raju Vegesna, chief evangelist at Zoho, who warns that running your business through a platform you don’t own means slowly losing ownership of the value it produces. He calls it the difference between a customer and a hostage.
Agents extend that concern from data ownership to workflow ownership, raising the question of who controls the work once software carries out the process.
Working backwards from the customer
That makes ownership more than a governance question. If an agent is carrying work across departments, someone has to decide what the work is ultimately meant to change.
Dave Aeri, CIO and CTO of Shoptravel Group Inc. and a national board member at the CIO Association of Canada, framed that test around the customer.
For him, the measure of technology work is about whether it changes how the business operates and reaches the customer.
“If that outcome impacts horizontally and vertically throughout the organization, plus it impacts the customer, you’ve done something bigger than just what your role is,” he said.
That is a useful way to think about a job that now includes answering for agents. A CIO who measures success by organizational impact is already in the habit the agent era demands, because an agent’s work shows up as impact across the business, not as a tidy deliverable with one owner.
Aeri pairs that with a caution about speed.
“Even if you define a use case today, it could fundamentally change within 90 days,” he said.
An agent can get better and better at carrying work from one end to the other, and the question of who owns that work still has to be answered by the people around it. McCullough and Ell build these agents and work alongside the companies deploying them, so the gap they point to is one they’ve seen up close.
That opening voice was convincing because you couldn’t tell who, or what, you were dealing with. The same goes for a workflow once an agent is running it end-to-end. The ownership isn’t obvious, and it won’t assign itself.
A CIO who doesn’t decide in advance who answers for the agent’s work is the one who ends up answering for it by default.
Final shots
- The difficult part of agentic AI is connecting the people who know the business to the people who know the agents. It’s an organizational problem, and no vendor sells a fix for it.
- A CIO inherits responsibility for an agent’s work without the authority to change how it runs across departments they don’t control. The org chart hasn’t caught up to the way the work now moves.
- The technology resets on a roughly 90-day cycle, which makes any decision about how far to let agents run a decision a CIO will be constantly remaking. The ownership question is the one worth settling first.