My knowledge when it comes to AI is pretty limited.
I don’t think it is that useful. I don’t like Scammy Altman and all the other Silicon Valley bros. I hate the privacy and policy aspect of it. The list goes on.
However, I was asked to roll out AI for a small 30-employee company.
Currently, it is a little bit of a wild west. Some employees already have a ChatGPT subscription and now rightfully want a company-provided license. Some don’t use it at all. Some use it to write emails, some to sketch stuff, some to analyze contracts. We have a 500-page PDF that would be great to be able to ask questions about.
So here are my few questions.
How do you guys handle the privacy aspect of it? Just give up? Local LLMs? 2. We are not coders. According to our employees, ChatGPT works well for them. But nobody has given Copilot a try yet. How does Copilot compare to ChatGPT? 3. I was thinking of getting Teams subscriptions. These are mostly around $20 per user (ChatGPT, Claude). However, since we already use Office 365 Business Basic, the additional costs would only be $7 per user for Copilot. Seems like the by far best offer. 4. OneNote is IMHO a clusterf***. Our boss hosts them all and shares them with every single employee because there is no real way to back up that shi*. Is it possible to read from OneNote so the LLM can “learn” from that? Apparently, Claude does not offer that. ChatGPT with a plugin and for Copilot, the only thing I found is that you can use Copilot inside of OneNote, but found nothing if you can read from Copilot stuff from our OneNote. 5. Basically the same as before but with mail. Some want to auto-let read certain mail folders and ask the LLM questions about it. This one seems to work with all of them. 6. We have one 4090 for GPU tasks. That PC is a floating PC for everyone that wants to do a GPU-heavy workload. I was thinking about installing Qwen3.8-27B on Ollama, unfortunately Windows. The idea would be to feed the 500-page PDF, plus publish the link for all employees in the local network. That way, all could connect to that Ollama interface and ask questions (of course, only one person can really work that way). 7.
While we are at publishing the local Ollama interface, we could also get ChatGPT to connect to that, so we could maybe even share a ChatGPT subscription? We have started moving the people who don’t really use AI to DeepSeek, there is no subscription so it’s perfect for the people only use AI a couple of times per week.
xyz 3
I would recommend local and lock it down as much as possible. Most LLM’s are major security risks and providing them least privilege access is best practice atm.
If selfhost is not an option then I’d at least get formal agreement with a preferred vendor. I do not recommend copilot: Critical Copilot vulnerability allowed hackers to steal 2FA code from users - Ars Technica