Chatbots

Chatbots: how to choose

Conversational assistants are the most general tools in this directory, and that is exactly what makes them hard to compare. Two products can look identical on a landing page and behave very differently once you give them a real task. What separates them is rarely the chat window. It is the model underneath, the size of the context they can hold, what they are allowed to do beyond writing text, and how the vendor handles what you type.

What actually differentiates them

Start with the model, not the interface. Most assistants are a wrapper around one or several large language models, and the wrapper matters far less than the model does. Check which models are available on the plan you would actually pay for, because vendors routinely advertise their strongest model while restricting it to the top tier or to a small monthly quota.

Context length decides what you can do. An assistant that holds a few thousand words is fine for short exchanges and useless for reviewing a contract. If your work involves long documents, this single number will matter more than every feature on the marketing page.

Look at what the assistant can do besides producing text. Reading files you upload, searching the web, running code, calling external services: each of these turns a writing aid into something closer to a working tool. Each also introduces a failure mode, since a wrong answer sourced from a live web page looks more credible than a wrong answer produced from memory.

Language quality is uneven and worth testing yourself. Nearly every assistant claims multilingual support. Fluency outside English varies widely, and it varies by task: an assistant can translate well and still write clumsy original prose in the same language.

What to check before committing

Read the data policy, not the summary of the data policy. The questions that matter are whether your conversations train future models, how long they are retained, and whether a business plan changes those answers. For anything involving client data, this is not a detail.

Test on your actual work rather than on a demo prompt. Assistants are optimised, consciously or not, for the kind of prompt people use to evaluate them. The gap between a clean demo and a messy real task is where the differences show.

Expect to verify. Every assistant in this category produces confident, well written statements that are sometimes wrong, and the failure is worse on niche subjects where you are least equipped to notice. Treat output as a first draft by a capable assistant who does not know your field.

How pricing usually works

The common shape is a free tier with rate limits, one paid individual plan, and a team or business plan that adds administration and data guarantees rather than raw capability. Per seat pricing is standard, so a small team can reach a meaningful annual cost quickly.

Watch for quotas expressed in messages rather than in words or tokens. A cap of a few dozen messages sounds generous until a single long document consumes several of them. Where an API is offered, usage based pricing often works out cheaper than seats for occasional or automated use.