Enterprises plan six times more AI agents. Only 29% of today's agents talk to each other

CIO Dive reported on 25 September that an IDC survey of more than 400 enterprise technology decision-makers, commissioned by agentic AI company Leah, found two-thirds of organisations already running AI agents in production. They plan to increase their agent numbers sixfold by next year and to double agentic budgets. Only 29% of their agents interact with each other. Just 7% have advanced orchestration, where several agents collaborate on one piece of work. More than half rely on basic or workflow-based handoffs. Anurag Malik, Leah's president and CTO, described enterprise AI as "a thousand instruments playing at once without a conductor."

The bottleneck moved to the seams

The standard reading of agent adoption is a headcount story: every new agent retires a slice of a role. The survey points somewhere more useful. The enterprises in the survey are deploying agents department by department, and the work that carries the most value rarely stays inside one department.

IDC's Neil Ward-Dutton made the point in a statement: complex business processes do not usually belong to one department or system. He advised enterprises to manage agents as a portfolio rather than as departmental projects. A customer complaint touches support, product and legal. A contract moves through sales, finance and operations. When each function automates its own slice, the handoffs between slices stay basic or workflow-based. That is where context gets dropped and delays accumulate.

Deloitte's August study of 501 US senior managers and C-suite executives shows the same gap from the leadership side. Nearly three-quarters expect about half of their business processes to be redesigned or rebuilt around agents within four years. Only one in five say their organisation is prepared to redesign processes to run autonomously. First on the respondents' list of obstacles: "poorly documented and understood processes." Fragmented data and entrenched ways of working came after it.

The agents are ready to be deployed faster than organisations can describe the work they are meant to do.

Oversight has not kept pace

New Relic's 2026 Observability Forecast, which surveyed 2,575 IT and engineering leaders with Enterprise Technology Research, found that one in four AI agents runs unmonitored. Eight in ten organisations plan to deploy AI application observability within the next 12 months. The dashboards will arrive. A dashboard cannot say what a correct outcome looks like across a whole process, only whether one step behaved as configured.

Deloitte found that 75% of leaders see more value in human-agent collaboration than in automation alone. In that model, employees direct agents, review outputs, validate quality and decide when human judgement is needed. Separately, 61% expect most agents to be largely autonomous, with people serving mainly in oversight.

Read those findings together. Organisations intend to hand execution to agents and keep a human layer above them — one that understands the full process, owns the handoffs and answers for the result. The surveys describe that layer in detail. None of them name who staffs it.

AI does the routine work. You do the thinking. In an agent estate, the thinking is the map: which steps exist, who owns each one, what passes between them and what counts as done.

Your Next Move

1. Write down one cross-team process this month. Choose a process that leaves your team and comes back. Record every step, every handoff, every system and every sign-off. Deloitte's respondents put undocumented processes first on their list of obstacles. The person holding the only accurate map is the person any redesign has to run through.

2. Mark the seams and take them to the budget holder. For each handoff, note whether an agent, a person or a spreadsheet carries it, and what gets lost in transit. IDC found that only 29% of agents interact with each other; count how many of yours do. Bring the marked-up map to whoever controls the agent budget before the next deployment is scoped without it.

3. Propose one outcome measure and volunteer to own it. Ward-Dutton's advice was to measure results at the outcome level instead of counting deployments. Offer your manager one end-to-end number for your process: cycle time, rework rate or time to resolution. Agent counts are easy to report and say little. Ownership of the outcome across the seams is the work that keeps its value as the agent count climbs.