What Should You Delegate to AI/your agent?

Introduction

AI adoption in most organisations did not begin with a grand redesign of how work gets done, but with simple tasks: employees used LLMs to summarise documents, conduct research, prepare presentations, and produce first drafts. The tools worked, so their use spread.

The productivity gains were easy to see. Research involving two major consulting firms found that junior consultants could complete desktop research that once took days in 30 minutes, and analyses that once took weeks in hours. But the research also revealed a less visible consequence: faster production did not remove the need to validate outputs, catch errors, and maintain quality standards, those responsibilities remained with the people.

The recent Deloitte Australia case illustrates what can happen when those responsibilities are not effectively carried through the workflow. In 2025, the firm used generative AI in preparing a report worth $440,000 for the Australian government, and fabricated references and an invented Federal Court quote made it into the final document. The report was later corrected, highlighting a familiar challenge with generative AI: the model can produce the work, but the organisation remains responsible for its accuracy and quality.

Verification is only one part of that responsibility. As AI spreads across an organisation, they also need to control what it can access, what it can do independently, when human approval is required, and how its actions are recorded. Where no system exists to manage these controls consistently, the decisions fall to individual employees and managers, working across disconnected AI tools and teams. The problem, then, lies at the organisational level: we have introduced a new way of doing work without fully redesigning how that work is managed.

The Human Governance Layer

For nonprofits, that problem is already taking shape. According to the 2026 Nonprofit AI Adoption Report, 81% of nonprofit organisations use AI for ad-hoc tasks and yet 76% have no AI governance policy. AI, in other words, is spreading faster than the organisational systems for managing it.

For now, people are filling that gap. Consider a composite portrait of Meera, an operations manager at a mid-sized development nonprofit:

• She starts her morning reviewing an AI-assisted donor report, checking its claims and sources before it goes to the funder.
• A program team asks whether beneficiary data can be used with a new AI tool. Another needs access to internal data for an AI-assisted analysis. Neither request fits neatly within an established process, so Meera has to decide who can access what with the relevant teams.
• By midday, an error surfaces in an AI-generated document that has already passed review. She now has to trace where it entered the process, who reviewed it, and whether it appears anywhere else.
• Before the day ends, she documents what happened, clarifies the rules for future use, and makes sure IT and program colleagues are working from the same guidance.

No single task is overwhelming. The problem lies in their accumulation. Employees decide how to use AI. Managers review outputs and handle exceptions. IT manages access and security. Legal teams establish what is permissible. In the absence of clear, embedded controls, people are having to carry out these functions manually, often case by case. As AI spreads across functions, these people become the connective tissue between otherwise isolated tools and organisational controls.

Organisations cannot scale AI indefinitely by adding another human check, another handoff, or another rule for employees to remember. At some point, the controls themselves have to become part of how the work gets done.

Every Workforce Needs an Operating System

As AI takes on more of the workflow, organisations need a way to manage how that work moves between agents and people. A governed AI workforce provides that operating layer, with five capabilities:

• Workflow orchestration determines how work moves between agents and people, including what happens next and when a task should be escalated.
• Organisational memory gives agents access to approved policies, decisions, documents, and prior work rather than rebuilding context through prompts.
• Role-based permissions limit agents to only the information and systems required for their roles.
• Human oversight and auditability determine where approval is required and preserve a record of what the AI did and what informed its work.
• Policy enforcement and model flexibility keep organisational rules consistent even when the underlying AI models change.

Put these controls in place, and Meera's day changes. The donor report reaches her with its sources and AI actions recorded, and verification can be required before the work moves to the next stage. Requests for beneficiary data are governed by predefined permissions rather than decided case by case. If an error surfaces, she can trace it through the workflow instead of reconstructing what happened across teams. Meera remains in the loop, but her attention shifts from carrying all controls herself to handling the judgments and exceptions that require her.

For mission-driven organisations, this division of labour is consequential. Donor requirements, beneficiary information, internal policies, approval authorities, and program knowledge all place boundaries around how work gets done. If AI is going to participate in that work, those boundaries need to travel with it.

From AI Tools to an AI Workforce

That is the problem CURA is designed around. Rather than another chatbot employees interact with independently, CURA acts as the control system between AI capabilities and organisational work: coordinating agents, maintaining organisational context, enforcing permissions and policies, inserting human approvals where necessary, and preserving a record of how work was completed.

With that layer in place, organisations can manage AI at the level of work rather than one tool at a time. A chatbot gives an employee a tool for completing a task. A control system manages how that task moves through the organisation, including which agents and people are involved and which controls apply along the way.

This also changes the decisions organisations make about AI adoption. The question becomes which work should remain human, which should be AI-assisted, and which can be delegated to agents operating within defined boundaries.

Managing Intelligence, Not Just Building It

For the first phase of generative AI, much of the competitive conversation concerned intelligence: which model could reason better, write better, or handle a larger context window.

As AI gets embedded in organisational workflows, another question becomes just as important: how well can the organisation manage that intelligence?

Every organisation is likely to develop its own division of labour between people and AI. The organisations that make this transition successfully will be those that can decide where autonomy belongs, put controls around it, and manage human and AI work as one operating system.