What are AI agents? A plain English guide for New Zealand business.
What an AI agent is, how it differs from a chatbot or automation, what agents can do for a New Zealand business today, and how to keep them under control.
The short answer
An AI agent is AI that works towards a goal by taking steps on its own.
It plans, uses tools such as email, files and business systems, checks the result and keeps going until the task is done.
A chatbot answers and waits. An agent acts, so it needs clear limits and a person approving what matters.
Key facts.
- Anthropic separates workflows, where models and tools follow steps set in advance, from agents, where the model directs its own process and tool use.Anthropic, December 2024 · source
- The Model Context Protocol (MCP), the open standard for connecting agents to tools and data, was launched by Anthropic in November 2024 and donated to the Linux Foundation's Agentic AI Foundation in December 2025.Anthropic, December 2025 · source
- 40 percent of large organisations (revenue over US$1 billion) report scaling AI agents, up from 27 percent a year earlier. Among smaller organisations the figure is flat at 22 percent.McKinsey, The state of AI in 2026, 1,719 respondents, May to June 2026 · source
- Gartner predicts over 40 percent of agentic AI projects will be cancelled by the end of 2027, because of cost, unclear value or weak risk controls.Gartner, June 2025 · source
- "Excessive agency", giving an AI too much functionality, permission or autonomy, is one of the top ten security risks for AI applications.OWASP Top 10 for LLM Applications, 2025 · source
- Before using or disclosing personal information, the Privacy Act 2020 requires reasonable steps to check it is accurate, up to date, complete and not misleading. That includes information an agent produces.Privacy Act 2020, section 22, principle 8 · source
What is an AI agent?
An AI agent is software that is given a goal rather than a single question, and works out the steps to reach it.
Ask a chatbot for a summary of a supplier contract and it writes one. Ask an agent to check every supplier contract due for renewal this quarter, flag the ones with price rises above inflation and draft a note to each account manager, and it will find the contracts, read them, compare the terms, write the notes and put them in front of you to approve.
That is the difference in one line: a chatbot produces an answer, an agent produces an outcome. Agents are a kind of agentic AI, and most are built on the same large language models behind ChatGPT, Claude and Copilot. What makes them agents is what sits around the model: access to tools, a way to plan, memory of what has happened so far, and permission to act. If you are new to the basics, start with our guide to what AI is.
Anthropic, which makes Claude, draws a useful line in its guide to building effective agents: workflows run models and tools along steps written in advance, while agents decide their own steps as they go. OpenAI's practical guide to building agents puts it more simply, as systems that independently accomplish tasks on your behalf, built from a model, tools and instructions.
How does an AI agent work?
Every agent runs the same loop: plan, act, check, repeat, with a person at the gate for anything that matters.
- GoalA person sets the outcome and the limits.
- PlanThe model breaks the goal into steps.
- ActIt uses tools: search, files, email, your CRM.
- CheckIt reads the result and decides what is next.
↻ Repeat until the goal is met, or the agent asks for help
The tools are what turn a model into an agent. A tool can be a web search, a spreadsheet, a calendar or an API into your accounting or property system. The agent decides which tool to use at each step, reads what comes back, and adjusts. That flexibility is the point, and it is also the risk: an agent can take a path nobody planned for.
So the design questions are practical ones. Which tools can it use? Can it only read, or can it also write? What has to come back to a person before it happens? And how will you know if it starts getting things wrong?
How are agents different from chatbots and automation?
Automation follows your steps, a chatbot answers your question, and an agent chooses its own steps towards your goal.
| Rules-based automation | AI chatbot or assistant | AI agent | |
|---|---|---|---|
| You give it | A fixed set of steps | A question or a task | A goal and limits |
| It decides | Nothing | What to say | Which steps to take, in what order |
| It can act | Yes, exactly as scripted | Rarely, beyond writing | Yes, through the tools it is given |
| Handles the unexpected | No, it stops or fails | Within one conversation | Yes, by re-planning |
| Best for | High-volume, predictable work | Drafting, research, analysis | Multi-step work across systems |
| Main risk | Breaks when inputs change | Wrong answers | Wrong actions |
These sit on a scale, not in boxes. Many good "agents" are mostly a fixed workflow with an AI step or two, and that is often the safer design. Robotic process automation is still the right answer for work that never changes. Use an agent where the work genuinely varies.
What is MCP, and why does it matter for agents?
MCP is an open standard that lets an AI agent plug into business tools and data without custom code for each one.
Anthropic released the Model Context Protocol as an open standard in November 2024, and in December 2025 donated it to the Agentic AI Foundation, a fund under the Linux Foundation co-founded with Block and OpenAI. The MCP project lists support across Claude, ChatGPT and widely used developer tools. In plain terms, MCP does for AI what USB did for devices: one standard plug, so a connector built once works with many AI tools.
For a business, it means connectors to the systems you already use are increasingly off the shelf, and an investment in connections is less tied to one AI vendor. It also means every connector is a door into your data, so treat installing one like installing any other software with access to client records. Our earlier explainer, what is this MCP thing, covers the basics.
What can AI agents do for a NZ business?
The agents working well today handle bounded, repeatable work with a person checking the output.
Client case story · Property management
Aspire Property Management: a virtual property manager
Aspire's assistant reads inbound tenant and owner messages, looks up the company's procedures, the Residential Tenancies Act and its insurance policies, and drafts the reply for a property manager to check and send. Read the Aspire story.
Publicly reported · Appliances
Fisher & Paykel: customer support agents
Fisher & Paykel announced in November 2024 that it would use Salesforce Agentforce agents to answer customer questions and schedule service appointments.
Publicly reported · Telecommunications
One NZ: agents across the business
One NZ's chief executive told Mi3 in October 2025 the company had deployed close to 100 agents across marketing, customer service and network operations.
Customer service
Inbox triage
Sorting inbound email by topic and urgency, pulling the customer's history, and drafting a reply for approval.
Sales
Quotes from email
Reading an order request, checking stock and price lists, and building a draft quote or sales order.
Finance
Reconciliation checks
Matching invoices to purchase orders and flagging the exceptions a person needs to look at.
Research
Prospect and market briefs
Researching a company before a meeting and writing a one-page brief with sources.
Aspire is an AI New Zealand client; Fisher & Paykel and One NZ are as publicly reported, and we have not verified their results. The other examples are illustrative, by sector. Platforms for building agents include Microsoft Copilot Studio, Claude Cowork, workspace agents in ChatGPT and the Wellington-built Numa. For time saved per task across the use cases we have scoped with clients, see AI examples from New Zealand businesses.
What are the common myths about AI agents?
- "Agents replace staff." In our work, agents take on the drafting and the looking up. People move to checking, deciding and handling exceptions, and the work they keep gets more valuable, not less.
- "You can tell an agent to behave." Instructions in a prompt are suggestions. Real limits are technical: what the agent can access, what it can change, and what needs approval.
- "Agents need perfect data first." They need clear, current procedures and a defined task more than a data lake. Encoding how you work is usually the first job.
- "An agent is a product you buy." It is a configuration of a model, tools, knowledge and rules. The platform matters less than the process it runs. See our comparison of AI tools for the options.
- "Set and forget." Agents drift as data, models and processes change. Someone has to own their quality.
How does New Zealand law apply to AI agents?
There is no agent-specific law in New Zealand. Existing law applies, and you stay accountable for what an agent does in your name.
The Government's AI Strategy takes a light-touch, principles-based approach that relies on existing law, and we have found no New Zealand guidance written specifically for agents yet. The Privacy Commissioner's AI guidance applies to them in full, and recommends a privacy impact assessment before AI is used on personal information. Principle 8 of the Privacy Act 2020 matters most for agents: you must take reasonable steps to check information is accurate before you use or disclose it, so an agent's output needs checking before it reaches a customer or a decision. Public sector agencies also have the Algorithm Charter for Aotearoa New Zealand, a useful model for any organisation using automated decisions.
In practice: if an agent drafts a customer reply, you are responsible for what it says. If it uses personal information, the Privacy Act applies as if a staff member had done it. Our guide to AI and the Privacy Act goes through the principles in plain English.
Not legal advice. This page explains the general position. For decisions about your organisation, talk to a lawyer.
Our view
Agents are where AI starts to touch the operating model. A chatbot makes one person faster. An agent changes how a piece of work flows through the business, which is where the value is, and where the risk is.
Two rules follow. Prompts are not controls. Limit what an agent can reach, keep a person approving what leaves the building, and log what it does. And agent assurance is a budget line, not an optional extra. Once agents run real work, someone has to sample their output, watch for drift and tighten the rules when quality slips, the same way a factory assures its line.
Start with one workflow, one number to move and one owner. Our AI Operating System is how we build agents for clients: governance first, then process, knowledge, connections and, last, the skills and agents themselves.
What are the risks of AI agents?
Agents carry every risk of a chatbot, plus the risks that come with acting.
- Prompt injection. An agent that reads email, web pages or documents can be hijacked by instructions hidden inside them. It is number one on the OWASP list of AI security risks. See prompt injection in our glossary.
- Excessive agency. Too many tools, too much permission or too much autonomy. Give each agent the least access it needs, as OWASP recommends, and require approval for anything that sends, pays, deletes or changes a record.
- Confident mistakes. Agents can hallucinate and then act on the error. Read why AI models hallucinate and how to prevent it, and keep a person checking outputs.
- Cost that grows quietly. Many agent platforms bill by usage, such as Copilot Studio's Copilot Credits. Set budgets and watch them. Cost and unclear value are two of the reasons Gartner expects many agent projects to be cancelled.
- Drift and accountability. Agents change as models, data and processes change. Name an owner for each one, sample its work, and log what it does. OWASP's Top 10 for Agentic Applications is a fuller checklist for technical teams.
The free AI Champions Committee resource pack gives you a governance structure to own these risks, and our AI policy guide covers what to write down.
Where should a NZ business start with AI agents?
Start small, read-only and supervised, then widen what the agent can do as it earns trust.
- Find out where you stand. The free NZ AI Benchmark is 40 questions, takes about 20 minutes, and shows whether your governance and processes are ready for agents.
- Pick one workflow. Repetitive, multi-step, with a clear owner and a number you already track, such as response time or hours per week. Measure it before the agent starts.
- Write the procedure down. An agent can only follow a process that exists. Our free setup checklists for Claude and Copilot show how to encode procedures, tone and templates.
- Start read-only, then measure. Let the agent read and draft, with a person sending, and compare time, quality and errors with the baseline. Add write access one tool at a time, as the evidence allows.
- Build the skills. The Claude at Work and Microsoft Copilot courses cover the hands-on side, and the AI Leaders Course covers the governance.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers a question and waits for the next one. An agent is given a goal, plans the steps, uses tools such as email, files and business systems to carry them out, and checks its own progress. A chatbot's main risk is a wrong answer; an agent's is a wrong action, so agents need tighter limits and human approval.
What is an example of an AI agent?
A virtual property manager that reads an inbound tenant email, looks up the company's procedures and the relevant law, and drafts a reply for a property manager to approve. Other common examples are inbox triage, building quotes from order emails, and researching a prospect before a sales meeting.
Is Microsoft Copilot an AI agent?
Copilot is mostly an assistant: you ask, it helps. Microsoft also lets organisations build agents with Copilot Studio, which can take actions across Microsoft 365 and connected systems. ChatGPT and Claude now offer agent features alongside chat too, through workspace agents in ChatGPT and Claude Cowork.
Are AI agents safe to use with customer data?
They can be, with the right setup. Use business plans with clear data terms, give the agent the minimum access it needs, keep a person approving anything that goes to a customer, and log what it does. The Privacy Act 2020 applies to personal information an agent handles, just as it would for a staff member.
Do I need a developer to build an AI agent?
Not for simple ones. Tools such as Copilot Studio, Claude and Numa let non-developers build agents that read documents and draft work. Connecting an agent to core business systems, and making it reliable and secure, usually needs technical help.
Will AI agents replace jobs?
In our experience roles change more than they disappear. Agents take on drafting, looking up and sorting, and people move to checking, deciding and handling the exceptions. New work appears too: someone has to own each agent's quality.
Sources
- Building effective agents, Anthropic (December 2024; accessed 19 October 2026)
- A practical guide to building agents, OpenAI (2025; accessed 19 October 2026)
- Introducing the Model Context Protocol, Anthropic (November 2024; accessed 19 October 2026)
- Donating the Model Context Protocol and establishing the Agentic AI Foundation, Anthropic (December 2025; accessed 19 October 2026)
- modelcontextprotocol.io, Model Context Protocol project (specification July 2026; accessed 19 October 2026)
- The state of AI in 2026, McKinsey (August 2026; accessed 19 October 2026)
- Gartner predicts over 40 percent of agentic AI projects will be canceled by end of 2027, Gartner (June 2025; accessed 19 October 2026)
- Top 10 for LLM Applications 2025, OWASP GenAI Security Project (2025; accessed 19 October 2026)
- LLM01: Prompt injection, OWASP GenAI Security Project (2025; accessed 19 October 2026)
- LLM06: Excessive agency, OWASP GenAI Security Project (2025; accessed 19 October 2026)
- Top 10 for Agentic Applications 2026, OWASP GenAI Security Project (December 2025; accessed 19 October 2026)
- Artificial intelligence (AI), Office of the Privacy Commissioner (September 2023; accessed 19 October 2026)
- Privacy Act 2020, section 22, Parliamentary Counsel Office (version as at 1 October 2026; accessed 19 October 2026)
- New Zealand's Strategy for Artificial Intelligence, MBIE (July 2025; accessed 19 October 2026)
- Algorithm Charter for Aotearoa New Zealand, Stats NZ (2020; accessed 19 October 2026)
- Copilot Studio, Microsoft (product page; accessed 19 October 2026)
- Claude Cowork, Anthropic (product page; accessed 19 October 2026)
- Introducing workspace agents in ChatGPT, OpenAI (April 2026; accessed 19 October 2026)
- Fisher & Paykel and Agentforce, Salesforce (November 2024; accessed 19 October 2026)
- One NZ goes fully agentic, Mi3 (October 2025; accessed 19 October 2026)
Related topics
Keep going.
Your next step
Find the first workflow worth giving to an agent.
Book a 20 minute AI Chat and we will look at where an agent could take real work off your team, and what it would need to be safe.
Changelog: 19 October 2026, first published.