You run a small business and every tech expert in the universe is telling you to “embrace AI” and warning that if you don’t, you’re going to fail.
They’re a little dramatic. But they’re not wrong. So you have two choices. You can wait for your software vendors to develop AI features on top of their applications. Or you can build your own.
Unless you’re a huge corporation with lots of resources, building your own AI system was not a great option even as recently as a year ago. But there’s been a significant change in the tools available.
Let’s say you’re using an accounting system like Epicor, Dynamics, Sage, or QuickBooks Enterprise for your quotes, orders, invoices, inventory and payables management. And let’s say you’re using a CRM software like Zoho, Salesforce, Insightly or HubSpot for your sales and marketing. Can you turn these systems into an AI based system so that you can query all of this data like you can do with ChatGPT?
The answer is yes. But it’s still complicated and expensive for many small businesses. However, if you want to consider this option, then here are 6 not-so-easy steps you’ll need to take to do so.
Step 1: Hire a developer.
You’re not going to do this on your own and all the remaining steps will require a human to bring the apps, tools and data together. A good developer will cost you $150K-$200K annually. The good news is that – thanks to AI – many tech firms (see Meta) and the government (see DOGE) are laying off developers. So there’s talent available.
Step 2: Define Your Deliverables.
How do you want to use your data like ChatGPT? You’re going to want to think of all the prompts you’ll be doing like “give me the status of a customer order” or “based on prior projects how would I price this project?” or “what materials can we substitute to manufacture this product?” or “how much machine time should I consider for this job?” This will keep you focused on results.
Step 3: License an LLM Platform.
Just like you would license a database like SQL Server to write an application, you’ll instead need to license an LLM (Large Language Model) to be your ChatGPT. There are lots of options available and that’s an entirely different article I could write. But you should know the big players: OpenAI, Llama, Anthropic, Google, Microsoft. All of these companies will deliver to you a ready to go LLM model that can be populated with your data and trained.
Step 4: Create Your Data Integration and Index.
Now the really hard part.
You will need to use the APIs (Application Programming Interfaces) provided by your accounting and CRM vendors to move data from these systems into your LLM. Or you can consider out-of-the-box data integration tools like Zapier, Fivetran, Airbyte, Talend or Make. Or your developer can write custom scripts.
LLM’s are different than databases. A database has tables with rows and columns of data. In an LLM data is unstructured. So instead of having an invoice record showing date, customer, amount and description, an LLM would have the same information but in the format of “on June 12, 2024 Acme Corporation was invoiced $354.60 for dry goods.” This information needs to be indexed so it can then be prompted. Tools like LlamaIndex or LangChain can turn this data into unstructured chunks that can be queried.
Key information about this data needs to be stored in a “vector” database using tools like Pinecone, Weaviate or Chroma. Different than an LLM – which is considered to be the “brain” that understands and generates language – a vector database is a filing cabinet, or a type of database that helps computers quickly find things that are similar to each other – like finding a photo that looks like another photo, or a sentence that means something similar to another sentence. It works by turning things like text, images, or sounds into sets of numbers (called vectors) and then comparing those numbers to find the closest matches. LLMs and vector databases work together.
Metadata filters then need to be created to make that happen. You’ll also need to schedule workflows to update data from your accounting and CRM systems into your vector database and LLM so that your queries are using the most recent information available.
I’m simplifying this whole process. There’s more work involved. And there are licensing costs which can be significant depending on your data needs and usage. All of this should be investigated in advance and before you get started.
Now do you see why you need a developer?
Step 5: Build Your User Interface.
ChatGPT has a very simple user interface. But there’s a lot going on behind the scenes so that it can translate the prompts we submit into a format that its LLM understands. You don’t have to build this from scratch. You can use dashboards like Streamlit, Gradio or ChatGPT’s custom interfaces. Or you can develop and embed the interface into Microsoft Teams or Google Gemini. Your custom interface would need security to limit data access to users based on their logins and roles. Authentication, encryption and audit logs also need to be created and maintained.
Step 6: Train, Train, Train.
Do you notice how ChatGPT and other chatbots are becoming better at answering questions? That’s because, since their introduction in 2022, there’s been countless millions of queries made to their LLMs and with each query the system gets smarter. The same goes for your system. Your users will need to test and train and review and help your AI system become smarter too.
I recently showed these steps to a client and they were overwhelmed by its complexity. They’re not wrong: it is complex.
But it is something that can be done with the right developer in place using the right tools, including the ones I’ve mentioned above. For smaller companies that don’t want to wait for their software vendors to catch up, or for those that have data stored in multiple places and need to consolidate to truly build a meaningful generative AI system there is now a (somewhat) affordable path.
All of this will take time. Probably a year. Do you do this? Do you wait? At least now you have options.
Read the full article here
