A recent Harvard Business Review study revealed only 6% of companies fully trust AI agents to autonomously run their core business processes.
That number should give every leader pause, not because the technology has failed but because the way we're deploying it has.
The deficit in trust isn't a capability problem. Today's AI agents can handle complex tasks and synthesise information at speed.
The problem is implementation: the absence of the guardrails, shared structures, and organisational context that make any collaborator — human or artificial — reliably effective.
Chief Product Officer, Asana.
This matters because it reframes the entire conversation. Autonomy has become the dominant goal in discussions of agentic AI — the idea that the most valuable agent is one that requires the least human involvement. But autonomy without context is a liability.
The real unlock is building the shared infrastructure that allows humans and agents to work together in ways that are visible, accountable, and productive.
The single-player problem
To understand why trust in AI agents remains so low, look at how most organisations are actually deploying them. The dominant model today is ‘single-player mode’: AI tools assist one person, in one conversation, disconnected from the wider business - and detached from the workflows that give work its meaning and direction.
It may accelerate output, but it simultaneously amplifies the coordination burden with more outputs to review and more decisions to align. Without shared structure, the productivity gains are consumed by the overhead of keeping everyone on the same page.
The deeper problem is organisational. When leadership can't see what an agent is doing, why it's doing it, or how its output connects to business priorities, trust inevitably stays low. And when trust stays low, ROI becomes difficult to track, adoption remains uneven, and the technology gets siloed in pockets of individual use rather than scaled across the workforce.
The agent becomes a personal productivity tool rather than a genuine driver of organisational transformation.
Agents as teammates, not tools
The alternative is ‘multi-player’ mode, where agents operate as participants in shared plans, workflows, and accountability structures. AI agents are increasingly capable of handling the repeatable, template-driven work that consumes time - but they cannot replicate taste.
For example, the strategic judgment that comes from understanding an organisation's history and ambitions. Or the ability to weigh competing priorities and make a call that reflects not just the data but the culture. These capacities remain distinctly human and the goal of effective AI deployment should be to protect and amplify them, not to replace them.
That means rethinking the role of agents entirely. Rather than tools that individual users pick up and put down, agents should function like members of the team: visible to their colleagues, accountable within shared workflows, and improvable through collective feedback.
And this learned knowledge builds institutional memory that doesn't reset with every conversation, ensuring the AI Teammate gets smarter for everyone and preserving valuable context as an organisation scales.
Context as a competitive advantage
For agents to function as genuine teammates, they need more than data. They need layered context to enable an understanding that connects individual actions to broader organisational goals.
Think of this as a ladder. At the bottom rung, an agent understands the immediate task: what needs to be done, for whom, and by when. Most agents operate here, and it's useful, but it's also where outputs are most likely to be technically correct yet strategically misaligned.
Climb higher, and agents understand how tasks connect to workflows across departments and divisions. Higher still, agents grasp organisational strategy: which goals are being pursued, which trade-offs have been made, and why. At this level, they can exercise judgment aligned with organisational intent, rather than just following individual instruction.
Most companies are stuck at the bottom rung, not for lack of AI capability but because they haven't built the infrastructure to climb. Climbing from task-level awareness to genuine organisational understanding requires investment in three interconnected areas: controls, checkpoints, and context:
Controls - the role-based permissions that mirror human boundaries, making agent impact legible and manageable.
Checkpoints - the shared frameworks where stakeholders can review agent reasoning, intervene early, and course-correct in real time — an auditable collaborator running on rails.
Context - making organisational knowledge explicit through defined projects, clear ownership, stated goals, and transparent priorities.
The path forward
The agentic enterprise won't be built on model capability alone. It will be engineered through clear structures, stronger governance, and a commitment to the kind of organisational clarity that allows both humans and agents to do their best work.
The organisations that get this right won't just have smarter AI, they'll have AI that knows how to be a good teammate. This unlocks something more durable than just incremental productivity gains - an AI that understands how their business actually works, that gets better the longer it operates within their systems, and that frees people to focus on the work that only humans can do.
In a landscape where every organisation has access to roughly the same underlying models, the quality of the context you provide — and the infrastructure you build to support it — may turn out to be the most important competitive variable of all.
We feature the best IT automation software.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit