We're all managers now. And not all of us know how to do it well.
We recently attended the SNN (Signal Not Noise) Conference in Brisbane, and the keynote speaker, Cherie Curtis, the Chief People Officer of Trilogy Care, shared her journey with AI, and quite a few things she said resonated strongly with us!
Treat your AI Agents like employees - Role Description, Regular Reviews, PIPs and Separation if needed!
Her guidance is that if you ask AI to research a problem, analyse information or draft a recommendation, you are assigning work and judging performance.
The trap is that you might never have managed an employee. You might never have been taught how to delegate. Yet here you are, directing work that someone else will rely on.
If you apply that same lens to managing AI agents: define the work, establish boundaries and decide what good looks like before accepting the result, then you will be miles ahead!
The intern analogy is useful because it reminds you to provide direction and oversight before expecting independent delivery.
Wharton professor Ethan Mollick put it succinctly in a 2023 Microsoft WorkLab interview: “You have to think of it like an intern.”
In On-boarding your AI Intern, Mollick also wrote: “Just like any new worker, you are going to have to learn its strengths and weaknesses”. That is a useful starting point for leadership: learn what you can responsibly delegate before expanding the remit.
The analogy has limits. AI is software, with no employee's lived experience, professional responsibility or welfare needs. Managing it like a new starter means borrowing useful management disciplines. It does not mean treating people like machines or assuming AI develops judgement through experience as a person would.
Use role clarity, onboarding, proportionate delegation, specific feedback and team coordination to structure your agents' work.
Give each AI agent a clear purpose, expected outcome and boundary. This applies management by objectives: start with the result you need, then define acceptable work.
Replace “research this market” with a brief specifying the decision, audience, approved sources, required evidence and output format. State what the agent must not decide.
For example: “Prepare an evidence-backed market brief for our leadership team. Separate facts from assumptions. Flag missing information. Do not recommend expenditure.” Core State Consulting's focus on decision clarity provides a useful foundation for that brief.
Onboard your agent with the context a new team member would need to do the assignment responsibly. Explain your customers, terminology, priorities and constraints.
Provide approved reference material and examples of acceptable work. Include an example that fails the standard, and explain why. Use only information your organisation permits in the system.
Ask the agent to restate its brief and identify missing inputs before proceeding. Your role includes making meaning from the information available. Do not confuse giving an agent access to documents with giving it sufficient direction.
Match an agent's authority to its demonstrated reliability on the specific task and the consequences of failure. Borrow this task-specific supervision principle from situational leadership.
Start with research and recommendations. Move to drafting for review. Consider bounded action only after testing representative cases and defining escalation rules.
A reliable internal summariser has not automatically earned permission to send client advice. Set access limits, approval gates and a way to stop the process. Maintain leadership visibility through records of actions and exceptions. Reassess authority when the task, data or configuration changes.
Review AI work against an explicit standard, then explain the gap. Use a coaching discipline: observation, correction, example and another test.
“This is too generic” gives little direction. “Replace broad claims with verified evidence relevant to Australian professional services; flag anything you cannot substantiate” defines an actionable revision.
Check the revision yourself. Record useful corrections in reusable instructions, examples and evaluation cases. Do not assume a chat correction permanently trains the underlying model or persists across sessions. This is where human judgement remains essential: a confident revision still needs scrutiny.
Give your cohort of AI agents a shared outcome, distinct responsibilities and explicit handoffs. Apply the same coordination discipline you would use across a human team.
For example, a research agent supplies sources and uncertainties; an analysis agent tests interpretations; a drafting agent prepares a brief. Carry evidence and unresolved questions through every handoff.
Assign a human owner and review gates before consequential decisions or external release. One agent's approval is no substitute for independent verification. Core State Consulting explores how shared alignment connects separate contributions to an outcome.
You remain accountable for deciding whether AI-produced work is accurate, appropriate and useful. Measure accepted work, avoidable rework and decisions improved, rather than the volume of drafts produced. That keeps attention on decision value.
Before adding another agent, ask: have I defined the role, supplied the context, set authority, established feedback and connected the handoffs? Those questions expose management gaps you can address.
Core State Consulting helps leaders build clarity and capability. Start with one workflow and strengthen those five practices. You are responsible for turning the agents' activity into work your organisation can trust.
No. The employee comparison is a practical management analogy, not a claim that AI is a person. Use it to improve briefing, supervision and review. Human employees need respect, development and support; AI requires appropriate configuration, testing and controls.
Include its purpose, expected output, quality standards, approved sources, permitted actions and escalation rules. Name the human accountable for the result. Keep the brief specific enough to test whether the agent stayed within its remit and delivered acceptable work.
Task-specific supervision offers a useful approach: adjust oversight to demonstrated reliability and risk. An agent can perform one task reliably and another poorly. Test each use case rather than granting broad autonomy because a different assignment went well.
Do not assume a correction permanently changes the underlying model. Preserve improvements through saved instructions, approved examples and evaluation tests, as your system supports. Check what persists between sessions, and verify performance rather than relying on an agent's assurance.
A second agent can help identify issues, but its agreement does not establish accuracy. Check important claims against original evidence. Retain human review appropriate to the stakes, particularly before customer-facing communication, financial commitments or other consequential actions.
Measure whether the workflow delivers useful, verified outcomes after review effort is included. Track errors, exceptions and rework alongside completion time. Core State Consulting's discussion of human alignment offers a related perspective on valuing meaningful results over additional content.
Core State Consulting builds leadership capability around clarity and delivery.
Reach out if you need support or guidance.