AI Chatbot Platform Comparison Matrix: Claude, ChatGPT, Copilot, Gemini and Perplexity

This is a side-by-side comparison of Claude, ChatGPT, Microsoft 365 Copilot, Gemini and Perplexity as business platforms. It covers what each one costs once the licensing fine print is included, how it reaches Microsoft 365 and other business systems, which security and compliance controls exist at which tier, how ready each is for MCP and agent write actions, and where each one tends to fail. Every cell carries the sources it rests on.

I use this matrix in vCIO work, where the same document has to serve the owner deciding what to buy and the IT lead who will have to deploy and govern it. That is why it is ordered the way it is. The first sections answer what each platform is and what it costs, the middle sections cover Microsoft 365 integration and connectors, and the later ones move into compliance scope, MCP, failure modes and rollout effort. A reader can stop wherever the detail stops being useful to them.

Downloads

The Excel workbook is the fullest version and the one to open if you are making a decision with it. It carries the matrix, a pricing calculator where you set your own seat count, the full change log and every source with the date it was checked. The JSON holds the same data in one document for anyone who wants to automate against it.

How to read it

Each capability cell opens with a mark that summarises it, followed by the claim itself. Rely on the claim and its sources, which resolve to the full list at the foot of the page.

Two cautions apply to almost every row. Facts here are scoped to a plan: business and enterprise tiers of the same product differ in identity controls, logging, retention and residency, so a feature listed for one tier should not be assumed for another. And prices are public US list prices before tax, reseller discounts and promotions. They are a starting point for a quote, which is why the workbook calculator exists.

What each platform is and what it costs

Start here if you are deciding whether a platform belongs on the shortlist at all. The identity rows describe positioning, context limits and the agent products each vendor ships, and the pricing rows show the compared business tier for each platform at the default seat count.

The headline seat prices are not like for like, and that is the most common way these comparisons mislead. The Copilot figure is an add-on that needs a Microsoft 365 base license underneath it, while the Workspace figure buys the whole suite with Gemini included. The tables below set out the tiers organisations actually end up buying, including the bundles and enterprise paths, with what each price does and does not include.

Microsoft 365 integration and connectors

For most organisations the first technical question is how a platform reaches the documents, mail and chat already sitting in Microsoft 365. The answers differ more than the marketing suggests. Some platforms query content live with the user’s own delegated permissions, some build an index on the vendor’s side, and some copy files into their own storage when a user imports them. Each model has a different answer for where copies live, how quickly a permission change takes effect and what an administrator has to consent to.

Connector counts are reported here as each vendor states them, and they are not comparable. Catalogs mix first-party and partner connectors, read-only search with write actions, and curated lists with open directories. Compare the specific operations a workflow needs.

Security, compliance and admin controls

Certifications are scoped to named services. A SOC 2 report or a HIPAA business associate agreement covers the products it lists, and new agents, browser extensions and preview features are frequently outside that scope even when they ship inside a covered product. Where a row says to confirm, the vendor’s own documentation either conflicts or is silent, and the open questions further down say what to ask.

MCP readiness

The Model Context Protocol is how these assistants call tools and reach systems that have no built-in connector. The exposure grows with write access: a connected server can carry instructions the model follows, and a write action approved once can keep running. These rows track authentication, write approvals, logging and what has changed recently, because this is the part of the market moving fastest.

Failure modes and operational reality

Every platform here can produce a confident wrong answer, expose data through permissions that were already too broad, and change its pricing or terms between one renewal and the next. These rows record the specific, sourced version of each risk for each vendor, so the conversation can be about what has actually happened to each vendor.

Rollout effort

These rows are practitioner estimates and are labelled that way in the sources column. Rollout effort depends on the state of the environment far more than on the product, and a tenant with years of loose sharing will take longer to prepare for any assistant that searches it.

The real tradeoffs

Each choice gives something up. The hidden catch column is the one worth reading twice, because it is the cost that tends to surface after the contract is signed.

Solution patterns

Most organisations end up in one of the shapes below. The last one, MCP-led extensibility, is only worth starting once one of the others is in place and governed.

A rollout sequence

What changed in this refresh

The evidence date and the rebuild date are shown separately at the top of the page on purpose. A rebuild is not new evidence. Each entry below records what the previous version said and what is true now, and marks whether the difference came from the vendor changing something or from the earlier version being wrong.

Open questions

Where vendor documentation conflicts with itself or says nothing, the matrix records the gap and leaves the answer open. These are the questions to put to the vendor in writing before purchase.

Methodology and limitations

Facts come from vendor pricing pages, help centres, documentation, trust centres and official announcements, opened and read on the date recorded against each source. Press reporting is used only for dated events such as lawsuits, incidents and court rulings, and third-party captures only to flag a conflict that the vendor has not resolved. Each is labelled with its type in the source list.

The matrix records what vendors document and what reporting establishes. It does not measure answer quality, and it cannot tell you how a platform will perform on your own data, which is what a bounded pilot is for. Vendor documentation also changes without notice, which is the reason the page is refreshed on a schedule and states the date its evidence was checked.

The matrix is free to use with attribution under CC BY 4.0. Corrections are welcome and useful, particularly from vendors who can point to documentation that shows a row is out of date. There is more about me on the about page.