AI is not just abstract code and algorithms anymore, it has become real "virtual companions" that we can chat with, collaborate with and innovate with. Chatbots and conversational interfaces have become the gateway to a world of extraordinary potential.
Imagine being able to communicate with a virtual assistant who not only understands your language but can also anticipate your needs and respond with intelligence and creativity. This is just the beginning. Gen AI is revolutionizing user interfaces in ways that will radically change our digital experience.
From previous posts, we understand that with Generative AI, the right verb to use, at least for me, is Hire, not adopt or implement.
I think that AI is a new type of resource among those currently available in the company which, although Generative AI does not represent the right solution for all seasons, it is able, with its "conversational" ability, to give us a new approach to digitization.
Asking a Generative AI what it doesn't know how to do - for example making calculations - is like asking a Boeing to transport just a sheet of paper for a few metres. It is therefore appropriate to always ask ourselves which existing technologies it should be combined with.
So we understand, I hope you all agree, that the right framework to give it is that of a multi-degreed, omniscient, arrogant intern, available 24/7, inexpensive (for now), and always hungry for your data, otherwise at risk of hallucinations. And as an intern, it should be monitored and controlled because you can never trust it too much. Especially if you use it "as is".
But what is it made of ?
Of course we are far, or almost, from having a real person in the flesh, very close to having a very realistic virtual representation, as Mark Zuckerberg recently showed us, soon to have a "likeable" avatar but, mainly, we are facing, often, a simple chat-style interface. Chat that should not be confused with those we are used to, very boring and precise, chat whose use between private and professional must be strictly understood. And how we access the AI can make big difference in results and productivity.
Generative AI models are therefore accessed in several main ways:
Model producer chat
APIs
Private or Enterprise Chat
Workflow Automation
Co-pilot app
Native AI App
Let's analyze them in detail!
AI Model Producer's chat
Summarizing it in a very simple diagram, may the more technical folks forgive me as always, there are only two levels between us and the model:
AI Model → Chat User Interface
To empower the capabilities of Generative AI (Through NLU - Natural Language Understanding and NLG Natural Language Generation), the interface that has emerged, given the current average technical limits, is that of a chat in a browser window.
I won't explain how a chat works, but for the first time there is a direct written dialogue with software that is able to 'pretend' to be human and, for the most distracted, easily pass a Turing test (i.e. not make the other side of the chat realize there is artificial intelligence rather than natural intelligence). A way of working that allows you to use any language and take advantage of it to improve your way of communicating and being clear.
Open AI and others, tempted by the possibility of making their AI do more and more things, then began to 'stuff' the interface with commands, options, menus, plugins, just like any other software, while still trying to keep everything as smooth as possible.
Disadvantages? In addition to the effort of working with so much text, to date all interactions between you and the AI are maintained and managed by the producers who will be able to use them to train new models (for example GPT 5.0). This means that if you tell all your secrets associating them with you or your company to the bot, it makes it plausible that in future releases, they will become public domain. #nowyouknow
It's an API
AI Model → API
If someone still doesn't know the concept of API - Application Program Interface beyond asking a GEN AI or looking it up on Wikipedia, just say it's a way for two or more software to communicate with each other. This video explains it quite well.
In our case, Chat GPT can be queried through these APIs by software of different kinds that can make 'calls' to OpenAI APIs without having an interface like a chat. Enterprise chats described above implement this mode.
APIs have different authentication mechanisms that often come down to an API KEY, i.e. an alphanumeric key (e.g. a0gw2hjwrhkajrhaksharharha) that identifies, for OPEN AI, who is the user making the requests, and then charge them. In fact, with APIs you often pay 'on consumption', so based on how many tokens you ask the AI and how many it uses to respond.
The fact that you can have our arrogant intern available via an API means that any internet connected application will be able to request its services.
Furthermore, OpenAI and Anthropic (for Claude) guarantee that in this case it will not use your data to train algorithms. So if you trust the developer of the app you have no one in between and you can 'safely' confess anything to your AI.
However, you must be able to write in some language capable of 'invoking' them. Because as mentioned above they are application dialog modes that use code and often data in JSON format to transfer them between each other.
You can experiment with this world with a very nice tool called GPT4All. It is software that you can download to your PC/MAC and, simply by providing the OPEN AI API KEY, allows you to ask anything of the model, confident that the data will only be on your PC and, for the time needed to transit, on the Internet.
It also gives you the ability to use other models (LLaMa, Falcon and many others) which, if you have great hardware, will run locally on your PC/MAC without ever transiting the network. Let me know if this is clear or if you need more information about it.
Disadvantages: disadvantages? (there are some but... negligible compared to the infinite possibilities they offer and that we will analyze in the rest of the post. First and foremost, what you need to pay attention to is the amount of data you pass through in your prompts and responses: even if you can monitor usage almost in real time there is a high risk of hefty bills at the end of the month!)
It is a "Private" or "Enterprise" chat
AI Model → API → Software Producer → Chat User Interface
Strangely surprised by this need for privacy, more or less all model producers have begun to provide solutions to solve the issue, and I will talk about them in the next points. But many independent companies, or open-source communities, have begun developing chat systems in which the data is 'safe', 'segmented' and private.
Safe means that whoever provides you with the solution takes care of managing the information you provide them, not just chats but also any data needed to train them (documents, web addresses, Q&A you create, etc.) with the compliance criteria required by the various local laws. And that the environment in which they reside is secure from a cybersecurity point of view. But of course, all this will need to be verified.
Segmented, because they offer the ability to decide which of the various company accounts connected to the chat can access which data. To avoid any famous problem of salespeople having access to production costs. Or to ensure that sensitive data, think for example of all the content of an HR office, is available to anyone through the chat.
Private, simply because they are treated in accordance with the GDPR or similar laws.
Quite a few solutions have been born this year: from gpt-Trainer.com to CustomGpt.com to chatfast.io. But on theresanaiforthat.com you can find many others. Just try them out and analyze them.
Disadvantages? You have to trust the contract with the producer of the system which, placed between you and the LLM producer (which you will also have to trust), will know all the company secrets you decide to entrust to it. But if you use it to create a chat with public data (e.g. your catalog, your public KB) it won't have any particular problems.
Workflow Automation
AI Model → API → Workflow Automation
An average adult can type 200 to 350 characters per minute. Those of you who have started using Gen AIs will have realized that the need to 'type faster' on keyboards has increased. Those in IT have been aware of this for decades; and only Google with Alexa, Apple with Siri and finally OpenAI with the Chat Gpt app for smartphones give us the ability to dictate our desires to save a little time.
But there are use cases where between writing the prompt, copying and pasting from one app to another, getting a headache picking up the thread again, using a Gen AI through chat, especially for repetitive tasks (Imagine using it to review resumes coming into the company), it certainly isn't a useful way of working and you risk spending more time asking the question than trying to do the activity yourself.
And this is where Workflow Automation systems come into play.
What are they?
They are tools, mostly web based, that allow you to create repetitive tasks by replacing manual activities with software that performs them and, except for exceptions or specific needs, without writing code.
In essence, they allow you to call the APIs of your preferred LLM in a simple and intuitive interface.
In recent years there has been a great explosion of these solutions: zapier.com, Microsoft Power Automate, n8n.io, Pipedream.com, Dataddo.com, ifttt.com, workato.com just to name the main ones.
How do they work?
Essentially, starting from a Trigger - an Event such as receiving an email, inserting a cell in a Google spreadsheet, a certain time going off, etc. - they allow you to build a series of subsequent steps to be performed by a workflow.
For example
→ Whenever you receive an email from GMail (Trigger)
→ Check if the subject says Application or CV
→ open the attached file and extract the text
→ Send it to Chat GPT with the entire body of the email inside this [prompt] to understand if it meets [these requirements], make a [general assessment] of the person and give a [recommendation score] from 1 to 10
→ if the [recommendation score] is higher than 7, forward the email to [Head of Personnel] adding the word "RECOMMENDED BY AI" + [recommendation score] to the subject and add the [general assessment] to the body of the message
By the way, Zapier recently put an AI inside, so if you copy and paste the text above (i.e. the first four points) it will already prepare the structure which will be similar to this:
At this point all you have to do is, by clicking on the individual points, configure access to your mailbox, check the conditions it proposes, enter your OpenAI API KEY and write a nice resume analysis prompt and... test it all by sending an email.
The first time it will take you about an hour but... you got a real-time CV analyst!
Understood the potential?
Comparison
Here the analysis of disadvantages and advantages would be very long to do.
I had Bing Chat do some research and a comparative table of the main features that I hope will help you understand that:
they are often cheaper than developing an ad-hoc integration;
they are sometimes too expensive because they work on a per-transaction basis. And if you don't know how to plan the numbers, you risk ugly surprises with the API bill at the end of the month;
you can work on them too: no special IT skills required, just a little desire to tinker;
taking them into production in complex environments remains a task best left to professionals.
Co-Pilot Apps
AI Model → API → Co-Pilot
Co-pilot apps are applications that were already born to do things other than AI. But with AI they can benefit from hundreds of new possibilities.
Think of the unjustly maligned old Microsoft Clippy which clumsily tried to guide us in using Office.
For several months now, all software producers, from ERPs to Office tools, from image management tools to software code management systems, have been equipping themselves with co-pilots. That is, they are adding features to their applications to allow the data they are processing to interact with a GEN AI and increase your productivity in a more or less complex way.
For example, to write this post I already used one when Zapier asked me to describe how I wanted to build the tasks in the previous section. Then in recent weeks I got certified to implement Microsoft's Co-pilot (which promises to be one of the biggest software launches in history by going directly into Office 365 to maximize automation of your Excel, Word, PowerPoint and your emails), I use the new version of Adobe Express integrated with Generative AI with some joy, I sometimes let myself be assisted by Notion's assistant to summarize my meetings.
But I also talked about AI at a workshop where a company that deals with knowledge management, ekr.it, is creating its co-pilot to help its customers get the most out of the combination of Generative AI with their systems. (This is not a sponsored mention but I find them very nice and it's the closest case of a co-pilot on an Italian system that came to mind 🙂)
In essence, prepare to be flooded with upgrade proposals for your current favorite software to take advantage of Gen AI in their use.
AI Native Apps
LLM→APP
Obviously there are applications that are born WITH the purpose of providing you with native GEN AI solutions. Apps that are built around AI and allow you to maximize its use in various ways.
Here the list would be too long and as I've already written and as I repeat in every speech, soon you'll need an AI to navigate the thousands of AI-based apps that are emerging every day.
You'll find a nice list on my favorite directory theresanaiforthat.com.
Among these, an app I find particularly useful is rows.com.
It’s a cross between a spreadsheet, a database and workflow app in some ways.
Its formulas, once you have entered your OPENAI API KEY, allow you to call it in the same way you do Excel formulas.
So for example
Copy code=ASK_OPENAI (Prompt, [temperature], [max_tokens], [model])
will allow you to make direct calls to the APIs from within a spreadsheet, referring, to fill in the content, to any cell.
So what...
Again this classification is personal and subject to change over time because the GEN AI world is growing too fast.
But I hope I have given an idea that the Chat is for now the reference model for using an AI when you want to experiment or interact on new topics while paying attention to privacy and confidentiality of the information you send.
APIs are tools that enable different solutions, such as creating complex workflows easily to automate running the prompts you find effective, and are allowing existing applications to be 'augmented' by AI.
Thousands of new applications are emerging every day that make AI increasingly pervasive in new domains and application areas.
It will take a little more time to have a digital avatar connected to the brain. But they're working on it!
As always, I invite you to reflect, comment, and spread ideas by sharing this post with people you think might be interested.
To stay up to date on my content:
Read Glimpse, my novel on AI.
Or contact me here.
See you next time!
Massimiliano Turazzini
P.S. I mentioned many tools and producers today. Obviously I am not affiliated with any of the tools mentioned.
Comments