AI Assistants & Chat: Reviews & Rankings

AI Assistant buying guide

What should I actually expect to pay per person?

Three tiers, and the middle one is where most companies land. Personal plans sit at $20 a month — ChatGPT Plus, Claude Pro and Perplexity Pro all price there, and none of them give you admin control, so treat them as a way to evaluate rather than a way to deploy. Business and team plans run roughly $25 to $30 per seat per month on annual billing, with Microsoft 365 Copilot at $30 the reference point most CFOs already recognise. Enterprise means a phone call and a quote. Watch for two things the pricing page won't lead with: seat minimums (Claude's Team plan has historically required five) and the gap between the annual price advertised and the month-to-month price you'll actually pay if you're not ready to commit for a year.

Is the free version good enough to run a business on?

For evaluation, yes, and you should use it that way. For deployment, no — and the reason isn't capability, it's control. Free accounts are personal accounts. You can't see who's using them, you can't revoke access when someone leaves, and the data handling terms are usually weaker than the paid ones. Sensible pattern: run a two-week bake-off on free tiers with six volunteers, pick a winner, then buy seats properly instead of letting a shadow rollout happen one personal credit card at a time.

What does SOC 2 actually mean, and do I need it?

An independent auditor has examined how the vendor handles security and confirmed the controls do what the vendor claims. It is not a government certification and it guarantees nothing about model quality. What it does guarantee is that your legal review has something to read. If you're in finance, healthcare, law, or anything touching regulated client data, treat a SOC 2 Type II report as the entry ticket and end the conversation without one. For a twelve-person consultancy it's a nice-to-have — but separately check whether your prompts get used to train the model, because that's the question that actually bites, and on business plans the answer is almost always no.

How big a context window do I really need?

Smaller than the marketing suggests. Here's the conversion that matters: 100,000 tokens is roughly 200 pages. Nearly every business document you'd hand an assistant — a lease, a board pack, a contract, a quarter of meeting notes — fits inside 200,000 tokens comfortably. Above that you're into whole-codebase and whole-archive territory, which is a developer's problem, not yours. Our scoring caps at 256K for exactly this reason. Buy on setup, security and integrations. The context number stopped being a real differentiator somewhere in 2025.

My IT support is one part-time contractor. Which of these can I set up myself?

Most of them, genuinely, and that's changed in the last two years. Buying seats, inviting your team by email and switching on single sign-on are point-and-click in every major business plan now. Where you'll want help is connecting the assistant to your own systems — hooking it into your CRM or file store is usually an hour of somebody who knows your admin passwords, not a development project. Anything we rate developer required is a genuine build, and without an engineering team you should skip it. The Setup Effort column exists so you never discover that in week three.

Do I have to standardise on one tool, or can different teams use different ones?

Standardise on one for security and billing, then let people keep a second on the side. The models have noticeably different personalities and your teams will figure that out — writers gravitate one way, analysts another. Fighting it costs more goodwill than it saves in licence fees. The line worth holding is that anything touching customer or financial data goes through the sanctioned, SSO-protected account. What someone uses to rewrite a headline isn't worth a policy memo.

Everyone keeps saying MCP. Should I care?

Only for what it lets you avoid. MCP is an open standard for plugging an assistant into other systems, and the practical effect is that a tool supporting it can reach your own files, databases and apps without anyone writing custom integration code. Before it existed, connecting an assistant to an in-house system meant a developer and a project plan. Now it's often a configuration screen. You don't need to understand how it works. You need to check the box exists, because it separates an assistant that talks about your business from one that operates in it.

How is SquadScore calculated, and can anyone pay to rank higher?

No, and the method is published so you can check our working. Every spec in the table is either a raw published figure or a clearly-labelled editorial rating. Each gets normalised to a 0-100 scale — reversed where lower is better, as with price — then combined as a weighted average using the weights printed at the top of this page. When a vendor doesn't publish a figure we record it as unknown, drop it from that product's calculation and re-balance the remaining weights, so a gap in our research can never quietly score as a zero. The role lists reuse the same raw data with different weights, which is why one tool can win a list and finish sixth on another. That's the point.

What the SquadScore measures

The complete formula, bounds and data rules are published on the methodology page.