The Government Automation Paradox

29 Sept

Why proportionate governance matters from macros to agentic AI

Government organisations are actively exploring artificial intelligence, including systems that may eventually act across processes and applications. Yet many of the same organisations cannot provide a controlled pathway for a simple document automation. That mismatch reveals a wider governance problem: we often regulate the name of the technology instead of the capability, action and risk involved.

A simple process with an unnecessary obstacle

Recently, I was working on a Word document designed to support a business process. The requirement was practical. The document needed to capture information consistently, guide the user through relevant considerations and produce a structured result.

A modest piece of automation could have made the process easier. The user would enter the required information, select a button and the document would assemble the relevant output. The result would be faster to complete, easier to repeat and less vulnerable to manual handling errors.

Visual Basic for Applications was the most direct option within Word. Depending on the wider architecture, a database-backed solution could also have been considered. Neither approach required artificial intelligence. The underlying logic could be defined, tested and explained in advance.

The barrier was not the business requirement. It was the organisation's restriction on macros. Instead of a controlled automation, the process required more mapping, more manual handling and more design effort to achieve a similar result.

At the same time, the organisation was encouraging the use of AI and beginning to discuss how agentic capabilities might be incorporated into the work environment. The contrast was difficult to ignore.

Why macros are treated with caution

There are sound reasons for controlling macros. A macro is executable code embedded in an ordinary-looking Office file. Malicious macros have been used to deliver malware, exploit user permissions and compromise organisational systems. A document can also be copied, emailed or changed without the controls that would normally apply to an enterprise application.

The Australian Signals Directorate includes restrictions on Microsoft Office macros in the Essential Eight. Its guidance says macros should be disabled for users who do not have a demonstrated business requirement. It also supports a risk-based approach to exceptions, including documented approval, compensating controls, monitoring and review.

That distinction matters. The guidance does not require every macro to be prohibited in every circumstance. It recognises that some users may have a legitimate business requirement and that exceptions can be governed.

The problem arises when an organisation implements the prohibition but does not establish the exception pathway. A security control then becomes an absolute barrier, even when the proposed automation is narrow, testable and valuable.

Agentic AI sharpens the inconsistency

The inconsistency becomes more pronounced when the organisation begins considering agentic AI. A conventional document macro follows instructions written by a developer. Its logic may be complex, but its intended behaviour can be inspected and tested before release.

An AI agent may be designed to interpret information, choose between actions, retrieve data, interact with applications or initiate workflow steps. Its authority can be restricted, and good design should limit what it can access and do. Even so, the governance challenge is potentially greater because the system may operate with more discretion and across a wider environment.

This does not mean that macros are safe and AI agents are unsafe. Their risks are different. It does mean that a blanket prohibition on a small deterministic automation sits uneasily beside enthusiastic discussion of systems that may act with greater autonomy.

The different responses often reflect organisational history. Cybersecurity teams have years of evidence about malicious macros and established controls for restricting them. AI arrives through a different door, supported by executive interest, vendor investment and a strong productivity narrative. The technologies may also have different sponsors, owners and approval pathways. As a result, nobody necessarily places the two decisions side by side and asks whether the treatment is proportionate.

Govern the capability and the action

Technology labels are a poor substitute for risk assessment. The word macro can trigger an automatic refusal, while the words artificial intelligence or agentic can trigger an innovation conversation. Neither response tells us enough about the proposed use.

A more consistent approach would examine what the technology will actually do. The same questions can be applied to a macro, workflow, script, low-code application or AI agent:

What information can it access, and is that access necessary?
What actions can it perform, and what is the consequence if an action is wrong?
Who owns the solution and remains accountable for its operation?
How will it be tested, approved, monitored and changed?
Can its actions be explained, recorded and audited?
Can the process be stopped, corrected or reversed?

These questions focus attention on the actual exposure. A macro that formats an approved template presents a different risk from one that calls external services or modifies large numbers of files. An AI assistant that drafts text presents a different risk from an agent authorised to update records or initiate decisions. Governance should be capable of recognising those differences.

A controlled path for practical automation

The answer is not to enable macros for everyone. It is to create a controlled pathway for legitimate automation. Depending on the risk and operating environment, that pathway could require:

a documented business requirement and an assessment of the expected benefit
a named business owner and a named technical owner
development by approved personnel within a controlled environment
signed code, trusted locations and restricted permissions
testing, peer review and change control before release
version control, documentation and a supported fallback process
periodic review to confirm that the automation remains necessary and safe

Other tools may sometimes be more suitable. Power Automate, Office Scripts, Power Apps, forms or a purpose-built application may provide better controls and support. But selecting a larger platform solely because a smaller tool has been prohibited can increase cost, complexity and delivery time without necessarily reducing the underlying business risk.

The governing decision should begin with the process, the information involved and the consequence of failure. The technology choice should follow from that assessment.

The question leaders should ask

If an organisation is prepared to explore AI agents that may interpret information and act across systems, it should also be capable of assessing a simpler automation on its merits. Otherwise, the organisation risks building sophisticated governance for emerging technology while leaving routine process improvement trapped behind blanket controls.

This is not an argument against cybersecurity controls or in favour of macros as the default solution. It is an argument for proportionate enablement. Known risks deserve strong controls, but strong controls should include a safe route for legitimate use where the business value justifies it.

Governance is decision infrastructure. Its purpose is not only to prevent unacceptable activity. It should help an organisation make a defensible decision, assign accountability and proceed under clear conditions when a useful capability can be introduced safely.

The question should not be whether a technology is old or new, familiar or fashionable. The question is whether the proposed use creates value, whether its risks are understood and whether the organisation can establish controls proportionate to those risks.