AI agents are moving beyond providing information. They can now act across applications, access data, initiate transactions and influence business processes. This creates new opportunities, but also new questions about identity, authority and accountability.
Explore two perspectives on how organisations can manage AI agent risk, strengthen digital trust and give AI agents greater autonomy with confidence.
“Identity and access management capabilities have not yet adjusted to nonhuman actors such as AI agents.”
Sources: PwC 2027 Global Digital Trust Insights: Australian insights; October 2026; APRA letter to industry on artificial intelligence, April 2026.
Why this matters in Australia
Regulators have set the bar
APRA expects Multi-Factor Authentication (MFA) or equivalent controls for high-risk activities and privileged access, and enforceable preventative controls rather than reliance on policy or after-the-fact detection. MFA shows a person was present at login, not that a person is directing each action that follows.
AI is making fraud cheaper
AUSTRAC's Fintel Alliance identified potentially hundreds of millions of dollars in suspected fraudulent home loans across 10 major banks, supported by falsified documents. Any organisation that accepts applications, claims, invoices or account changes online faces the same pressure, at a scale and speed manual checks were not built for.
Obligations and expectations are rising
Most Scams Prevention Framework obligations commence on 31 March 2027. Consumer Data Right action initiation will create a sanctioned route for accredited parties to act for consumers. The Government's guidance for AI Adoption calls for risk management that accounts for agentic AI, and from 10 December 2026 privacy policies must explain where computer programs make or inform significant decisions.
Start with every agent in your environment
Agent risk has two sides: the agents you deploy and the agents you don’t control. Govern your own firmly, and be equally deliberate about which external agents you allow into your environment.
Govern the agents you deploy
Decide your stance on the agents you do not control
Baseline your exposure to external agents, legitimate or not, and agree a risk appetite: is this a practice you condone?
Building trust in AI agents across financial services
For banks, lenders, insurers, superannuation trustees and payments providers
Govern the agents you own. Verify and authorise the agents you do not.
Permission to access an account does not authorise every action that follows. Whether an agent is the institution's own or acts for a customer or member, institutions need to know which agent is acting, what was delegated and whether each action reflects the intended authority.
The perspective covers how to:
govern the agents you deploy and agree your stance on external agents
know which agents reach your customer and member channels
authorise high-risk actions, not just the login
detect, stop and trace agent activity, including AI-enabled fraud.
It sets out an illustrative control pattern from customer authentication to agent identification, delegated authority, action-level authorisation and attribution.
Building trust in AI agents across the enterprise
Govern the agents you own. Verify and authorise the agents you do not.
Agents now affect sensitive data, business processes and customer interactions. Unintended or excessive actions create risk, not only malicious ones, and agent access can outpace the controls needed to verify who or what is acting, and with what authority.
The perspective covers how to:
- govern the agents you deploy and agree your stance on external agents
- know which agents reach your systems, data and workflows
- authorise each action, not just the login
- detect, stop and trace agent activity across applications.
Read the cross-industry perspective
One AI Trust conversation, three dimensions
- Trusted AI
Are we deploying and operating agents we can trust, with clear accountability? - Identity and agent security
Can we independently verify and enforce what agents may do? - AI-enabled security and monitoring
Can we detect and respond to agent risk, knowing detection is not authorisation?
Next step: understand your agent exposure and control gaps
Identify the agents across your environment, those you deploy and those you do not control, agree your stance on external agents, assess identity, authorisation and fraud controls, and set priorities for secure adoption.
Explore PwC’s AI Trust services
Contact our team
Contact us
