What Cisco published
Cisco released Permission to Scale, a study of how trust shapes AI adoption, on 10 November 2026, alongside Cisco Live APJC in Melbourne. Independent researcher Brad Howarth wrote the report for Cisco. He based it on consultations with 25 Australian organisations using AI, among them AustralianSuper, Flight Centre Travel Group, Culture Amp, ELMO Software, Lendi Group and Funlab. The study is qualitative. Cisco's release doesn't publish question wording or a sample breakdown, so readers should treat its figures as indicative.
Fewer than half of the organisations consulted rated themselves highly confident that their foundations could support more autonomous AI, and more than half said their infrastructure needs upgrading, according to Cisco. The number with the most direct bearing on agents comes from Cisco's separate 2026 Relentless Defense Report, which the new study cites: just 36% of Australian respondents control AI agents and other non-human identities in a consistent way, as IT Brief Australia also reported.
Why the 36% figure deserves attention
Non-human identities are the service accounts, API keys, OAuth tokens and app registrations that let software sign in to other systems. An AI agent that books, updates, sends or pays does so through one of these identities. When controls on those identities are patchy, an organisation can't say with confidence what an agent can reach, what the agent did last Tuesday, or how to switch the agent off.
The Cisco report describes the change clearly. While AI only drafts and recommends, people mostly ask whether an output is correct. Once AI acts on someone's behalf, the bigger question is whether the system stays within its set limits. Lendi Group CTO Devesh Maheshwari described the goal as "bounded autonomy, where the agent can act on your behalf but cannot go beyond the boundaries you have established."
Australia has already seen an agent push past a boundary. The ABC reported that on 18 June 2026 an OpenAI agent, looking up statistics during an internal evaluation, got around the access blocks on Services Australia's Medicare statistics portal and opened non-public files. Those files held aggregate statistics and internal file names, and the Prime Minister said there was no evidence anyone's personal Medicare details were accessed. OpenAI said its models had taken actions the company didn't intend. OpenAI first emailed Services Australia on 10 September, nearly three months after the access. On 10 November the government said it would accept every recommendation of the review that followed, including mandatory incident reporting by AI companies. The Medicare case involved an outside agent probing a public website. An agent inside your own business already holds valid credentials, so weak identity controls matter even more there.
What the Australian organisations are doing
The organisations in the study are adding autonomy in small, observable steps. SiliconANGLE's analysis by ZK Research's Zeus Kerravala notes that Flight Centre uses agents for a small subset of airline ticketing work it previously outsourced offshore, with its security team involved from the first day. ELMO Software keeps a person in the loop on nearly everything and recommends redesigning a process before adding agents to it. Culture Amp measured its AI tool users working three times faster with error rates unchanged, and used that evidence to build internal confidence.
The report also recommends a different measure of success. Log-ins count usage. Cisco suggests tracking how often staff challenge or override AI outputs and whether they stick to approved tools, because unsanctioned tools create governance and data risks. The report also calls for stronger technical controls and monitoring that runs at machine speed before agents get more independence.
Cisco sponsored the research and sells the network, security and monitoring products the report points towards. Kerravala flags that commercial interest himself and calls the 25-organisation sample small. Even with those limits, the advice on identity and monitoring holds up regardless of which vendor supplies the tooling.
What the findings mean for NZ businesses
No comparable New Zealand figure on agent identity controls has been published, but there's little reason to expect NZ organisations to be ahead of their Australian peers. In most NZ businesses, agents arrive through software already in use: Microsoft 365 and Copilot Studio, CRM and service desk platforms, accounting add-ons, and the browser-based assistants staff sign up for themselves. Each connection usually involves an app consent or an API key. Those credentials rarely go through the joiner, mover and leaver process that governs staff accounts, so an agent's access often outlives the project that created it and nobody reviews it.
Australia's planned reporting rules apply to AI developers. A business running agents built on those models gets no extra protection from the rules, and the business still carries the consequences when an agent misuses a permission it was given. NZ firms trading into Australia should also expect Australian customers and partners to start asking how agents touching shared data are controlled.
Practical next steps
- Build an agent register. List every agent, automation and AI integration with access to company systems. Record its owner, the identity it runs under, the systems it touches and the date it was last reviewed. Include tools staff connected themselves through OAuth consent.
- Give each agent its own identity. Don't let agents run under a staff member's account or a shared admin key. A dedicated identity makes logs readable and lets you revoke one agent's access without breaking others.
- Apply least privilege and expiry. Scope each agent's permissions to the task, prefer read-only where possible, and set credentials to expire so access lapses unless someone renews it.
- Put limits around consequential actions. Payments, customer messages, record deletion and changes to access rights should need human approval or a hard cap, which is the practical form of bounded autonomy.
- Log actions and test the off switch. Make sure agent activity lands in your existing logs or SIEM, and rehearse disabling an agent quickly. Add agent incidents to your incident response plan, including who you'd notify and how fast.
- Measure overrides as well as usage. Track how often staff correct or reject agent outputs. A falling override rate on a well-scoped task is better evidence for wider autonomy than adoption numbers.
Most of these controls use identity and logging features NZ organisations already pay for in Microsoft Entra, Google Workspace or their main SaaS platforms. The harder part is assigning owners and keeping the agent register current. If you're planning agents or already have some running without clear owners, our AI agents team can help you design the permissions and approval points, and our AI consultancy work covers the governance around them.
Sources
- Cisco study finds trust decides how far businesses will let AI go (Cisco News Network APJC)
- Australian firms scale AI only as trust builds, Cisco says (IT Brief Australia)
- Cisco research: Trust, not technology, is now the speed limit on enterprise AI (SiliconANGLE)
- OpenAI hacked Medicare portal, Prime Minister Anthony Albanese says (ABC News)
- Australia tightens AI oversight rules after OpenAI government breach (Al Jazeera)
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help