launchlayer
WhatsApp

Testing generative AI in an SME: ChatGPT, Claude, Gemini or Perplexity, where do you start?

Opening ChatGPT and asking for a poem about your company is not a test. A test is a task your team already does every week, handed to an assistant for two weeks, with a score at the end. Here is how to go about it.

  • Updated on 16 November 2026
  • 3 min read
01

What a test has to prove

A free trial shows that an assistant can write. You knew that already. The real question is narrower: on this task, with your documents, how much time do you save once the review is counted?

One impressive result on a hand-picked example proves nothing. What counts is the tenth try, on a badly scanned PDF or a half-readable customer email. That is where you see whether the assistant holds up or whether someone has to redo its work.

02

ChatGPT, Claude, Gemini, Perplexity, Le Chat: how they differ

They can all draft, summarise and rephrase. The differences are more about where they plug in and what your team already uses. We do not give prices: plans change often, so read them on each vendor's site on the day you choose.

  • ChatGPT (OpenAI): the best known. Chances are some of your staff already use it, on their personal account.
  • Claude (Anthropic): often picked for writing and for long documents, contracts or reports.
  • Gemini (Google): worth a look if you already work in Gmail, Docs and Drive, where it is built in.
  • Microsoft Copilot: the Microsoft 365 counterpart, inside Outlook, Word and Excel.
  • Perplexity: a search engine that answers with its web sources cited. Useful for research, less so for working on your own files.
  • Le Chat, by Mistral AI: the French vendor's assistant, worth including if provenance matters to you.
03

The method: two weeks, three real tasks

Pick three tasks that come back every week: for example summarising a customer's emails, pulling the details out of a purchase order, drafting a first version of a meeting report. Write down how long they take today. Then give them to two assistants, no more, for two weeks, on the real documents of the moment.

After each try, one person fills in the same grid. Five minutes is enough. After two weeks, you have your own numbers instead of a review site's opinion.

  • Time spent, review included, against the usual time.
  • Number of corrections before the result could be used.
  • Mistakes that would have slipped through without a review.
  • What had to be copied and pasted by hand to give the assistant context.
04

What you can hand to an assistant

Before pasting a customer file, read the vendor's page on how data is used. Depending on the plan and settings, your conversations may be used to improve its models. Team and business plans usually commit to the opposite: check it in their terms, not in an advert.

For the test, replace names, addresses and numbers with "Customer A" or "Supplier B". You are testing the task, not the assistant's memory. Personal data falls under GDPR as anywhere else, and the French data protection authority (CNIL) publishes guidance on generative AI. If you are unsure, ask before you paste, not after.

05

When the test shows AI belongs inside your software

Look at the last line of your grid. If the assistant does the task well but your team spends its time pasting extracts of files into it, the problem is no longer the AI. It is the copy and paste.

In that case, the AI has to read your data where it lives: in your business software, your shared inbox, your folders. That is what we build for SMEs, see how to build AI into your business software. If you do not know yet which tasks to test, our free AI assessment suggests three uses in two minutes. For funding to get support, see public funding for an AI assessment.

Published on 16 November 2026

Questions we get asked

Other guides

Want to look at your own case?

Show us the file and the process around it. In 30 minutes, we will tell you what is worth keeping, connecting or replacing.

Book a call