How many agent coworkers are you responsible for? And how many are running outside of your control?
89% of IT leaders say their company has at least one AI agent in production, according to our latest survey of 300 technology leaders for Governing the Agentic Workplace. Most of those agents are built in-house: only 4% of leaders with agents in production rely mainly on outside vendors to build them.
But what happens after launch, when the person who built an agent leaves, the process changes, or the agent needs a different scope? The work an agent does used to have a named owner. How does that change when an agent (or a fleet of agents) is now charged with the same task? When nobody owns that work, IT is usually the team left to track it down and fix it.
In the extreme case, one company Retool’s CEO David Hsu recently described had built roughly 500,000 agents, more than its own headcount, with no idea where the money was going. Those agents aren’t accountable for paying that bill. IT is.
Most companies aren’t at that scale, but unknown agent sprawl is easy to imagine. Agents are quickly becoming a default way of working—Gartner predicts (opens in a new tab) that 40% of enterprise apps will be integrated with task-specific AI agents by the end of 2026. Among leaders with agents in production, 42% say teams own agents in their domain with some central oversight, 27% say one team owns them all, 21% say teams manage agents independently, and 10% have no formal structure. Half of all leaders named agents making decisions with no clear owner as a top concern. In July 2025, an AI coding agent deleted SaaStr’s live production database (opens in a new tab) during a code freeze, despite explicit instructions not to make changes. The freeze was a policy, but nothing in the agent’s permissions enforced it.
Every agent should have an owner and a defined scope, with access tied to both through guardrails IT sets. The organizations running the most agents are already moving those rules out of individual review and into shared controls, because one team can’t approve hundreds of agents by hand. AI’s made work easier for some, but for most IT teams it has meant the same headcount with far more to manage.
Onboarding and governing AI agents the way you would a human workforce solves two problems at the same time:
- It maximizes benefit by making agents easier to manage and measure.
- It minimizes risk by making sure the right agents reach the right data, and nothing more.
- It controls cost by giving IT visibility into what each agent spends and what it returns.
Does every one of these business problems even need an AI agent?
Among 268 leaders with agents in production, 28% leave it to a central AI governance committee, 26% to the requesting business unit, and 24% decide case by case. Only 3% let anyone build an agent.
Centralizing the decision making on whether a business process needs an agent at all is a good place to start. Experts like the IT leaders we surveyed can help determine whether the process really requires an agent or can be solved with a workflow. A central review also surfaces duplicates, so teams can share an agent that already works instead of building ten that do the same thing. A workflow follows fixed steps and doesn’t need the same accountability structure.
Agents are more expensive than workflows, and making an informed choice up front can save a business money down the line. In the example above, roughly 95% of those 500,000 agents were better suited to workflows.
AI agents work best with the right information
As agents take on more real work, the old ways of assigning and tracking that work will need to change.
An org chart is an easy comparison to reach for, since it shows at a glance who at a company is accountable to whom. But most org charts live with HR, and they don’t capture what an agent needs to do its job: context, instructions, and access to the right systems.
McKinsey has argued that org charts will give way to “work charts, (opens in a new tab)” built around tasks and outcomes rather than reporting lines. Agents can only work with the data, systems, and tools they’ve been given. A work chart gives IT the ability to assign and supervise an agent, its task, the systems it needs. “Treat AI like a new hire” is well-worn advice at this point, but it still holds true as more businesses scale their agent fleets.
The companies building agents describe the same needs a manager would. OpenAI says agents on its Frontier (opens in a new tab) platform need shared context, onboarding, feedback, and clear permissions. In Anthropic’s multi-agent research system (opens in a new tab), each subagent needs an objective, an output format, guidance on tools and sources, and clear task boundaries.
Both companies manage agents in ways that resemble an org chart, too. OpenAI’s guide to building agents (opens in a new tab) describes a manager pattern, where a central agent delegates to specialists, and a decentralized pattern, where peers hand work to each other.
Examples of how companies are managing AI agents
The companies furthest along are giving agents the structure people already have: a manager, an identity, and a scope for their activities.
There will ultimately be as many unique solutions to managing AI agents as there are companies. But the important thing is that they’re managed—with each one having an owner, a clear scope, and access that matches that scope.
Why ownership matters
Governing agents is more than organizational hygiene. Getting governance wrong—or worse, not implementing it at all—is expensive.
43% of technology leaders in our report on AI costs from earlier this year are already spending more on AI than they budgeted, and agentic workflows use 5 to 30 times more tokens per task than a basic LLM query, according to Gartner (opens in a new tab). That bill usually falls on technical leadership. In our State of AI Governance report, 32% of CTOs, CIOs, and CISOs said engineering leadership or the CTO is held accountable for incidents caused by AI-generated tools, more than any other single answer.
Delegating ownership lets a company measure how effective the agents are at delivering what they’re supposed to, and it’s where accountability lands when one does something nobody intended.
You can’t measure what nobody owns
Agent ROI can be unpredictable. At one Fortune 10 retailer, a pricing agent that photographed competitors’ shelves produced close to nothing, while a simpler automation monitoring competitors’ earnings calls produced an estimated $50 million. Someone had to watch both agents and cut the one that wasn’t working.
Of the IT leaders we surveyed who have agents running in production, 70% measure success by time saved, but only 43% track business outcomes like revenue or cycle time. Another 18% track usage without success metrics, and 12% can’t measure success at all. Time saved only counts when it shows up elsewhere in the business.
Unowned agents can also run unchecked. One AI-powered out-of-office responder cost $10,000 a day before anyone noticed, and just 5% of CIOs, CTOs, and CISOs in Retool’s State of AI Governance report said they have full visibility into AI tools in production.
If AI agents can’t be held accountable for their actions, who is?
13% of leaders with agents in production said one took an unintended action in a production system in the past year, and 7% don't know whether they had an AI incident at all. An AI agent can’t be held legally liable for its actions by a regulator or a court, so responsibility falls to the people who deployed it—IT teams and other technical transformation leaders.
We can expect more government regulation around AI agent accountability as they become more commonplace. Singapore’s framework for agentic AI (opens in a new tab), launched in January, says each agent’s identity should be linked to a person, a department, or a supervising agent, with human approval for high-stakes or irreversible actions. The National Institute of Standards and Technology (NIST) released a concept paper (opens in a new tab) in February 2026 that applies identity and authorization standards to agents.
Still, more guidance exists today than actual law. But it shows where requirements are heading, and an agent with no owner and no defined access won’t meet them without the right governance and accountability structures in place.
How to protect your business as AI agent fleets grow
Most IT leaders we surveyed haven’t settled how agents reach company data. 40% of leaders with agents in production set up data credentials one agent at a time, and 20% have had credentials shared insecurely to get an agent working (something that shouldn’t ever happen). A few practices make agents easier to manage at any scale:
- Decide whether each process needs an agent, or whether a fixed workflow would do the job for less.
- Give every agent a named owner and a defined scope before it launches, and update both when people or processes change.
- Set access rules once, on the data. Write table-, column-, and row-level policies on the data source and check every query from every app, workflow, and agent against them, instead of managing separate credentials for each agent.
- Review agents on what they produce, and use that information to inform decisionmaking on costs.
None of this requires slowing down. In the WSJ Tech Council Summit, McKinsey’s Kate Smaje (opens in a new tab) argued for bending the cost curve rather than shutting AI off, so usage rises while cost per unit falls. Proper AI agent management is how you tell which agents earn their place on that curve.
Structuring work for the next hundred AI agents (and beyond)
AI was supposed to lighten IT’s workload, not add hundreds of agents to review by hand. The organizations that keep up will be the ones that set AI agent management standards early. 55% of leaders with agents in production already lack the staff to govern agents at scale, and 54% say business teams move faster than IT can review.
Retool gives IT one governed platform for the apps, workflows, and agents teams build. Because Data Access Enforcement applies IT’s policies at the data source, a new agent inherits those rules the moment it launches, without its own set of credentials or a separate review. That frees IT’s time for the decisions that still need a person, like which processes need an agent and who owns each one.
Learn how 300 IT, engineering, security, and AI leaders are building, scaling, and securing agents in Governing the Agentic Workplace. We also recently opened up the waitlist for our enterprise agent harness, which gives agents the access they need to do real work across your business. Sign up today.
