ChatGPT vs Claude vs Copilot: which one should a New Zealand business pick?

Start with where your work already lives

The fastest way to narrow the choice is to look at where your documents, email and calendars sit today. Microsoft Copilot is built into Word, Excel, PowerPoint, Outlook and Teams. It grounds its answers in your tenant through Microsoft Graph, and it only shows organisational data the user already has permission to view⁵. If your firm runs on Microsoft 365 and most of the value you want is inside those apps, Copilot starts with an advantage the other two have to work to match.

ChatGPT takes a different route. It is a separate workspace that reaches into your other systems through connected apps. With company knowledge switched on, a Business or Enterprise user can ask a question and ChatGPT searches connected tools such as SharePoint, Google Drive, Slack, GitHub and HubSpot, then answers with citations back to the source². It respects the permissions each user already has in those systems, and an admin decides which connectors the workspace can use². That suits a business with a mixed stack: Google Workspace for documents, Xero for accounts, a CRM, and Slack or Teams for chat.

Claude sits closer to ChatGPT in shape: a separate workspace with its own projects and file handling, used mainly for long reading, careful writing and analysis. On the data side, Anthropic says it does not use inputs or outputs from its commercial products, including the Team and Enterprise plans, to train models by default⁸.

• Mostly Microsoft 365, and the work happens in Outlook, Word, Excel and Teams: test Copilot first.
• Mixed systems, a lot of drafting, research and spreadsheet analysis outside Office: test ChatGPT first.
• Heavy reading and writing of long documents, such as contracts, tenders and reports: put Claude in the trial as well.

Where ChatGPT is the strong choice

ChatGPT is the most general of the three. People already know it, and that matters in a twenty-person business where nobody has time for a long change programme. Its working features cover most office tasks: Projects keep related chats, files and instructions together so the context carries across a week of work¹. Projects can be shared with colleagues on Business, Enterprise and Edu, with chat or edit access¹.

Skills let one person package a repeatable task once and share it with the team when workspace settings allow, and plugins bundle skills with connected apps¹⁵. Typical examples are a quote-letter writer that knows your terms of trade, or a policy helper loaded with your staff handbook. Older advice points to custom GPTs for this, but personal plans can no longer create new GPTs¹⁴, and OpenAI is retiring custom GPTs in favour of plugins¹⁶. The current retirement date is 11 December 2026¹⁷. Workspace agents go further and can own a whole multi-step workflow that a team shares¹¹. Treat those as a second-stage tool, after people are confident with the basics.

A generic worked example. Imagine a fifteen-person Tauranga property management office that uses Google Workspace, a separate property system and Xero. Copilot would add little, because very little of their work sits in Microsoft 365. In a ChatGPT Business trial they build a shared project for tenancy correspondence. They also make a skill that turns inspection notes into a tidy report, and use data analysis on an export of arrears. The test that matters is not how clever the output is. It is whether the property managers still use it in week three.

Where Copilot or Claude is the better fit

Be honest about the cases where ChatGPT is second best. If the main goal is to summarise Teams meetings, triage a full Outlook inbox or build a deck from files already in SharePoint, Copilot does that inside the apps people already have open⁵. Its prompts and responses also fall under Microsoft's enterprise data protection when staff sign in with their work accounts⁶. For a Wellington consultancy that lives in Teams all day, that convenience usually beats a better chat experience somewhere else.

Claude is worth a serious look when the core work is long-form reading and writing: tender responses, policy reviews, research summaries, legal or technical drafting. Some firms run Claude next to ChatGPT for exactly that reason. You do not have to pick one tool for every job. You do have to pick one approved tool per job, or staff drift back to personal accounts.

• Copilot: best when the value is inside Microsoft 365 apps and your permissions in SharePoint and OneDrive are already tidy⁵.
• ChatGPT: best as a general assistant across mixed systems, with Projects, skills and connectors^1,2,15.
• Claude: best for careful work on long documents, with commercial plans excluded from training by default⁸.

The data question: compare plans, not brands

Most of the privacy risk sits in which plan your staff are on, not which brand they chose. For ChatGPT the line is clear. On Free, Go, Plus and Pro, content may be used to train models unless the person opts out³. On Business, Enterprise, Edu and the API, OpenAI does not use your content to improve its models by default³. A business that lets staff use personal ChatGPT accounts for client work has a governance problem whichever tool it later buys.

Data residency is the next filter, and here the three differ in ways a New Zealand buyer should check. OpenAI offers data residency for ChatGPT in a list of regions that includes Australia, but only for eligible Enterprise and Edu customers⁴. ChatGPT Business is not on that list. Microsoft offers Copilot data residency under its Product Terms in some countries. New Zealand is not on the Product Terms list, so local storage needs the Advanced Data Residency or Multi-Geo add-on⁷. If your contracts or sector rules require data to stay in New Zealand or Australia, ask each vendor for the specific commitment in writing before you choose.

The Privacy Commissioner expects a business to check whether a generative AI tool is necessary and proportionate, to get senior leadership approval, and to have a human review outputs before acting on them¹². Those duties are the same whichever product you pick. A tool comparison that ignores them is not finished.

Run a two-week test, not a debate

Opinions about AI tools are cheap. A short, structured trial with your own work gives a better answer than any comparison table, this one included. MBIE's responsible AI guidance for businesses suggests starting from your reason for using AI and the problem you want to solve, which is a good discipline for the trial design¹⁰.

Keep the trial small and real. Pick four to six people who do different work. Give each of them the same three to five tasks they already do every week. Use business plans or trial workspaces only, never personal accounts. Set a simple rule for data: no client personal information until you have checked the plan settings and finished a privacy assessment.

• Spend the first morning listing the five most common writing, summarising and analysis tasks in the business.
• Write down how each task is done today and who checks it.
• Run the same tasks in each tool you are testing, using anonymised or public material.
• Score each result on accuracy, edit time and whether the person would use it again.
• Ask your IT provider to check sign-in, admin controls and connector options for each plan.
• Decide on one approved tool per kind of work, and write that into your AI use policy.

Bring your staff with you

Choosing a tool changes how people work, so treat it as workplace change. Under the Employment Relations Act 2000, employers must act in good faith. When a proposal could affect someone's duties or role, that means giving the affected staff information about the change, time to comment and real consideration of their feedback before deciding¹³. Many small rollouts will not change anyone's duties. When one will, consult first.

The NCSC's advice for small businesses is to set an internal AI use policy that says clearly which data cannot be put into AI platforms, and to check vendors' data handling against your own needs⁹. Write that policy before the trial ends, not after. A one-page version that names the approved tool, the plan, the banned data types and the person to ask is enough for most small firms.

A quick decision guide

Use the questions below as a short checklist at the end of the trial. If two tools score the same, pick the one your staff kept using without being reminded.

• Do most of our documents and meetings live in Microsoft 365, with tidy permissions? Copilot earns a place⁵.
• Do we need one assistant that reaches several systems, plus shareable Projects and skills? ChatGPT Business fits^1,2,15.
• Is the main job long, careful reading and drafting? Add Claude to the shortlist.
• Does a contract or regulator require local data storage? Check Enterprise-level residency options in writing first^4,7.
• Are any staff using personal accounts for work today? Fix that before anything else³.