AI is here. Are your staff using it too much?

Companies are discovering that the same technology they encouraged employees to embrace is now multiplying beyond anyone's control – and the consequences are landing squarely on the desks of human resources leaders.

The sprawl nobody planned for

Not long ago, the challenge facing most employers was persuading their workers to adopt artificial intelligence. That problem has largely been solved – perhaps too well. Across industries, a new predicament is emerging: employees are building and deploying AI agents faster than anyone can track them, creating what technologists have taken to calling "agent sprawl."

The Wall Street Journal recently reported that companies including Lyft, DaVita, and GitLab are actively trying to manage the proliferation of AI agents inside their organisations – navigating the delicate challenge of reining in usage without discouraging innovation. At Fair Isaac, the credit-scoring firm better known as FICO, employees are creating dozens of new AI agents every single day, spanning everything from personal email assistants to large-scale data management tools. At kidney-care company DaVita, employees have created more than 10,000 agents in total.

The scale of what is coming is difficult to overstate. Gartner, the technology research and consulting firm, predicts that by 2028, the average Fortune 500 company will be running more than 150,000 AI agents – up from fewer than 15 just a few years ago. By the end of 2026, Gartner forecasts that 40% of enterprise applications will have integrated task-specific AI agents, compared with less than 5% in 2025.

A problem HR cannot ignore

For human resources professionals, agent sprawl is not purely a matter for the information technology department. The organisational, cultural, and legal dimensions of uncontrolled AI proliferation fall squarely within HR's mandate – and the window for getting ahead of it is closing.

Gartner survey findings show that just 13% of organisations believe they have the right governance in place to manage AI agents. Meanwhile, a 2025 report by Menlo Security found that 68% of employees are using unsanctioned AI tools, and 57% are feeding sensitive corporate data into them.

The governance gap is not just a cybersecurity headache. It creates a tangle of overlapping accountability questions that land in HR's lap: Who owns the output of an AI agent – the employee who built it, the manager who assigned the task, or the company whose data it consumed? What happens when two agents produce conflicting results and a business decision gets made on flawed information? And how should performance be evaluated when it is no longer clear how much of an employee's work is their own?

The 2026 trends report from Deel, highlighted by HRD, underscored that HR teams should assess their readiness in AI literacy, compliance exposure, and technical integration in order to operationalise AI safely and effectively – describing HR as the "architect of the human-machine enterprise."

The workforce implications

The people dimension of this shift is significant. A global survey by Workday found that while 82% of organisations are rapidly deploying AI agents, employees are already drawing clear boundaries around the technology. Three-quarters of workers say they are comfortable working alongside A.I. agents, but only 30% say they are comfortable being managed by them. And fewer than half – 45% – believe AI agents will become true members of the workforce.

At the same time, a Moveworks poll of 200 IT executives at large US companies found that non-technical, frontline employees are now driving agentic AI initiatives from the bottom up, with 91% of respondents confirming this trend. This is not a story about technologists experimenting in isolation – it is a story about the entire workforce reshaping how work gets done, often without any formal approval or oversight.

A Salesforce survey of 200 chief human resources officers, covered by HRD, found that 89%  AI agents will empower them to reassign employees to new roles, with roughly 23% of the workforce expected to be redeployed as a result of the technology. CHROs consider AI literacy the single most important skill workers will need in the years ahead, and three in four HR leaders said AI agents will increase demand for soft skills such as collaboration and adaptability.

The hidden costs

Beyond the human dimension lies a financial one that is only beginning to register on corporate balance sheets. AI agents do not run for free. They consume computing resources – measured in tokens – every time they process a request. When dozens of employees independently build agents that perform the same task, those costs multiply without any corresponding increase in value.

Gartner forecasts that global software spending will surge 15.2% in 2026, reaching $1.43 trillion, with much of that growth driven by AI-related costs that bypass traditional procurement processes. The firm also warns that more than 40% of agentic AI projects will be abandoned by the end of 2027 – not because the technology fails, but because of escalating costs, inadequate governance, and an inability to demonstrate business value.

There is also a subtler cost that researchers are starting to quantify. A term has entered the enterprise vocabulary: "workslop" – the low-quality output that results when employees are pressured to produce more, faster, using AI tools that are not equal to the task. Research cited by GoSearch found that workers spend an average of nearly two hours cleaning up each instance of substandard AI-generated output they encounter.

What employers should do

Gartner has outlined a practical framework for organisations seeking to address agent sprawl without stifling the innovation that makes AI valuable in the first place. The core principle is that blocking or restricting AI agents is not a sustainable strategy – employees who cannot use sanctioned tools will simply reach for unsanctioned ones, creating far greater risks through what is known as "shadow AI"

The recommended approach involves six interconnected steps: establishing clear governance policies that define who can build agents and on what platforms; creating a centralised inventory of all agents in use; managing agent identity and access permissions carefully, with regular reviews to retire redundant bots; governing what information agents can access and for how long; monitoring agent behaviour continuously for anomalies; and building a culture of responsible AI use through training and shared best practices.

Some organisations are already putting these principles into practice. DaVita has built an internal platform that gives it the ability to scale back AI spending when needed while concentrating resources on the highest-performing agents. Lyft has developed an IT-approved system for sharing the instruction sets that tell its AI agents how to handle specific tasks, reducing duplication and improving oversight. Anthropic, whose Claude platform underlies many of these tools, says it has introduced features to help IT administrators manage role-based access, spending controls, usage analytics, and audit logging.

The opportunity inside the problem

For HR departments willing to lean in, the current moment of controlled chaos carries genuine opportunity. As reported by HRD, Deel's 2026 trends report positioned HR as uniquely placed to design environments where people and AI can genuinely thrive together, rather than merely coexist.

That means HR professionals need to be asking hard questions now: Do we have a policy governing which AI tools employees can use and build? Do we know what data those tools are being fed? Are our managers equipped to evaluate work that was partly produced by an agent? And are we prepared for the redeployment and reskilling that will follow as agents absorb more routine tasks?

The organisations that will be best positioned to capture the opportunity AI presents are those that start building the governance, literacy, and cultural infrastructure today – before the sprawl becomes unmanageable.