As AI agents move from experimental curiosity to operational reality, the question is no longer simply whether to deploy them, but how to deploy them responsibly. A new article from The Institute of Chartered Accountants in England and Wales (ICAEW) sets out six practical safeguards that organisations should build into any AI agent deployment from the outset. For Chartered Accountants working across every sector and jurisdiction, these principles carry real weight.
The core message is straightforward: the technology may be capable, but most failures trace back to the organisation rather than the model. Without the right processes, governance structures, and human involvement, even a well-configured AI agent can cause significant disruption to systems, data, and trust.
The first safeguard is narrowing the so-called blast radius. Every AI agent should be given the minimum access it needs to perform its function and nothing more. An agent built to route and summarise emails has no legitimate need for database write or delete permissions. Treating agents with the same discipline applied to user access management for human employees is a practical and proportionate approach.
Second, sandboxing remains essential during the deployment phase. Before an agent goes live, it should operate in a secure, isolated environment that mirrors the live system as closely as possible. This contains any unexpected outputs, whether hallucinations, flawed code logic, or irregular interactions with external data, without risking the host system.
Third, every deployed agent should have a named human accountable for its performance, its access levels, and its outputs. Defining who owns which decisions is not a bureaucratic exercise. It is one of the most frequently overlooked factors in scaling AI safely. Full auditability, strong data governance, and clearly defined failure modes follow naturally from getting decision rights right.
Fourth, backups. The advice here is blunt: if you have read accounts of agents deleting company data or wiping inboxes, you already understand why disconnected backups, ones the agent cannot reach, are non-negotiable.
Fifth, a kill switch should be built into the infrastructure of every AI agent deployment. This is a hard control that sits outside the agent’s reasoning processes, meaning it cannot be overridden or ignored. It removes the agent’s access to systems immediately, serving as a last line of defence when other controls have not held.
Sixth, organisations need a clear escalation procedure for when things go wrong. This includes understanding the legal exposure. Breaches involving personal data or AI systems operating contrary to EU regulation can attract fines of up to 35 million euros, or seven per cent of global turnover, whichever is higher. Research published in 2026 by the Aithos Research Foundation found that none of the leading AI models currently demonstrate acceptable levels of compliance with EU AI Act and privacy legislation. Erring on the side of caution is not timidity. It is sound professional judgement.
For Chartered Accountants Worldwide and the global profession, the significance of these six safeguards extends well beyond any one market. AI agents are being adopted by finance teams, audit functions, and advisory practices in virtually every economy. The risks of ungoverned deployment, from data loss and reputational damage to regulatory sanction, are not hypothetical.
What these safeguards collectively point to is the enduring importance of professional judgement alongside technological capability. AI can augment the work of Chartered Accountants significantly, but only when it operates within a framework that preserves accountability and transparency. The human element is not a weakness in the system. It is its most important feature.
Chartered Accountants are well placed to lead this conversation in their organisations, applying the same rigour to AI governance that they bring to financial reporting, risk management, and regulatory compliance. Establishing these safeguards is not a technical task reserved for IT teams. It is a leadership responsibility, and one that the profession is uniquely positioned to champion
—
Source: https://www.icaew.com/insights/viewpoints-on-the-news/2026/jul-2026/setting-up-an-ai-agent-correctly-part-2-six-safeguards
Source institute: ICAEW
Content type: article
Hub category: AI & Technology
