Solving your business problems, or any problems, with AI sounds like a great idea until you realize you don’t know how to start.
When you watch “AI-influencers” on YouTube, everything seems easy. They type “write a Snake game in Python,” and ChatGPT writes it in Python. Some of them attempt to run the code; some don’t. Sometimes, the generated code works, so after 30 seconds of playing, they announce that the programmers are done. We don’t need them anymore. Such a claim makes a great clickbait-y title, but it is still far from the truth. The software we create nowadays is far more complex than a Snake game. Also, those AI-generated Snake games tend to fail when you encounter an edge case, such as hitting a wall or your snake’s tail.
Similarly, any naive attempt to deal with business problems using AI will become a disappointment if we expect AI to know what we want, deliver a solution, and account for all edge cases without our intervention. It’s not going to happen. At least, not today.
Solving business problems with AI requires working around AI’s shortcomings. Here is the method I follow:
What is the goal?
The most challenging question comes first. What are you trying to achieve? Today, everyone wants to do something with AI. We don’t need more generic chatbots or web assistants. What’s the goal?
Let’s say you are running a brick-and-mortar business. Let’s assume your business is a tire and wheel shop — a large one. You employ dozens of car mechanics and have 200 customers daily. Many of them post reviews online. You want to figure out what you could do better but don’t have time to read those reviews daily. You would love to get a summary once a month so you can spend the next month fixing the things customers complain about. What’s the goal here?
I think the goal is to get an actionable summary of customer complaints.
You don’t need a list of all the problems they mentioned in the reviews. You won’t have time to deal with all of them at once anyway. Let’s assume you want to see the three most common problems and the three problems with the largest increase in the number of people who complain about them compared to the month before. When you have such a report, you will know the biggest issues and spot the emerging problems, too.
Forget about computers. How would a person do it using a pen and paper?
When we know what we want to achieve, imagine you can’t use electronic devices to solve the problem. Too scary? Ok, I will be generous. You can print anything you want, but after printing those documents, you go to a room with a single desk, pens, empty sheets of paper, and all of the documents you printed earlier. No computers, no phones, no AI. What do you do?
Does it sound silly? Yes, but this is an analogy of what AI will do.
AI has the data you provide in the prompt (imagine those printed documents on your desk) and some space to write the output (pen and empty paper sheets). Even when AI uses tools like access to a database, the Internet, or APIs, it doesn’t access them directly. Conceptually, when AI uses tools, it’s as if we were writing a piece of paper with instructions, handing it over to an assistant, and getting a printout with the results delivered. At best, this is your setup when you use AI. At best! Maybe your AI model doesn’t use any tools and needs to accomplish the task with whatever information we provide in the prompt.
Now, imagine yourself sitting in the room with printouts of reviews. You got a single sheet of paper because the AI’s output window is limited, too. You are bad at math, and you don’t have a calculator. What can you do?
You can do only the smallest portion of the task and give someone the result so they can count them and return them to you later. So, instead of tackling the big job, you read the reviews individually and highlight the parts where someone complains about your business. After highlighting the text, you write what happened on the side of the paper. You write a short description like: “Rude staff,” “Too expensive,” or “Wheel fell off.”
The task you accomplished is like asking AI to classify the problem in a review and extract a quote about the issue. If you store the results in a database, you can refer to them later when you write the monthly report.
How to make the task easier for the person using a pen and paper?
What would make your task easier?
In the case of classifying reviews, you don’t need an assistant who can bring you more documents. It means our AI system doesn’t need access to a database or other external software.
What could help you? You could get someone to get rid of all the positive reviews first. Reading positive reviews about your business feels good, but positive reviews don’t help you achieve the goal. If the reviews come with a score, you can trust people to know how to use a star-based scoring, and we can remove everything with more than three stars. If not, you can quickly read them earlier and throw all positive reviews into a waste bin.
In the IT world, we would achieve the same by retrieving only the data with the correct score from the database or using AI for sentiment analysis before we use AI (with a different prompt) to extract the problems reported by the customers.
We want to remove as much input data as possible. The process will be faster and cost less if we can use a database to filter the data before passing the data to the AI model. If not possible, maybe we can use a simpler (and cheaper!) AI model to preprocess the input.
How does a person create the solution to the problem?
You classified the reviews. What’s next? You still can’t count, so requesting all of them again and counting isn’t a good idea. You are a wordcel, and comparing two numbers confuses you. You will delegate all of such tasks to your assistant.
We won’t use AI to count or compare numeric values. Instead, we can write a database query to count the categories of problems and return the three most common issues or the issues with the largest increase in cases compared to the previous month.
After getting such information from the assistant, we would put the values into the report. We could even use a template because the report would look the same every month. The only variables are the problems we want to report and the date we write the report.
Is this a creative task? No. Anyone can fill out blanks in the report template with the provided values. Therefore, we wouldn’t use AI for this job. Instead, we should write code to place the values in the report. What if the solution were different every time and required some creativity? We would use an AI model. Maybe we would split the solution generation task into subtasks.
The mental model for building an AI system
A Large Language Model, the thing we call AI nowadays, is an equivalent of a forgetful wordcel sitting in a room without access to any electronic devices. You have to give them information printed out on paper. They have a sheet of paper where they can write their response or ask you to do something on their behalf. You can’t trust them with numbers. They earn way more than your other employees, so it’s best to bother them with work only when you can’t do the job without them.
AI works in the same way. AI doesn’t remember anything between prompts. You have to provide all the data in the prompt. AI has limited space for writing the answer. When AI wants to use a tool, it has to communicate with your code and request the tool being used. Later, AI gets the response from the tool as part of the subsequent prompt. AI makes mistakes with numbers and costs more than old-school software.
If you can imagine a part of your business process being done by such a quirky employee, maybe you can use AI. If you can’t, AI is not for you.
Do you need help building AI-powered applications for your business?
You can hire me!