Introduction
We here at Practically have been playing with AI for a couple of years now and seeing what works best for us and for our clients. As such we get asked all the time about how best to start – see the bottom of this article for some of the questions we get.
The above slides form some of our advice – a little bit like an FAQ.
Case study - data fixed with AI
In the early days of AI a trade organization working within the education and eco space asked to us to advise on a data project. We said “Woah there”. Their plan was take all their historic and current data and use AI to reformat, collate and for this to form their cleaned DB.
Arguably AI has got a lot better at this sort of thing, but the flags were and are the following:
Uploading customer and company data to AI is never a good idea.
AI can merge and cleanup data, but it is not reliable.
Why? Uploading data to any AI other than a self-hosted LLM should not be trusted. Even on a paid for GPT account data goes into the main brain. Can you 100% say that data won’t be used?
AI can manipulate data well, as long as you can convince it to do the right thing. But, It gets bored, hallucinates and needs checking manually. Large language models by design will never return the same answer twice. This is not what you need and not how you might think about how tech should work. For example we experimented with using AI to mark simple questions on GCSE Prepper, but we found it not to get the most accurate answers.
A bigger question was “What is in this older data” and why do you need it. Or perhaps what is this old data worth to your organization. And here the conversation stalled as every new development piece should start with the WHY.
Saying that most organizations have data issues, formatting and mopping up requirements. We repeat, please do NOT upload your data to
Instead you should…
If you want to use sensitive data then you should either:
Make sure the paid for service really is secure. Read the terms and privacy. Make sure your data is siloed.
Or
Install a local or private LLM. The results are not as great but you know you have security of data.
Or
Get AI to write scripts to fix your data and run those on your server or locally. Create a sample of your data (preferably with fake data). Get AI to build the scripts and make sure it is doing it right. Remember the above note about AI not givining you the same answer twice? Well scripts and computers in general DO. So running scripts will do their thing, over and over again, as long as they can. So you had better make sure they are doing the right thing!
When it comes to data manipulation it should come as no surprise that this latter option is the preferred.
But it is not just data we get asked to look at. Some common questions we get asked to look at:
Can I have an AI chatbot on my site (yes!) that does this (maybe!)
Can I build my website with AI? (yes but…)
Can AI help with my content (yep but beware)
I have built this prototype using AI. Can you make it real please?
We shall spin out some more articles on the above.
