Use Microsoft Copilot if your policies, letters and staff queries live in Microsoft 365, because it answers from the files and permissions you already have. Use ChatGPT Business if your documents sit elsewhere or you want a separate drafting workspace. Seat cost is close: $20 a user for ChatGPT Business, $21 for the Copilot Business add-on, and Copilot Chat is free.
Either way, staff records are the sensitive part. Pick a business plan, not a free account, and decide what HR data never goes in. The vendor scenario pages that fill the search results show what each tool can do, but not how a small people team should choose. Four questions about where your documents live and who may see them settle most cases.
Four questions that settle most small-team choices
Answer these in order, and stop as soon as one gives a clear result.
| Question | If yes, lean towards | If no, lean towards |
|---|---|---|
| 1. Do your policies, contracts and handbook sit in SharePoint, OneDrive, Outlook or Teams? | Copilot, because it can work from those files with their existing permissions | ChatGPT Business, with documents added to a project or connected from another store |
| 2. Does your organisation already have a Microsoft 365 business plan that includes Copilot Chat? | Start on free Copilot Chat and see how far it gets | ChatGPT Business, since the Copilot add-on needs an eligible plan underneath |
| 3. Is most of the work drafting new text, or answering from existing documents? | New drafting: either. Answering from your own files: Copilot, if the files are in Microsoft 365 | Drafting without files: ChatGPT Business is the simpler workspace |
| 4. Who may see the output, and who may see the source files? | If access must follow existing folder permissions: Copilot | If you want a closed project with named members: ChatGPT Business shared projects |
Most small teams get a clear answer at question one. A people team of three inside a driving school with 45 staff, all of whose policies are in SharePoint and whose email is Outlook, lands on Copilot. A photography studio of six, with its contracts and handbook in a shared drive from another provider, lands on ChatGPT Business. The broader comparison of ChatGPT and Copilot for a Microsoft 365 business covers the same choice for whole-company use, where the people team's needs are one voice among several.
Seat costs for a three-person people team
Costs come from the vendors' pricing pages as read on 1 November 2026. Microsoft sells the paid Copilot, listed as Microsoft 365 Copilot Business, as an add-on to an eligible Microsoft 365 business plan on an annual commitment, so the figures below assume the base plan is already paid for the whole company. Microsoft renamed the Microsoft 365 Copilot app to Microsoft Copilot in August 2026, so you will see both names on its pages.
| Option | Price per user a month | Three users a month | Notes |
|---|---|---|---|
| Copilot Chat (included with eligible Microsoft 365 business plans) | $0 extra | $0 | Web-grounded chat, in Outlook on mail and open files, with enterprise data protection |
| Microsoft 365 Copilot Business add-on | $21 paid yearly; $25.20 paid monthly | $63 yearly billing; $75.60 monthly | For organisations up to 300 users; Microsoft shows an $18 first-year offer for purchases between 1 July and 31 December 2026 |
| ChatGPT Business, Standard seats | $20 paid yearly; $25 billed monthly | $60 yearly billing; $75 monthly | Two seats minimum, 2 to 200 users |
| Microsoft 365 Business Standard with Copilot bundle | $23.50 paid yearly | $70.50 | Includes the base plan, so it only compares fairly when you have no Microsoft 365 plan yet |
On yearly billing the paid options for the three-person team sit within $11 a month of each other, from $60 to $70.50. The $18 Microsoft offer brings three Copilot users down to $54 a month in year one, then to $63. That is why the price should be the last factor and not the first. The breakdown of Microsoft 365 Copilot pricing for business and enterprise shows the licence tiers if you outgrow the small-business plan.
Seat price is also not the whole bill. Copilot works with whatever your permissions allow, so tidying them may take a few hours, which is a real cost to include. ChatGPT Business needs two admins' time for settings and a written rule. Say six hours of tidying for Copilot against four of settings and policy for ChatGPT Business: at an internal cost of $50 an hour that is $300 against $200, once. Neither changes the answer, but both belong in the comparison.

The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhat free Copilot Chat already does for an HR team
Microsoft describes Copilot Chat as a web-grounded assistant included at no extra cost with eligible Microsoft 365 plans. Its enterprise data protection means that, when staff use it signed in with their work account, prompts and responses are covered by Microsoft's data protection addendum and product terms, are not used to train foundation models, and sit inside your tenant's permissions, retention and audit settings. Microsoft Learn also notes that Copilot Chat in Outlook works on the inbox, calendar and chats without the paid licence, and on files you have open.
For a people team this is enough for several routine jobs. You can ask it to draft a job advert from a bullet list, tidy a long-winded policy paragraph, turn a meeting note into an action list, or write the first version of a reminder email. What it does not do on the free tier is search across all your SharePoint HR files for an answer. That is what the paid licence adds, by grounding answers in your work data through Microsoft Graph. The comparison of Copilot Chat and Microsoft 365 Copilot explains what you get free, and a month on the free tier is a cheap way to see whether the paid add-on is worth it.
One check before anyone relies on the data protection: make sure each person is signed in with the work account, not a personal Microsoft account. A personal login sits outside your organisation's agreement with Microsoft, so an HR officer who pastes a letter into it has bypassed everything above.
The same leave-policy question in each tool
Take the question staff ask most often: "Can I carry over unused leave?" Here is how each tool should be set up to answer it, and how it can go wrong. Illustrative throughout.
With paid Copilot, the HR officer asks in Copilot Chat: "Using the current leave policy on the HR site, can an instructor carry over unused leave? Quote the clause." Copilot searches the files that person can open, finds the policy, and answers with a quote.
With ChatGPT Business, the policy lives in a shared project with the instruction below. The project holds the current handbook as a PDF and nothing else.
You answer staff questions from the HR documents in this project only.
1. Quote the clause number and the exact words you rely on.
2. If the documents do not answer the question, reply "Not covered in the
documents; ask HR" and nothing else.
3. Do not give legal advice or predict how a decision would be made.
4. Never ask for or repeat an employee's name, pay or health details.
Both approaches can fail the same way. In the driving school, the SharePoint folder held "Leave Policy v4" and an old "Leave Policy 2022 (do not use)". Copilot's illustrative answer quoted a carry-over limit of five days from the old file, while the current policy says three. The fix was administrative: move old versions to an archive folder with narrower access, and add "use only the file marked CURRENT" to the prompt. ChatGPT Business cannot make that mistake with a single-file project, but it can answer from its general knowledge if instruction 2 is missing, so the instruction matters more than the product.
Letters: a warning letter drafted with the names stripped
Letters are where employee data creeps in, because the natural prompt is the whole file. Strip it first, in either tool.
Before: Draft a first written warning for [full name], an instructor, for
three late arrivals in October, referencing the call with [manager name]
on 14 October and her health condition.
After: Draft a first written warning letter template for an employee who
arrived late on three occasions in one month. Include: a factual summary
using [Employee], [Dates] and [Manager] placeholders, the standard
expected, the improvement needed within [Period], how to respond
or appeal, and the consequence of no improvement. Neutral tone, no
reference to health or personal circumstances.
The after version takes about a minute longer to write, produces a reusable template, and keeps the name, the dates and the health detail out of the prompt. The people team fills in the placeholders afterwards in Word. What the assistant must never do is decide whether a warning is justified or fair. A human reviews the final letter against the policy, and a qualified adviser checks anything that could end in a dispute or a legal claim. The tutorial on whether free AI plans are safe with staff data is the companion for why the free tier is the wrong place for any version of that letter.
Job adverts: a sample output and the lines a person must remove
Adverts are the safest place to begin, since they hold no employee data and every draft is reviewed before publication. The prompt is the same in either tool.
Draft a job advert for a part-time office administrator at a driving
school with 45 staff. 20 hours a week, mornings. Duties: bookings diary,
answering phones and email, invoicing. Salary band [band]. Use plain
language, under 250 words, and include a closing date placeholder.
An illustrative first draft opens with "We are looking for a young, energetic team player to join our fast-moving school" and later says "ideal for a recent graduate". Both phrases hint at an age preference, and neither was in the instruction. The assistant also left the salary band as a vague "competitive salary" and made up a start date.
The person reviewing should do four things. Delete the age-coded wording. Put in the real salary band and hours. Replace the invented start date with the closing-date placeholder. Finally, add the equal-opportunities line your adviser has approved. This is a good first test because both tools make the same kind of mistake, which shows that the human review step, not the product, is what protects the school. Log how many minutes the edit took, and compare it with how long a blank page takes.
Staff queries: what a tool may answer and what goes to a person
A people team gets the same questions repeatedly, and assistants can handle some of them. The table separates what to automate from what to keep human.
| Staff question | Who answers | Why |
|---|---|---|
| How many days' notice do I give for holiday? | Assistant, quoting the policy | It is in a document, and the answer is the same for everyone |
| What is my leave balance? | The HR or payroll system | It is a record lookup, not a language task |
| How do I submit expenses? | Assistant, quoting the process | Procedural and low risk |
| I think I was underpaid last month | A person | It needs the payroll record and a human reply |
| I want to raise a concern about my manager | A person, quickly | Sensitive, and mishandling it has serious consequences |
| What happens to my pay if I am off sick for a month? | Assistant for the policy wording, then a person for the individual case | The policy is general, the case is personal |
| Can I work flexibly after my probation ends? | A person | It involves discretion and possibly legal rights |
| Where is the new starter checklist? | Assistant, linking to the file | Findability only |
Whichever tool you pick, publish that split to staff. An assistant that answers a grievance question in a friendly tone is worse than no assistant. The tutorial on answering staff HR questions with an AI helpdesk shows how to build the routing in more detail.
Employee-data handling rules that matter more than the tool
The choice between ChatGPT and Copilot is narrower than the choice of rules around them. A short set covers most risks.
- Business plans only for anything about a person. OpenAI says Business workspace content is not used for training by default, and Microsoft says prompts and responses in Copilot are not used to train foundation models. Personal and free accounts have different terms.
- Minimise before you prompt. Use role titles and placeholders, not names. Leave out health, family, disciplinary history, pay and ethnicity or other protected details unless the task truly needs them.
- Know who can see what. Copilot respects permissions, so it exposes whatever a user can already open. That is the reason to read the guide to cleaning up SharePoint permissions before Copilot before you give anyone the paid licence. ChatGPT Business admins can view workspace conversations according to OpenAI's privacy page, so tell staff.
- Keep memory and sharing under control. In ChatGPT Business, shared projects use project-only memory, and the workspace link gives all members chat access, so invite named people only.
- Check retention. Microsoft applies your own retention policies to Copilot interactions. OpenAI says deleted Business conversations leave its systems within 30 days. Record both in your data-handling notes.
- Write down the rule and date it. Employee data may fall under data-protection law that treats some categories as sensitive, so ask your adviser what applies to you, and show staff the one-page rule.
Microsoft's page on enterprise data protection for Copilot is the official reference for the first Microsoft rule, and OpenAI's enterprise privacy page is the matching one for ChatGPT Business.
Scoring the choice for two example firms
When the four questions do not give a clear answer, score the options. Weight each criterion 1 to 5 for importance, score each tool 1 to 5, and multiply. The table is filled in for two illustrative firms to show how the same method produces opposite answers.
| Criterion (weight) | Driving school: Copilot | Driving school: ChatGPT | Photography studio: Copilot | Photography studio: ChatGPT |
|---|---|---|---|---|
| Documents already in the tool's reach (5) | 5 | 2 | 1 | 4 |
| Tenant-level protections and permissions (4) | 4 | 4 | 2 | 4 |
| Seat cost for the people team (3) | 4 | 4 | 2 | 4 |
| Draft quality in a one-week test (3) | 3 | 4 | 3 | 4 |
| Admin effort to set up (2) | 3 | 3 | 2 | 4 |
| Weighted total (max 85) | 68 | 56 | 32 | 68 |
The driving school scores Copilot higher because its files are already in Microsoft 365. The photography studio, with no Microsoft tenant to ground in, scores ChatGPT higher for the same reason in reverse. The tool did not win either contest on quality, which came out nearly level in both. Location of documents and the cost of preparing them decided the result, and that is true for most small teams.
Three first-month failures and their fixes
- The licence goes to everyone. A firm buys Copilot for all 45 staff, when only three do HR work, and pays for 42 seats nobody needs. Fix: license the people team first, check usage after 30 days, and add others only where a task justifies it.
- Staff paste from a personal login. An officer under time pressure uses a personal account because the work one hit a limit or was slow. Fix: write the "work account only" rule, make the work account easy to reach on every device, and ask people to tell you when it fails, not hide it.
- The assistant answers beyond the policy. It cheerfully answers a redundancy question from general knowledge. Fix: the "not covered, ask HR" instruction above, and a monthly check of five real answers against the policy.
A one-week trial to settle it
- Monday: list five repeating tasks (a job advert, a policy rewrite, a staff-query answer, a meeting action list, a reminder letter) and five staff questions.
- Tuesday: run the five tasks in free Copilot Chat. Log minutes to a usable draft.
- Wednesday: run the same five in a ChatGPT Business trial or one month of two seats ($50 monthly).
- Thursday: run the five staff questions, with the leave-policy set-up above in each tool, and check every quote against the source.
- Friday: total the logs, score the table above, and pick one. If Copilot Chat is enough for the five tasks, the right amount to spend may be nothing.
If you trial the paid Copilot rather than the free chat, note that Microsoft's one-month Copilot Business trial turns into a paid subscription automatically, so put the end date in the diary on the day you start it and check the current trial terms first.
That last point deserves emphasis: for many small teams the best plan is the free one in the Microsoft account they already pay for, plus a firm rule on what never goes in, with a paid seat added for the one person whose work genuinely needs it.
Questions small HR teams ask about ChatGPT and Copilot
Can an HR team use the free Copilot Chat instead of paying?
Often yes for drafting and general questions, if your organisation has an eligible Microsoft 365 business plan and staff sign in with their work accounts. Microsoft says Copilot Chat comes at no extra cost with eligible plans and applies enterprise data protection. It does not search all your SharePoint files the way the paid licence does.
Does Copilot show HR files to people who should not see them?
It only surfaces what the signed-in person already has access to, which is the problem: if an HR folder is shared too widely, Copilot makes that visible. Fix the permissions on policy, case and payroll folders before you switch on the paid licence for anyone.
Is ChatGPT Business safe for employee records?
OpenAI says workspace content is not used for training by default, and the plan has SSO and MFA, but admins can view workspace conversations. It suits anonymised or minimal data. Keep names, health details and case files out unless your policy, your contracts and an adviser say otherwise.
Could a team use both tools?
Yes, and some do: free Copilot Chat for everyday drafting inside Microsoft 365, and ChatGPT Business seats for the two people who need a project workspace. The cost is a second set of rules to teach and police, so only do it if the first tool leaves a clear gap.
Further reads
- HR Automation AI Tools: What a Small Business Can Automate Safely — What a small business can automate in HR safely.
- AI for HR in Small Businesses: 12 Tasks You Can Hand Over Safely — Twelve HR tasks that are safe to hand over.
- Using AI for HR: Probation, Renewal and Right-to-Work Deadline Checklist — A deadline tracker for probation and renewals.
- ChatGPT for HR Professionals: Job Ads, Reviews and Warning Letters — Prompts for job ads and review write-ups.
- How to Roll Out Microsoft 365 Copilot to a Small Team — A rollout plan if you choose Copilot.
- AI Tools for HR Professionals: Policies, Onboarding and Staff Questions — Policies and onboarding work AI can speed up.
- AI Survey Analysis for HR Consultants: Coding Employee Comments Safely — Code employee survey comments into themes with AI, spot-check the sentiment, and apply anonymity thresholds before anything reaches a slide.
- Microsoft AI Chatbot for Business: Copilot Chat vs Microsoft 365 Copilot — Free Copilot Chat or the paid licence? A plain choice with list prices, a worked three-user cost and a data protection check.
- Can AI Draft Disciplinary Letters? What HR Consultants Must Check — A process-first way to draft invitation, outcome and appeal letters without inventing evidence or decisions.
- Care Staff Supervision Template: Notes and Action Logs With AI — A care-specific supervision agenda, record format and action log, with AI drafting the actions from your notes while staff data stays inside approved tools.
- AI HR Software for Startups: Which Tool Fits Which Job — Which AI HR software suits your startup stage? Compare Rippling, Deel, and Gusto across onboarding, policies, scheduling, and payroll admin.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft's Copilot for business pricing page, Microsoft Learn on enterprise data protection, OpenAI's business pricing page, its enterprise privacy page and help articles on Business, read on 1 November 2026.