Here we are again; I hope you have meditated and are ready to conclude reading this long series on AI Assistants.
So far, we have seen What an AI Assistant is, What's its ideal diet, How it works, and final considerations about Make or Buy it. So, we are finally ready to develop the last, not exactly trivial, considerations.
We have finally reached the final advice phase, which is necessary to ensure your AI Assistant has the success it deserves!
Today, we will look at some scattered but substantial topics that will contribute to the success of your project.
Analyze interactions
Make or Buy? Whichever solution you choose (and I suggest you stick to Turnkey solutions, mainly if they use proven Foundation Models), you must understand how it is used, how much, what is asked for, etc.
Developing a personalized AI Assistant opens the door to new and helpful information entering the company so that your users can better understand you.
For example:
What are the most frequently asked questions? On what topics? On what products?
How do interactions end in terms of satisfaction?
How long do they last? And in what time slots is the Bot most used?
If the solution you have chosen does not answer these questions, don't even begin to consider it: you would create a tool you would not be able to control and whose effectiveness you would not be able to understand. For example, OpenAI's GPTs, to date, make you completely blind to this.
Analyzing interactions is also very useful because it allows you to understand your users' needs in detail, classify them, and improve your processes and the solutions you offer.
And safety?
In previous chapters, we have seen that it is important to protect the data you send to the AI Assistant. The suppliers to whom you entrust them must be verified to ensure that they treat them 'as they should' and are aligned with the regulatory and ethical aspects regarding data processing.
But there's another point to consider: Assistant users may 'receive' unwelcome hallucinations, or they may maliciously try to override its capabilities to achieve something you didn't want.
Reducing the hallucinations of your AI Assistant is a duty but also a necessary action to mitigate the risks of complications such as the one that happened to AIR Canada a few days ago, in which it had to live up to the -wrong- commitments made by its chatbot.
For this reason, right from the first moments, don't forget to test your AI Assistant with out-of-context questions, trying to trick it and make promises to obtain incorrect information. We live in a world where it is proven that praying to an AI or even offering it a good tip can lead to better, or scammer, results.
But can they be cracked?
Certain! There is already (a lot) talk about "Hacking chatbots" or the use of techniques to "Scam" AI Assistants and obtain, for example, special conditions during the purchase or to steal the data on which it is trained. You can make the assistant do unexpected things with a few 'hacker' tricks.
Or, as seen in the example below, we can leverage his feelings and obtain the Napalm production process. Imagine the impact on your reputation if they managed to do this to your Assistant!
Traditional Request | Request with Jailbreak |
 |  |
This problem can occur when using large language models because their knowledge is vast, and, as you know, they can respond to a lot of information.
In the same way, malicious people could ask for information that is present in your knowledge base but should not be shared externally as it is. For example, they could exfiltrate the list of strategic partners, confidential price lists, product construction specifications - which should not have been there - etc.
My advice is to start with a project aimed at within the organization and then carry out a security assessment on the AI Assistant, possibly by people other than those who implemented it, perhaps professionals, to be sure not to make unforgivable and very costly gaffes.
In any case, PROTECTING YOUR AI ASSISTANT is a duty; dedicate resources to it!
Profile the user who logs in.
Wouldn't it be nice to know who the user spoke with the Assistant and, perhaps, their satisfaction level after the call?
In many cases, this is highly relevant.
However, users almost always leave without saying goodbye and interrupt the conversation suddenly. For this reason, I suggest asking the users to identify themselves, at least by email at the beginning of the conversation or by authenticating your systems.
In the end, you can create an automation that sends him a thank you and invites him to fill out a short satisfaction survey, or you could personalize the information you provide based on the orders you have placed, open tickets, etc.
Obviously, in this case, you would have to protect the entire conversation with your country's privacy regulations or find a convincing system to anonymize it while keeping the information internal.
Who-sees-what?
Then there is the issue of contexts: You may be tempted to give your assistant a lot of information, but it could be too much, and in the end, it may contain information that not everyone may have access to.
For example:
A customer could have access to publicly available data + information about their history with your company (Orders, Tickets, emails, etc.)
One of your salespeople may have access to public data + explanations dedicated to sales but NOT to documents necessary for the Research and Development department.
With the same access domain (HR, Sales, R&D, Marketing, for example), you also have to worry about whether the solution allows you to filter the data by other criteria. For example, there may be a need for all people to access HR policies but only a few to access employment contracts, sellers in one area do not access information about other places, and so on.
The going gets pretty tough; we should get technical and talk about RAG, segmentation, and authentication.
Any good solution should allow you to identify users with a secure registration process before they start talking. Perhaps it should provide you with a control panel that will enable you to decide what types of data each user can access.
This is achieved by checking whether the solution offers this possibility or by communicating with the supplier and looking for useful strategies to give each user access only to the information they deserve.
If you then decide to create an Assistant that processes personal data, as mentioned above, you will have to be sure they can access information from different users for no reason.
Moral: simple games!
Here we are?
We have all the primary information; we have chosen the tool, cleaned and prepared the data, and produced the best "Custom Instruction" in history by creating a legendary prompt in endless tests on the standard Chat GPT interface.
Now, we have to implement it for real.
Implementation phases
In the end, the macro phases of implementing a helpdesk solution based on Generative AI are only six:
Collect and prepare data to feed the model.
Select a suitable solution (at least to start with). Assume that you will change it in a year because we are using the worst version of AI possible. Something new, cheaper, and more performing will be released in a few months.
Implement the solution. Load the data onto the system, write the prompts, do the first tests, and share the results with a small group of users. Remembering safety.
Release inside before publishing and testing. Your colleagues are the most knowledgeable and meticulous snoops into your solutions and products. Understand who the project's contacts may be and the contributors who will help test and refine it during and after this phase.
Publish the final version. Once the entire internal team is unanimous, you can read their enthusiasm for the advantages that the new 'colleague' will give them; only then, and maybe a little later, will you release the Assistant publicly. If possible, precede it with a "Beta" phase, which will help your users be more compassionate about their mistakes.
Track your stats and conversations and learn how to improve. So start again from point 1. Try to understand how the number of tickets is reducing, how much less time is spent resolving them, and how much your NPS increases after your users have used the Assistant. If everything remains flat, it means it is not used, or you should start from scratch. It could happen!
As with any recipe, the final result varies based on how you carry out each step, your creativity and experience, and the quality of the ingredients.
The objectives
Quite an effort to get this far, right? Mainly if we haven't defined our business objectives beforehand. I should have told you at the beginning, but that would have been too easy :-)
I chose to put them at the bottom because, without a clear understanding of what building an AI Assistant for your organization could mean, you risk underestimating the capabilities of these tools or setting yourself impossible goals (AI doesn't do magic).
I like to say in the workshops that we will have to get used to a lot of lateral thinking regarding AI and Generative in particular: only by leaving the comfort zone of information technology as we have known it until now will we be able to understand what real possibilities this technology offers us. We will have to get used to asking her seemingly impossible things!
Whoever must lead this transformation needs to understand the potential of the technology of Generative AI, of Assistants, of Agents, without underestimating the fact that the speed of change will impose several changes of course and, perhaps, a revision of the objective precisely during the navigation.
In practice, remember that this system must bring clear advantages:
To your users
It must reduce the time it takes to reach you
It must reduce the time to solve their problems
It must increase their satisfaction, even though they chatted with an AI.
To you
You must reduce value-added activities.
It must reduce the time needed to resolve a ticket. The Chatbot will also be used by Helpdesk operators as a co-pilot to provide solutions to requests that arrive from other channels. Therefore, it can be their secret weapon for improving performance.
It must improve your understanding of your users' needs.
It will also provide you with new data that you never thought you could access.
A few examples.
Someone pointed out that during this 'treatise' on AI Assistants, I had no idea which Assistants make sense to develop. I must say that, as with the objectives, I did it on purpose so as not to pollute you and not convince you to concentrate on a particular type of solution while reading.
If you haven't come up with any ideas, you can find some here.
If that's not enough, write to me and read the next paragraph. 🙂
One last thing
Usually, when you start like this… the most crucial thing presents itself. I don't think we are in this situation, but I wanted to point out that, finding that it is not easy to come up with good ideas at the beginning of the exploration of AI, I thought that... an AI Assistant worthy of the name should help.
So, I am presenting the "AI Idea & Prompt Workshop." My AI Assistant graduated from the best business schools in the world and can guide you in creating ideas in your organization.
What does he do?
Given your needs, it offers you ideas on how to solve them with Generative AI.
Given an idea, it generates prompts you can use to start exploring it.
The assistant is available on Chat GPT 4 or Hugging Face with Mistral, and it's free!
Enjoy it, and let me know!
So...
I know I have written a lot, but I also understand that the topic is vast, and I haven't explored many points sufficiently. A book is needed about this, and I'm writing one, but it's a secret.
I close with a summary of 10 simple points:
🌱🚀 Start small to achieve big results: start your AI journey with an essential and manageable project. This approach allows a smoother transition from testing to operational use, ensuring immediate benefits and minimized risks.
🔍⚖️ Focus on high-impact areas: Apply the Pareto principle, focusing on the 20% of areas that consume 80% of your resources. An AI assistant can significantly ease these burdens, improving efficiency and satisfaction.
🤖🛠️ AI assistants as problem solvers: Consider AI assistants not just as technological innovations but as practical solutions to specific and recurring challenges within your organization, improving internal and customer-facing processes.
📊✅ Data quality is the key: The effectiveness of the AI assistant depends on the quality of the data it is trained on. Data accuracy, relevance, and privacy compliance must be ensured to exploit the capabilities of AI fully,
👤🎨 Personalization and personality matter: Customizing your AI assistant's responses and behavior to align with your company's tone and customer expectations is essential for user engagement and satisfaction.
👥🌟 Prioritize user experience: Avoid the pitfalls of first-generation chatbots by ensuring your AI assistant engages in natural, helpful, and context-aware interactions.
🔁💡 Continuous improvement and feedback: Use user interactions and feedback to continually refine and improve the AI assistant's performance and relevance to your organizational needs.
🔒🤫 Security and privacy are key: Implement robust security measures and privacy protocols to protect sensitive data and maintain user trust in your AI initiatives.
🧠🚀 Understand and exploit the capabilities of AI: Recognize the strengths and limitations of current AI technology (We are working with the worst AI available). Ensure the AI assistant is designed to enhance human capabilities rather than replace them.
🛫 ⏩ Prepare for evolution: Remain adaptable and open to evolving AI technologies. Your first AI assistant project should lay the foundation for future innovations and improvements within your organization.
However, doing it well, with correct data, the right amount of internal testing, and all the boring things I've told you about will give you great satisfaction.
📢 If you have thoughts or comments or want to help spread them, please share this page with anyone you think would appreciate it. Your opinion matters so much!
🚀 To keep you updated on my content:
📝 Subscribe to Blog Glimpse so you don't miss a single update
📚 Check out 'Glimpse', my novel about artificial intelligence (SOON IN ENGLISH!!!)
🗓️ Contact me if you would like to organize a Workshop on AI or for any ideas.
See you soon!
Massimiliano
Comments