Get an AI Quick Win in Five Weeks

Your company wants to adopt AI, but you’re stuck in discussions and POCs. Here is how you can get a quick win in a few weeks.

Welcome to Episode 36 in Prolego’s Generative AI series. I’ve shared case studies of companies like Vericant and ServiceNow that are getting fast results from AI. Getting these quick wins requires picking the right project, and taking the right steps. Let’s discuss each.

First, pick an opportunity that has this straightforward workflow. Pass documents or chunks of text to an LLM along with instructions for processing it. Store the LLM output into a database or file system. This batch processing workflow simplifies your prompts, data prep, and LLM interaction. Don’t start with complex workflows such as those requiring RAG or agents.

Begin with text summarization or classification tasks. LLMs are good at them and less likely to hallucinate. You can even use small, open source LLMs like Mistral 7B. Finally, start with low-risk data such as policies, contracts, and emails, and avoid sensitive or highly-regulated data.

Second, take the right steps, beginning with data. Compile a representative set of documents, ideally at least 50 examples. If you can ’t use real data, you can starta by generating it. Download our free generating data guide here: https://www.prolego.com/reports/report-discover-ai-opportunities-with-generated-data

Once you have a data set, begin writing and evaluating prompts until you get the results you want. This is both straightforward and fun. You don’t need fancy tools or frameworks. A custom GPT from OpenAI and a spreadsheet are enough. Here’s an example.

Suppose you’re a real estate agent and you want to automatically classify your emails as personal or professional. This is a straightforward workflow because we can pass each email along with a prompt to the LLM for classification.

I begin by generating a bunch of business and personal emails with ChatGPT. I created an example using the raw data from Gmail and told it to make several like it. I then pasted them into a Word document. This is our starting data set.

Next I create a GPT, build this simple prompt, and see what happens. The model seems to be working, but I need a consistent format. So I tell the model to convert it to JSON and update the instructions. Now it produces the output I want.

And you just keep going. Continue improving your data set by adding more emails, making them realistic, and generating the best prompts to get the output you need. Of course you can do this more efficiently with python scripts and evaluation frameworks, but you don’t need to.

Once you get the data and prompts in the right format, give them to your developers and explain the workflow. A team with basic web services skills will have no problem implementing this batch workflow solution.

And that’s how companies are getting a quick win with AI. Of course you might be thinking, “Thanks Kevin, but this only works for simple problems” — and you’re right. Problems with complex data or workflows, and competitive situations will require more investment.

But every company has many opportunities for leveraging this approach. And as models get better, you’ll find even more of them.