How we train our A.I. Assistants to automate tasks we hate doing.
Today we breakdown how to deploy an A.I. Assistant to automate tasks the cost you money and time in your business (and quite frankly nobody likes doing).
Let's be honest. 80% of work we do on a daily basis is just to keep the wheel turning, and 20% is actual value add to the business. Busy work is the mind killer.
If you’re in customer support you are spending way too much time dealing with frequently asked questions and not enough time talking to high value customers about what they want to see in your product. If you are in sales, you are probably spending your time entering data into a CRM or qualifying leads that go nowhere instead of spending quality time talking to prospective customers outlining a proposed solution for their challenges.
Every job has 5 minute tasks that eat up our day and hold us back from investing in the more important tasks at hand. If you are a small business or entrepreneur you may even find yourself unable to effectively do anything meaningful because you are juggling all of the above without much help.
Five Steps to Automation
If you are looking to leverage A.I. to help your business here are the 5 steps you can take to start automating those 5 minute tasks.
Step 1: Identify the Tasks you want to Automate
The first step is to identify the basic tasks you want to automate. In most cases there are two types of tasks:
Conversational - where the A.I. has a conversation with someone and asks and answers questions and provides results to you.
Creative - where A.I. takes instructions (or content) from you and creates new content (e.g. blog, summary of a document, presentation).
The tasks you identify need to be simplified so they can be easily given to an A.I. Assistant to execute. Here are some examples:
Qualifying Leads. Asking a series of predetermined questions and answering common questions asked by prospects.
Writing a blog. Taking a basic instruction or idea for a blog and then writing a full article.
Presentation or Meeting Summary. Taking the transcript of a meeting and identifying the key points and next steps to share with attendees post meeting.
All of these are singular tracks and easily implemented launching an A.I. Assistant for each task using OpenAI Assistants and a launch platform like RAIA.
Step 2: Training your Assistant
Once you have identified the task, you can start gathering any materials (websites, documents, transcripts, examples, etc) that can be curated and used for training your A.I. Assistant. The best way to think about training A.I., is to think of it as a new employee who needs to be trained or setup to learn on the job.
In order for your training to be effective you will need to convert your documents into compatible formats for your A.I. Assistant. Using a platform like RAIAbot.com can help with this process. There are four ways to train (or prompt) your A.I. Assistant in OpenAI.
Instructions in the Assistant. This is limited in size but the most important training when it comes to identifying the purpose and goals of the assistant. This is more of a framework on how the A.I. should operate.
Prompting via Conversation. For more dynamic data you may want to include in your training you can embed information into the prompt via API when your Assistant is talking with an end user. On RAIA we use this to power our Memory functionality so the A.I. can remember past conversations with a specific user.
Vector Store. The vector store allows you to store a large amount of data for retrieval when the assistant needs to access data to answer questions. For example a large database of frequently asked questions or examples of support tickets would be uploaded into the vector store.
Fine Tuning the Model. For much larger sets of data that may need to be embedded into the core model you can also use fine tuning. In most cases this may be overkill.
Training is more of an art than a science which is also why we provide in our RAIAbot.com platform easy ways for users to provide “human feedback” via our app and also launch real-time training for the A.I. to be trained by experts in your business as well.
Step 3: Testing your A.I.
Once you have your A.I. trained - it is time to test. Since A.I. has a “mind of its own” sometimes testing is important to continuously do even after going live. The best way to start testing is to simply converse with the A.I. as if you were a user and ask common questions you feel most people would ask and see how it responds. We recommend writing down a test script to be used on an ongoing basis so you can easily run through tests in the future. If you are not sure what to ask, you can also ask the A.I. to provide you a list of questions you should ask (yes, I realize it sounds counter intuitive but it reveals to you what it thinks it knows as well from a different perspective). We built testing tools with RAIAbot.com to help businesses track testing and add comments via human feedback training to make it easier to improve their assistants.
Step 4: Integrate your A.I.
If you are looking to have the conversations or content your A.I. produces go into your 3rd party apps (CRM, Blog, etc) you will need to integrate via API. In many cases this is relatively easy to do (especially with a platform like RAIA). The most comment integration is sending the results of the A.I. conversation or content creation to an app. This would happen once the A.I. has completed the work (usually via a Webhook). For example, the A.I. has a conversation with a lead (which could last 5 minutes) and when complete it sends the entire conversation to the CRM system.
Step 5: Go Live and Monitor
Once you go live with your A.I. Assistant you should continue to monitor the conversations and creative it produces. There are always edge cases that may not have been caught in training. The good news is that it is easy to add on additional training instructions to your A.I. assistant real-time.
Hopefully this gives you a basic guidelines on how to start using A.I. in your business leveraging OpenAI and a launch platform like RAIAbot.com.