When AI Acts: A Responsible Collaboration Playbook

AI has moved from responding to acting. That shift changes who's responsible when things go wrong, and what "working with AI" means for a team.

In June, I attended AI Summit 2026 at the National Communication Museum in Melbourne, where a panel of researchers, technologists, and a Luritja creative technologist spent the morning pulling apart the idea of agentic AI. The practical version, not the hype one.

The session was a useful reality check. AI capability is accelerating, but the gaps that matter right now are governance, accountability and how teams structure their collaboration with these systems.

The shift from tool to teammate

Most organisations still think about AI as a tool you query. Ask a question, get an answer. That mental model is already out of date.

Agentic AI systems plan and act. They can be given a goal, access to systems, and enough autonomy to get there without a human approving every step. Some organisations have already given AI agents their own email addresses, Slack access, and the ability to track commitments on behalf of teams.

Professor Jie Lu from UTS framed it clearly: we're moving from "a human using an AI tool" to "a human managing a team where some members are AI agents." Teams don't use agents; they're responsible for them.

One panellist noted that many organisations are still confused about what "agentic" even means. A client of theirs claimed to have 400 AI agents running. They were automation with a rebrand: fixed rules following the same path every time, with no real decision making involved. Getting the definition right matters, because you can't calibrate trust or assign accountability for something you haven't correctly identified.

The accountability gap

When AI acts and something goes wrong, who's responsible?

The honest answer right now is: it's murky. Dheeren Vélu from Capgemini named this directly. Accountability doesn't disappear with agentic systems; it gets harder to trace. The designer is responsible for architecture and design rules. The deployer is responsible for authority settings. The user is responsible for the permissions they grant. When something breaks, everyone points to everyone else.

He called this "authority laundering." Like its financial namesake, it makes responsibility hard to trace and harder to assign.

This isn't hypothetical. Australia's Robodebt scheme was raised as a real example. Technology enabled policy decisions that led to serious harm. The governance structure focused accountability on end users, while the people who set up the system faced few consequences. The technology didn't create that outcome, but it made it easier for everyone to point the finger elsewhere.

If you can't trace a decision back to a person, your governance model needs work

What responsible collaboration looks like in practice

There are practical things teams can do right now.

Start by defining what you're actually deploying. Agentic AI has three dimensions worth understanding: autonomy (can it make decisions independently?), agency (can it act on those decisions?), and authority boundaries (what is it allowed to do?). If you're evaluating a vendor or building internally, ask those questions before going further.

Give agents bounded authority. One panellist used a travel agent analogy: you might give your travel agent authority to book flights and hotels based on your preferences, but you wouldn't give them access to your entire financial accounts or let them send emails on your behalf without limits. Set a budget ("spend up to $200 on this task") rather than open-ended system access.

Build audit trails from the start. Authority laundering happens when there's no clear record of who set what permissions, when and why. Designing those trails before you need them is far easier than reconstructing accountability after something breaks.

And keep humans leading teams that include agents. This sounds obvious, but it erodes gradually. Agents take on more tasks, fewer people stay close to the work, and oversight becomes nominal. Keep a human in the loop for decisions with real consequences.

How you interact with AI shapes more than the output

One of the more unexpected threads came from Professor Mari Velonaki, who raised the question of whether it matters how we talk to AI systems.

Her view: it doesn't change the AI's output, but it changes us. The habits we form, whether we're commanding, collaborating or dismissing, seep into how we interact with colleagues and how children observe adults engaging with technology. Every interaction becomes part of workplace culture.

Matthew Ngamurarri Heffernan added a broader point. From an Indigenous perspective, agency isn't individual; it's relational. It involves representing community, context, and the people around you. That framing is worth borrowing. Technical guardrails matter, but so does the culture a team builds around how and when they use these systems.

We aren't passive recipients of whatever AI does to our organisations. We make choices about how to integrate it.

What to do next

The governance frameworks most organisations rely on were built for a world where humans made decisions and software executed instructions. Agentic AI doesn't fit that model cleanly.

You don't need to solve everything at once.

Map what you've actually got. Audit the AI-assisted processes your team relies on. For each one, identify who has oversight, what the agent is authorised to do, and what happens when it gets something wrong.

Ask the right questions about new tools. Before deploying any agentic system, ask the vendor: what is this agent's authority scope? What decisions require human approval? What audit trail does it produce? If they can't answer clearly, that's your answer.

Set cultural norms alongside technical limits. Decide as a team where AI helps and where you want humans in charge, because of quality and accountability as much as risk. Write it down. Revisit it as the tools change.

The capability side of AI is moving fast. The responsibility side is moving slower. Closing that gap is what matters.

Need support in deploying an agentic system in your business?