How to Manage AI Agents as Part of Your Team

November 17, 2026

Listen to this article

An AI-generated audio summary of this article, made with Google NotebookLM.

Short on time? Jump to the quick test.

Most organisations are treating AI agents as a technology rollout. The risk sits in the management gap.

Agents are moving from demos into everyday work. They draft proposals, research suppliers, reconcile reports and answer customer emails. Unlike a chatbot, an agent doesn't wait to be asked each time. It takes a goal, decides on the steps, and acts using whatever access it has been given.

When an agent sends the wrong email, deletes the wrong file or quietly gets worse over three months, someone has to have been watching. In many organisations, it isn't clear whose job that is. The question for leaders isn't only what agents can do. It's who authorises their actions, who monitors them, and who answers when something goes wrong.

Agents aren't employees, and accountability for what they do never moves to them. But some of the disciplines we use to bring a new person into a team transfer well: a defined role, a named owner, limited access, close supervision at first, and clear points to check in. For agents, they need to be applied more strictly.

What is an AI agent, in plain English?

A chatbot answers. An agent does.

When you ask ChatGPT or Copilot Chat to summarise a document, you get an answer and decide what to do with it. An AI agent is given a goal and the access to act on it. It works out the steps, uses your systems to carry them out, and keeps going until the job is done or it hits something it can't handle. (For the wider picture, see automation vs GenAI vs agentic AI.)

For many agents, the closest comparison is a role, not a tool. Like someone in a defined position, an agent has a purpose, works within set procedures, uses specific systems, and should hand certain decisions to a person. Here's what that looks like in practice.

The comparison has limits. An agent follows its instructions without the judgement, context or sense of consequence a capable person brings. It can be confidently wrong, and it won't tell you it's out of its depth unless it has been set up to. Read the examples below with that in mind.

Role the agent covers What it does Where it must stop and hand to a person
Accounts payable assistant Reads incoming invoices, matches them to purchase orders, flags mismatches and enters approved ones into the finance system Approving payment, or anything over a set amount
Tender coordinator Pulls relevant past answers, case studies and CVs into a first-draft tender response Pricing, delivery commitments and submission
HR helpdesk Answers policy questions such as leave balances and travel rules, and drafts standard letters Anything about an individual's pay, performance, health or a complaint
Customer service triage Sorts the shared inbox, drafts replies to routine questions and updates the customer record Complaints, refunds and anything a customer is upset about
Project administrator Compiles daily site reports, chases missing subcontractor documents and updates the project tracker Instructions to subcontractors, variations and safety issues
Research analyst Monitors supplier news and regulatory changes, and sends a weekly summary Recommendations that commit the business to a course of action

Not every agent maps to a role. Some of the most common, and easiest to overlook, don't.

Type of agent Example What to watch
Personal assistant Triages one executive's inbox, prepares meeting briefs and books travel It can reach everything that person can, often including confidential material
Process agent Runs a new starter's onboarding across HR, IT and payroll It crosses teams, so ownership falls between them
Monitoring agent Watches cash flow, contract renewal dates or security alerts and raises a flag when a threshold is hit It works in the background, so nobody notices when it stops working
Built-in agent Switched on inside software you already use, such as your CRM, finance system or Microsoft 365 It arrives as a feature update, often without anyone deciding to use it
Agent that runs other agents Splits a large job, such as a market report, between several agents and assembles the result One error can pass through the whole chain before a person sees it

The role comparison fits these less well, but the six rules apply just the same. Built-in agents are the ones most often missed, because nobody remembers deciding to use them.

Why agents aren't just more software

We've managed software for decades. Agents are different in three ways that matter to a manager.

They act, not just inform. A report shows you something. An agent with credentials can retrieve records, update systems, contact customers and instruct other agents. Deloitte's research found 69% of Australian organisations were using agentic AI, but only 22% reported a highly advanced model for agent governance. That gap is where the risk sits.

They inherit your permissions. An agent works with whatever the account it runs under can reach. Years of overshared folders and loose permissions become usable overnight. IBM's 2025 Cost of a Data Breach report found that of organisations reporting an AI-related breach, 97% lacked proper AI access controls.

They don't stay the same. Traditional software generally behaves the same way tomorrow as it does today. An agent's underlying model gets updated, its inputs change and its outputs vary. An agent that worked well in March can be quietly worse by June.

None of this is a reason to avoid agents. It's a reason to manage them deliberately.

Six rules for managing AI agents

1. Give every agent a job description

Write down what the agent does, what it doesn't do, and what good output looks like. One paragraph is enough. If you can't describe the job clearly, the agent isn't ready to do it.

Example: "Drafts first-pass proposal responses from our template library. Does not send anything externally, quote prices or commit to delivery dates."

2. Give it a named business owner

Every agent needs a named business owner, responsible for its purpose, operation and oversight. Not "IT", not "the project team", and never the vendor. That person reviews its work, approves changes to it, and answers for it when something goes wrong. IT may manage its systems and permissions, and high-stakes actions may need their own approvers, but the business owner is the first call.

Responsibility can be shared across people. It can't be transferred to the software.

3. Grant only the access it needs

Treat an agent's permissions like a new starter's on day one, not an administrator's. If it drafts proposals, it needs the template library, not the whole shared drive or the finance system.

Where your systems support it, give the agent its own identifiable credentials, rather than letting it borrow an employee's or administrator's login.

This is also the moment to clean up the overshared folders you've been meaning to get to. An agent can surface and use them at scale.

4. Set stop-and-escalate points

Decide in advance where the agent must stop and hand back to a person. Typical points: anything sent outside the organisation, anything involving money, personal information or a customer complaint, and anything it isn't confident about.

An agent that can escalate well is more useful than one that can do everything.

5. Review its work like a new hire's

Check closely at first, every output if you can. As trust builds, move to sampling. But keep sampling, because agents drift as models update and the work changes.

Agree what you'll measure: accuracy, time saved, rework, complaints. "Usage is up" tells you very little.

6. Register it to the organisation

Every agent should sit in a register owned by the organisation, with its owner, purpose, access, current status and a review or retirement date. When an agent is no longer needed, switch it off and remove its access, as you would when someone leaves. Never build an agent on someone's personal account or credentials.

I saw the earlier version of this during the dot-com boom: domain names registered to individual staff walked out the door when those people left. An agent running on a departed employee's credentials is the same problem, with higher stakes.

Like a new hire, but not an employee

A small number of firms, and the main platforms, are starting to treat agents as identities rather than people. BNY is the clearest example: more than 100 "digital employees", each with its own login, a human manager and a narrow scope. Microsoft Entra Agent ID gives agents identities alongside users, and Workday's Agent System of Record tracks them from registration to retirement. These are capabilities, not proof of common practice. In most organisations, the gap is still the norm.

What hasn't moved is accountability. Agents have no employment rights or grievance process, and they aren't legal persons. Legal responsibility generally rests with the organisation, and operationally the named owner answers for the agent's work. In HR, agents increasingly handle routine queries and drafts, but decisions on hiring, pay, promotion, discipline and termination stay with people.

So use "treat it like a new hire" for the controls: a named owner, its own identity, least-privilege access, logging, and a way to pause or switch it off. Don't let the "digital employee" language suggest more than that.

The work is moving to agents. The accountability isn't.

What this means for managers

When a manager starts directing agents, their span of control grows quickly. One person may be able to oversee work that used to take several. That's the opportunity, and it changes what good management looks like.

New skills matter. Writing a clear brief, spotting a plausible but wrong answer, and knowing when not to use AI become core management skills. Few managers have been trained in them, and training alone rarely changes how people work. Without real support, it's no surprise that some managers resist AI adoption, even when they understand the tools.

Role design can't be left to chance. Gartner found only 4% of CIOs surveyed had explicit plans to redesign roles for a human-agent workforce. Those that did were 2.4 times more likely to be in its higher-performing AI group.

The people conversation comes first. When agents take on work people used to do, staff will ask what it means for their jobs. Some of the answers will be uncomfortable: some roles will grow, some will change, and some may go. People cope far better with an honest "here's what we know and how decisions will be made" than with silence. That's AI change management, not a technology task, and it's where most AI rollouts stall.

Quick test: do you have an agent management gap?

Tick each statement that's true for your organisation today.

No obvious gaps from this quick check. Keep reviewing as you add agents.

Review the gaps and fix them in order of risk, while the number of agents is still small.

You have several gaps. Run a structured review before adding more agents, starting with a named owner, tighter access and a register.

You ticked 5 or 6. That needs attention straight away, whatever your total.

0: no obvious gaps from this quick check. Keep reviewing as you add agents.

1–2: review the gaps and fix them in order of risk, while the number of agents is still small.

3 or more: you have several gaps. Run a structured review before adding more agents, starting with three rules: a named owner for every agent (rule 2), tighter access (rule 3) and an organisation-wide register (rule 6). Then look at the people questions.

This is a quick health check, not a validated assessment. A tick on 5 or 6 needs attention straight away, whatever your total.

Before your first agent goes live

The greater an agent's access, autonomy and potential impact, the stronger its controls must be. A simple way to judge risk: what the agent can access, plus what it can do, plus the consequences if it gets it wrong.

Every agent needs a baseline, however simple: its own identity, a named owner, limited access, logging and a place in the register. Then add oversight according to what it can do:

  • Observes: reads and summarises, changes nothing. The baseline is enough, unless it can see sensitive data such as personal or employee records. Then tighten its access and review it more often.
  • Recommends: suggests actions a person decides on. Add regular review of its outputs.
  • Acts with approval: prepares actions a person signs off. Add clear approval and stop-and-escalate points.
  • Acts on its own within limits: completes tasks without sign-off. Needs everything on the list below before it goes live.

Applying the same controls to every agent either chokes the simple ones or under-controls the ones that act.

The list below is the bar for any agent that acts. Scale it down for the ones that only observe or recommend, but never below the baseline.
  • It has a one-paragraph job description, including what it must not do
  • A named person owns it, and they know they do
  • Its access is limited to what the job needs, and permissions have been cleaned up
  • Its stop-and-escalate points are written down, and a named person can pause or switch it off
  • Someone is reviewing its outputs, with an agreed measure of quality
  • It's registered to the organisation, not an individual's account
  • It meets your privacy obligations. New Zealand's Privacy Commissioner has confirmed the Privacy Act 2020 applies to AI tools, and Australia's OAIC advises against putting personal information into publicly available AI tools
  • The team affected has been told what it does and what it means for them

If you're setting standards across the organisation, this is the core of an AI governance framework. For more on why IT and the business need to build these foundations together, see my IT Brief article, AI isn't the next cloud. Don't let it outgrow IT .

Questions leaders ask about managing AI agents

Who is accountable when an AI agent makes a mistake?

The organisation, through the agent's named business owner. Agents aren't legal persons, so legal responsibility generally rests with the organisation. Operationally, every agent needs a named business owner who oversees its work and answers for it, supported by IT for technical controls.

Should AI agents be covered by the same policies as employees?

No. Two different sets of rules are involved. Agents must operate inside your organisational policies on privacy, information security, financial delegations and records, like any system that touches data, money or records. Employment policies, rights and grievance processes don't apply to software, and decisions about people stay with people. Agents also need controls employees don't: restricted permissions, logging, human oversight and a way to switch them off.

Is there a standard framework for governing AI agents?

Not a mature one yet, though agent-specific guidance is emerging. NIST's AI Risk Management Framework and ISO/IEC 42001 give a recognised structure for managing AI, and Singapore has published guidance written specifically for agentic AI. Use a standard when a regulator, insurer or large client asks for one. But a certificate doesn't replace the basics: a named owner, controls matched to what each agent can do, and a register.

How much access should an AI agent have?

Only what its job needs. Start with the narrowest access that works, review it regularly, and clean up overshared files before the agent goes live.

How do you know if an AI agent is working well?

Review its outputs closely at first, then sample them regularly. Measure accuracy, time saved and rework, not just how often it's used, because performance can drift as models change.

Planning to put agents to work?

Putting agents to work safely raises four questions. Matrix AI Consulting helps leadership teams across New Zealand and Australia answer them:

Book a conversation

AI BLOG

The three levels of AI adoption, from tool to task to workflow
By Glen Maguire • November 12, 2026
Rolled out Copilot, Claude or ChatGPT but usage is shallow? One of the most common barriers is relevance. Why it matters, and how to close the gap.
Why Managers Resist AI Adoption
By Glen Maguire • November 4, 2026
AI changes who knows, who decides and who matters. Why managers can stall AI adoption even when they understand it, and what leaders should do about it.
AI Comparison ChatGPT, Gemini, Copilot, Grok and Claude
By Glen Maguire • November 2, 2026
Gemini just topped the AI leaderboard, and most businesses can’t use it. How NZ and Australian leaders should choose AI on access, control and value.
More Posts