Who owns your AI agents after they launch?

A colleague at a major utility is releasing 200 agents next quarter. Most of them are being developed outside of IT, in places like marketing. Super-impressive. But when I asked him who owns the agent maintenance — what happens to it when the model updates, when the policy changes, or when the person who wrote it moves teams — he didn’t have a good answer.

We’re so focused on building, we aren’t thinking about who owns AI agents after they ship.

Kana surveyed 225 senior leaders at large U.S. enterprises in May. 70% already run custom AI agents on real marketing work. Only 3% run none. The “should we build” question is settled, and we can move on to the more interesting ones, like… what happens next?

That’s where it gets muddy. About 40% of those leaders say the chief AI officer should own agentic marketing. Among AI leaders, it’s 52%. Marketing executives pick their own function or a shared model. Which means in a lot of companies right now, two capable groups are each assuming the other one has it.

Both camps have a real argument. The centralizers will tell you that agents handle customer data, they carry regulatory exposure, and that no single function can be trusted to police its own output. True. The marketers will tell you that nobody in a central AI group knows whether an agent’s tone is right, whether the offer is current, or whether the segment logic still matches how you go to market. Also true.

Split the question. Central owns access, data, and the model layer. Marketing owns instructions, tone, and whether the thing is still saying something true. Both of those need names attached, and right now, most organizations are leaving it fuzzy.

Ivanti asked 1,500 IT professionals about agent ownership in February and March. 85% said a named owner exists for every AI agent. Only 42% said ownership is clear — a 43-point gap. 

Governance shows up for the launch

The same research found that 65% of organizations run a review before an agent is deployed. Governance shows up for the launch. Then oversight shifts to a quarterly rhythm while the agent keeps working every day.

There’s a related habit worth checking in your own stack. Ivanti found that permission sprawl starts on day one, because organizations spin up agents by cloning a human user’s profile. In marketing, that means your campaign agent may be walking around with the CRM access of whoever set it up. That was fine back when it launched. Is it fine now, and who would notice if it wasn’t?

What we can learn from the software world

From the 1940s into the 1960s, code was treated as something you did once and then put on a shelf. When organizations started running critical systems continuously, that approach fell apart. 

At the 1968 NATO conference in Garmisch, engineers from a dozen countries compared notes and found they were all stuck on the same problem — software gone out of date. What emerged was the lifecycle we still teach today: requirements, design, build, test, deploy, maintain, retire. 

Maintain and retire got added because the industry learned the hard way that software doesn’t maintain itself. In 1980, Bennet Lientz and Burton Swanson studied 487 organizations and found maintenance consumed roughly half the software budget. Half. What should have gotten more attention was the breakdown:

The largest category was perfective work, meaning the requirements had changed. Next was adaptive, meaning the environment around the software had changed. Fixing actual defects was the smallest of the three. Most of that work came from the world moving around software that was written correctly the first time.

Now read it with an AI agent in mind. Your content agent was written in March against a March offer and positioning. Your SDR agent learned an ICP that predates the pricing change. Your brand voice guardrails were tuned against a model version that got deprecated over the summer. None of that is a defect. It’s maintenance, and someone needs to say what changed and what to do about it.

Picture the agent introducing itself at your next stand-up. “Hi, I’m the offer agent. I was written in March, against a promotion that ended in June, and I’m still sending.”

Recommended Articles

Most marketers say AI works. Few can prove it.

Oct 30, 2026

The risks of optimizing SEO for AI over users

Oct 29, 2026

The latest AI-powered martech news and releases

Oct 29, 2026

5 questions to ask yourself

Start with one agent you already have in production and answer these five questions about it:

• Who writes the instructions for the agent?
• Who revises them, and on what cadence?
• Who watches model releases and translates what changed into something your team can use?
• Who tracks output drift, and against what baseline?
• Who decides when this agent gets retired?

Most teams I talk with can answer the first one right away. The other four go unanswered.

What I’ve seen work is spreading those five across seats that already exist.

• Marketing ops usually takes instructions and revision cadence, because they own the workflow anyway.
• Brand takes drift, because they can say whether the output still sounds like you.
• Your AI or platform team takes model releases because they’re already tracking them.
• Retirement is the one that usually goes unclaimed, so give it to whoever owns the budget line. They’re the ones who’ll notice it’s still running.

None of that requires new headcount. It requires named accountability — a much cheaper thing.

You also get to choose the unit. Per campaign? Per platform? Per agent? A small team with a handful of agents can assign a person to each agent. A big one with workflows spread across the whole stack needs a different shape, probably per platform, with the drift question handled by sampling instead of review.

Salesforce has started putting language around this with an agent development lifecycle and roles like Agent Supervisor, and other vendors will follow. The tooling will show up. The org design part is yours.

Build the habit while the number is small

That’s good news, because habits are easier to build at eight agents than at 80. The software industry needed a crisis and a decade to learn that shipping is the front end of something much longer. We get to skip that part.

Take two questions into your Monday: Who owns your oldest agent? When did anyone last look at what it’s producing?