I Know Everyone’s Obsessed With AI; Here’s When It’s NOT Useful

Unless you’ve been living in the woods, deep inside a cave and under a rock, you’ve been hearing A LOT about AI.

Is AI going to take our jobs? When should we be using AI and when are good ol’ fashioned humans better? When should we do a “first draft” with AI and then bring in people to feng shui it?

As someone who’s been working in tech for 10+ years and helped hundreds of clients launch their tech products, here are my thoughts about AI.

What is AI anyway?

Ultimately, AI is a series of models that learn through trial and error and make decisions based on data.

Think about your Instagram feed. Let’s say you’re scrolling happily along and you get served a pimple-popping video – gross, right? Despite the bizarre popularity of this type of content, it’s not your thing, so you click through and choose “see fewer posts like this.”

Then you go back to your feed and get served several reels about the L.A. Lakers, whiskey-based cocktail recipes, and baby videos. If you’re me, you’re liking these posts and maybe even saving them for future reference.

AI has tried through trial and error to show me different pieces of content and it’s looked at the data – Jared Neutel doesn’t like pimple-popper videos, he DOES like content about the L.A. Lakers and whiskey-based cocktail recipes. It’s going to show me content based on that data.

In any situation, AI processes data and calculates how to provide recommendations or outputs based on previous attempts. In this example, it might show me content about the Los Angeles Kings hockey team (because it knows I like L.A. pro sports teams) or recommendations for cool cocktail bars in L.A. (because it knows I like whiskey and I live in L.A.)

What’s the best way to use AI in software development?

AI is great for web scraping and structuring unstructured data; it can identify patterns and organize information consistently.

Particularly if you’re gathering data in many different formats, from many different sources, AI can identify which data goes where and structure it accordingly. No matter what the platform looks like, where the data came from, it’ll work for your application.

We’re currently working with a pharmaceutical company that wants to build an AI scraper and it needs to be HIPAA compliant. That scraper will be able to gather public news and statistics, but it won’t gather any information that can be associated with an individual. AI models can make that choice, even if it’s not clearly marked in the text – incredible, right? A good AI can read large texts, infer context from them and make decisions, instead of relying on keywords or single paragraphs.

When should you use humans?

The truth about AI is that it gets things mostly right, most of the time. But there are many situations where 95% accuracy isn’t good enough! When it comes to things like financial data, health data, or even shipping calculations, AI models do this adequately. But in these types of situations “adequate” simply isn’t good enough. 

For example, I wouldn’t want AI to be in charge of sending ACH payments on its own with a 95% confidence window. What if it accidentally calculated how much money I should send someone because it misread an invoice?

Even when using words, AI can experience what we call “hallucinations”. AI never says “I don’t know.” It’s great at answering questions about the things it was trained on, but if it lacks some information it will simply make things up.

It’s bad enough when AI gives you an obviously absurd reply, but it is worse when the reply makes sense but part of it is wrong. Click To Tweet

This is why you should never, ever forget to verify. 

In situations where we want 100% accuracy, it’s incredibly important to have real, live humans with decades of experience and insights into the nuance of business and coding to double check things. (That’s why we pair our super senior engineers with AI tools when we run our NeuCheck code audits.)

What are the different ways we can use AI in software development?

There are three levels of AI integration. A good agency will help you decide how deep you want to go down the AI rabbit hole. 

Using existing AI tools and models

This might look like using chatbots to answer common customer questions or recommend products to customers based on what they purchased. It might look like using AI to summarize large documents or respond to specific questions. Or it might look like having the ability to read and extract information from photos of receipts taken with a phone.

Making modifications or adding layers to existing models

Let’s say you like how ChatGPT works and you think that with a few tweaks, it could work for your business. Our team of engineers has the ability to personalize AI to suit your needs. 

When you create agents or characters in ChatGPT – where you give the context of your business or application so the AI will give better, more targeted answers – you’re personalizing the AI. You can also feed it specific documents, like books or scientific papers and tell it to use that library as the preferred source of information. That process is called Retrieval-Augmented Generation or RAG.

Creating entirely new AI models from scratch

This is much less common, more expensive, and more time-consuming but it can certainly be done.  (And we have the Phd engineers who can do it for you if that’s what you truly need!) 

We reserve this approach for really specific and narrow scenarios. We had a client who exported meat from Brazil where there are incredibly specific tracking requirements – importers needed to know when the meat was cut from the animal, when it was processed, what the temperature was in the processing plant when all this was happening. Obviously, this is not the same AI as “write 5 social media posts based on this Google doc!”

We designed AI that used images from cameras placed around the processing plant. The AI
knew by the on-camera movement which piece of which bull was being cut, what that cut would become, etc.

For this client, it was easier for us to train an AI model than use an off-the-shelf solution. 

When should you use AI?

If you need 100% accuracy, which is most of the time the case, or if a project requires nuanced analysis or creative thinking or any type of math – you’ll want to pair any AI you use with humans. AI is a good assistant and a time saver, but they can’t replace actual human intelligence – and that is a good thing!

AI is changing every day. But right now, the smartest thing you can do is consider using AI on a case-by-case basis. You can view it as a source for a really solid “first draft,” knowing that you’ll probably need a talented, experienced human to go in and interpret, double check, and tweak whatever AI has provided.

Interested in seeing how AI tools + super senior engineers can work together to make your code and site better? Grab a spot on my calendar and let’s chat!