Read this for your industry
Optional. Choose your industry and this piece opens with a situation from your world and closes with a first move for you. Your choice carries across the guides and analysis on this site.
You lead a licensee that is putting agents to work
ASIC's Report 798 reviewed 624 AI use cases at 23 licensees and found governance lagging behind adoption. According to Hall & Wilcox's summary, only 12 licensees had AI policies that referred to fairness or bias, only 10 had policies on disclosing AI use to consumers, and about 30% of use cases relied on models built by third parties.
The seven parts in this guide line up closely with what ASIC is asking of licensees: a defined role, current source material, a bounded task, a quality bar, a point to stop and ask, a named accountable person and a way to learn from corrections.
What this means for you
- Keep an inventory of every agent, its role and its accountable owner.
- Set the stop point, meaning the cases where an agent must hand over to a person, such as clients experiencing vulnerability.
- Report agent performance to the board or risk committee alongside your other advice quality measures.
You run stores where agents now touch customers and staff
Bunnings and Kmart both used facial recognition in stores to tackle crime and refund fraud, and the Privacy Commissioner found that both had breached privacy law (Bunnings, Kmart). When the tribunal reviewed the Bunnings decision, it upheld the findings that customers had not been properly notified and that the company had fallen short on its privacy governance.
Both retailers wanted to protect staff and stores. What was missing was the system around the technology: the assessment, the notice and the person accountable for how it was used. Agents working in your stores need the same kitchen pass.
What this means for you
- Write a short risk assessment for each agent that touches customer or staff information.
- Name the person accountable for each agent in your territory.
- Set a stop point for anything involving privacy, pay or safety.
You lead clinical governance as agents arrive
The RACGP has added AI to its Standards for general practices, with expectations for governance, patient consent and clinical oversight of the tools a practice uses. Aged care providers work under strengthened Quality Standards that include governance and a dedicated clinical care standard.
Those requirements describe the kitchen pass in clinical language. A clinician checks the work before it reaches the patient, someone owns each tool, and the team learns from what goes wrong.
What this means for you
- Assess each agent before it goes live, including privacy, consent and whether it could be a regulated medical device.
- Name a clinical owner for every agent that touches care.
- Review incidents and near misses involving agents in your usual clinical governance meetings.
You lead a delivery system that agents now feed
The 2025 DORA report found that AI acts as an amplifier: it magnifies the strengths of well-run organisations and the weaknesses of struggling ones. DORA's advice to leaders includes strengthening control systems such as automated testing, mature version control and fast feedback loops, because more change without them leads to instability.
In engineering, the kitchen pass already exists in your pipeline. The question for you is whether it is strong enough for the volume of change agents now produce.
What this means for you
- Check that automated tests cover the parts of the system agents change most often.
- Give each agent a bounded scope in the codebase or supplier base, with a named owner.
- Track change failure rate and recovery time as agent use grows.
You run an installer under close regulatory watch
After inspecting 1,278 battery installations, the Clean Energy Regulator found 60.8% substandard, and installations under the program must be carried out by an accredited installer. Quotes, designs and compliance paperwork all feed into that result.
Agents that prepare quotes, designs and paperwork need the same pass a good supervisor provides. They work from current requirements, they stop when something is unclear, and a licensed person signs off before anything is energised.
What this means for you
- Load the current photo and labelling requirements into every agent that prepares or reviews compliance paperwork.
- Set the stop point, so any design or site condition the agent has not seen before goes to a designer.
- Keep the licensed sign-off with the person who answers for it.
What happens after the draft
Once agent output starts moving through a real organisation, the questions arrive quickly. Did it use the right sources? Was the task clearly bounded? Who checks it before it goes further? Who answers for it if it turns out to be wrong?
Those are fair questions, and the people asking them are usually the ones who will carry the risk. They deserve good answers.
The kitchen pass
A commercial kitchen is a useful comparison. A skilled cook can put a plate together in minutes. Whether that plate is safe to send out depends on everything around the cook: stock rotated so nothing is past its date, fridge temperatures logged, a head chef at the pass checking each dish, and someone who signs off if a customer gets sick. A great cook in a kitchen without those things is a risk to the business.
The agent is the cook. Most organisations are building the kitchen as they go.
The seven parts
I have ended up with the same seven parts in my own use of specialist agents. Each time, the useful step was giving the agent a defined place in the workflow.
- A role. What this agent is for and where its job ends.
- Source material. The documents and data it should work from, so its answers are grounded.
- A bounded task. A brief clear enough that you would recognise a good answer when you saw one.
- A quality bar. What must be true before the output moves on. In the kitchen, this is the head chef at the pass.
- A stopping point. When the agent should stop and ask, so it does not fill gaps with guesses.
- A named owner. One person who is accountable for the result.
- A learning loop. Corrections written down so the same mistake does not come back. There is more on this in the correction log.
Why this matters now
Agents are moving from helping individuals to running pieces of work. Microsoft’s Work Trend Index describes three phases: AI as an assistant, then agents joining teams as digital colleagues that take on tasks at human direction, then people setting direction for agents that run whole processes. Each phase moves more judgement into the system around the agent, which is the kitchen in this comparison.
When that system is missing, results stall. Fortune’s coverage of MIT NANDA’s 2025 report on AI in business reported that the core issue was a “learning gap” for both tools and organisations, and that model quality was not the main problem.
What leaders can ask
For any workflow where agents produce something other people rely on, go through the seven parts and ask who owns each one. Where nobody does, output will move faster than anyone can check it.
Start with those gaps, the same way a new head chef checks the fridge logs before rewriting the menu.
List every AI tool your advisers use today, including the ones they chose for themselves. Give each a named owner and a stop rule before the next risk committee meeting.
List the agents and AI tools your stores use. For each one, write down who is accountable and what it must never do, and share the list with your store managers.
Pick one AI tool already in use at your sites, which is often an AI scribe. Check it against the RACGP standards or your clinical governance framework, and close the first gap you find.
Compare the change failure rate for agent-authored changes with the rest over the last quarter. If it is higher, strengthen the tests in that area before widening agent access.
Write the seven parts for your review agent on one page: its role, its sources, its task, the quality bar, when it stops, who owns it and how it learns. Pin it next to the job board.
Questions readers ask
Is this the same as AI governance?
It overlaps. Governance usually covers policy and approval. This covers how each piece of agent work is set up so the output can be used, which is closer to how you would run any team.
Does every agent task need all seven parts?
No. A personal draft needs a role and a quality bar. Anything that reaches a customer, a regulator or a decision needs all seven.
Sources
- 2025: The year the Frontier Firm is born (Work Trend Index), Microsoft WorkLab, April 2025
- MIT report: 95% of generative AI pilots at companies are failing, Fortune, 18 August 2025
The views here are my own, not those of my employer or its clients.