The Orchestrator Is a Person: Who’s Accountable When AI Agents Do the Work?
Search "AI orchestrator" and almost everything that comes back is software: frameworks and supervisor agents coordinating other agents. The orchestrator most organisations are actually missing is a person. Agents are being added faster than anyone is defining who's accountable for what those agents do, and the vocabulary for that missing role, "agent boss," "AI agent owner," is spreading faster than the role itself is being defined. This is the companion piece to our recent article on what makes AI authentic. That piece set out what a production-grade system needs: evaluation, monitoring, a disclosed human role. This one names the person who has to own that bar.
The human oversight role is being formalised, under a lot of different names
Microsoft's 2025 Work Trend Index, drawn from 31,000 workers across 31 countries, defines the "agent boss" as someone who builds, delegates to and manages agents, and finds that leaders expect their teams to be training agents (41%) and managing them (36%) within five years. Tellingly, 78% of leaders are already considering hiring for new AI-related roles, against just 33% considering headcount reductions. The title is arriving well before the accountability that should come with it, which is exactly the gap this article is about.
Agent scale is about to outrun governance by a wide margin
Gartner predicts the average global Fortune 500 enterprise will have more than 150,000 AI agents in use by 2028, up from fewer than 15 in 2025, while only 13% of organisations currently believe they have the right agent governance in place. It's worth noting that Gartner's own first recommended step for managing this is establishing agent governance, ownership and lifecycle policies, before anything else on their list. The scale of that gap between agent growth and governance readiness is difficult to overstate.
Leadership, not workforce readiness, is the actual constraint
McKinsey's Superagency in the Workplace research found that almost every company is now investing in AI, yet only 1% believe they've reached maturity with it, and the biggest barrier to scaling isn't employee readiness, it's leaders not steering fast enough. Our own view, based on where we see this play out with clients, is that the specific steering gap is evaluation: having clear metrics for what an agent's performance should look like, and the discipline to pull scope back when it doesn't.
Don't let the vocabulary do the thinking for you
The software industry took the word "orchestrator" for itself, leaving the human role with vaguer titles like "boss" and "owner." That's worth resisting, because an orchestration framework routes tasks between agents. It cannot decide what "good" looks like, whether a given outcome is acceptable, or when an agent should be shut down. That judgement is the human orchestrator's job, and it needs defining before the agents multiply, not after. In the same spirit, naming an "AI agent owner" is happening faster than anyone is defining what that person actually has to answer for day to day, and the order matters: decide who owns outcome-accountability for each agent-touched process before you name the role, not the other way round.
There's a useful discipline hiding in all of this, and it comes down to three things worth getting in place before scope grows any further. First, define who owns outcome-accountability for each process an agent touches. Second, set the evaluation bar that agent has to clear, in specific terms, before it earns more scope. Third, keep the willingness to pull scope back when an agent stops clearing that bar, rather than treating evaluation as a clean-up job once something has already gone wrong.
The orchestrator's job isn't running the agents
It's defending the bar for what "good" looks like. Without that bar clearly defined, HR ends up as the last line of defence, catching problems that should have been caught by design long before they reached anyone's inbox. Every ungoverned agent touching an employee process is a trust liability from the moment it gets something wrong and nobody can say why, and that's not a hypothetical risk given how fast agent numbers are projected to grow over the next two years.
None of this is an argument against scaling AI agents. It's an argument for scaling them in the order that actually holds up: define the accountability, set the bar, evaluate against it honestly, and only then let scope grow. The organisations that get this right won't necessarily be the ones with the most sophisticated agents. They'll be the ones who can say, clearly, who's answerable for what those agents do.
Contact us
