When the Algorithm Becomes a Market Actor

Financial markets have been automated for years.

What is changing is the nature of the automation.

Traditional automated trading systems generally execute rules established by humans.

Increasingly sophisticated AI systems can identify patterns, adapt behaviour and participate in decision processes whose individual outputs may not have been explicitly determined in advance by their human designers.

ASIC's 2026 reforms to Australia's Market Integrity Rules recognise that transition.

The regulator is strengthening requirements applying to automated and AI-enabled trading, including controls around the testing, monitoring and governance of algorithms. ASIC has specifically identified the possibility that trading algorithms may behave in ways that are opaque, unpredictable or potentially manipulative.

That development points towards a broader governance question.

What happens when the system being governed becomes capable of acting faster than the governance system overseeing it?

Speed changes governance

Traditional governance assumes that there is time between observation and intervention.

A risk emerges.

Information is escalated.

Management considers it.

A committee may review it.

Authority is exercised.

But algorithmic systems can act thousands or millions of times before conventional governance processes have recognised that something significant is occurring.

The issue is therefore not merely automation.

It is governance latency.

If operational systems act in milliseconds while organisational oversight operates in days, weeks or quarterly reporting cycles, the formal governance structure may remain intact while losing practical control.

Governance must therefore become capable of operating at more than one speed.

Human approval cannot govern every machine decision

The obvious response might be to require greater human supervision.

But direct human approval of every algorithmic decision would defeat much of the purpose of automation and, in many environments, be physically impossible.

The governance challenge is consequently not to insert a person into every transaction.

It is to govern the conditions under which autonomous decisions may occur.

That means defining:

• permitted objectives;
• prohibited behaviours;
• operating boundaries;
• risk tolerances;
• testing requirements;
• monitoring thresholds;
• intervention mechanisms;
• shutdown conditions; and
• accountability for outcomes.

This resembles the governance of delegated human authority.

But there is an important difference.

Human delegates generally understand that they possess authority and recognise that the authority has limits.

An algorithm does not possess institutional judgement merely because it possesses computational capability.

The limits must therefore exist in the architecture around it.

The board cannot govern the algorithm line by line

Boards should resist two extremes.

The first is technological disengagement:

This is too technical for directors.

The second is technological micromanagement:

Directors must understand how every model works.

Neither is necessary.

The board's responsibility is not to become the organisation's engineering department.

Its responsibility is to ensure that the organisation understands the systems upon which it depends, has established appropriate authority and risk boundaries, can detect abnormal behaviour and can intervene when necessary.

The critical board questions are therefore architectural:

Where can algorithms act on behalf of this organisation?

What prevents them acting outside the organisation's intended boundaries?

Who can stop them?

How quickly can that happen?

And how would the board know whether those controls actually work?

Assurance must become behavioural

Traditional technology assurance often asks whether systems are available, secure and functioning as designed.

AI creates an additional requirement.

Organisations increasingly need assurance about behaviour.

A system may be technically operational while behaving in an unexpected way.

It may satisfy its immediate objective while producing an outcome inconsistent with the institution's broader obligations.

It may discover a pathway that designers did not anticipate.

That means assurance must extend beyond:

Is the system working?

to:

Is the system behaving within the authority and risk boundaries we intended?

That is a substantially more demanding governance proposition.

Gatekeepers still matter

ASIC's approach to automated trading contains an important broader principle.

Technology does not eliminate institutional responsibility.

Market participants remain gatekeepers.

The organisation cannot transfer accountability to the algorithm merely because the algorithm generated the action.

The same principle will increasingly apply outside financial markets.

As autonomous systems enter procurement, credit assessment, insurance, customer interaction, cybersecurity, logistics and organisational decision-making, institutions will need to determine where machines are permitted to act and where human authority must remain decisive.

The next governance divide

The organisations best positioned for the next stage of AI adoption may not be those deploying the most advanced systems.

They may be those that understand most clearly the difference between automation and authority.

Automation determines what a system can do without continuous human intervention.

Governance determines what it is permitted to do, how its behaviour becomes visible and who remains accountable for the consequences.

As algorithms increasingly become actors rather than tools, that distinction becomes critical.

The question for boards is therefore no longer simply whether their organisations are using artificial intelligence.

It is:

Where inside this organisation has artificial intelligence acquired the practical ability to act—and has our governance system evolved quickly enough to remain in control?