At Appian Europe 2023, we spoke to Matt Calkins, the company’s CEO and one of its founders. As with most of the market, artificial intelligence is at the centre of the discourse, but Appian defends that it has been involved in this type of system for more than twenty years. Let’s take a look at what the company has to offer (especially in the area of private AI), how it got here and what the next steps are.
You say Appian has been working with AI for 20 years. What has changed in the last year?
AI has changed so much. And I don’t mean just in the last 12 months, but over the last 10 years. In fact, 20 years ago we were doing something that you could call AI. Actually, 30 years ago I was writing AI myself. But it was terrible, because there was no data attached.
But in the last years there has been a tremendous evolution, and a lot of that has to do with the amount of data you can train an AI with. And also the techniques we use to translate that data into useful output.
We will have to trust much more in the translation of data into results, and in the efficiency with which AI does this
In the future we will have to rely a lot more on that: on the translation of data into results. And on the efficiency that AI does this. Some of the Large Language Models out there has been trained on ALL of the valid data its creators could get hold on. So we have, more or less, maximized our ability to improve AI by giving it more information.
If you go back only a few years, you would not have predicted that the breakthrough AI technology would come from a mid-sized non-profit firm. You would have guessed it would be Google. Or perhaps Uber. There are companies that have invested a tremendous amount in AI, but they were not the winners in the initial AI race. How interesting is that?
I think that this is signal to us, that this is not a winner-take-all market, and that many different vendors will surprise us, by being ahead of the curve and the best at a specific dimension of AI. In the end there will probably not exist a dominant AI provider. Instead, it’s a business for the clever, to apply AI to specific problems and to train it specifically for that. And I find that this is a better future.
Is there going to be one AI model that applies universally and is so much better than all the other? I say that there won’t be.
Your way of using AI is with a plain vanilla LLM and providing the data with the questions each time, instead of training it. Is this going to solve the problem of Bias?
The thing is you are using a plain vanilla LLM. If you believe that that LLM is free from Bias, then the answer is “yes” because we are not going to add any. Because it is easy to edit out the bias from the data we are sending.
For example, simply don’t send the race column data. Or the gender column. So, even if the base model has bias, we are not feeding into it. We think that this approach is less biased than the usual one. AI is like a bridge, between the data of the past, and the data of the present. So if the data of the present does not contain bias, there will be none in the results.
AI is going to change many businesses and maybe destroy some of them
Well, Ai will change who the winners and losers are in business. Some businesses will thrive, and some will go out of business. Regarding Appian, our low-code is about speed. It allows you to do things you could have done slowly, but much faster. You can develop an application faster using a flowchart, than writing it in code.
In moments like this, time is of the essence. It is important for executives to prove that they can create value with AI. Everyone is in a hurry to make value out of AI.
So low-code is a very good match, in that we can allow to our customers and partners to win the race. And this begins their journey to create more AI. Creating an initial success could be the ticket to create a long-term successful AI program inside a business.
Why has it introduced its AI functionality this year?
We have been in AI for, maybe, two decades. So we have been in this been for much longer than others. However, what is new here is Large Language Models (LLMs). We were not doing the LLMs two years ago. And that is relatively recent.
Our emphasis today on private Ai is to take the most popular new form of artificial intelligence, which is the generative LLM and find a way to do it without sharing the private information of the client. We didn’t invent the LLMs, but our private AI approach is new and novel. I am not aware of another vendor who is doing what we are doing.
Everyone is in a hurry to get value out of these technologies, so low-code is a very good starting point
It is a very divergent approach, and it is way to solve this intractable problem in which our customers want to use generative AI, but they don’t want to share their information. So it’s hard to find a way to achieve both of those at the same time, and this is our answer to that problem.
What is accuracy and reliability of this private AI
You are correct in that private AI uses less data, and you might assume that, therefore, it is less accurate. But I want to explain why it is more accurate in some cases, and comparably accurate in most cases: If we were to take the high-data approach, and train the AI on our entire database, or on all the data in the company wouldn’t that give you the best results? Surprisingly, the answer is: not necessarily. Because, presumably your data changes.
Our approach to private AI is innovative. I don’t know of any other provider that is doing what we are doing
You get new data, old data gets invalidated and it is very difficult to go to an AI that you trained and say “please forget this data point, and that data group” because the strategy changed, customers changed or regulations changed. It is very difficult to “unteach” an AI.
Our approach doesn’t have that problem, because you are only sending the data with the question. And, as you never teach the AI in the first place, there is nothing to later forget.
We also don’t have the problem of security regulations requiring you to have different permission levels. Imagine having trained an AI on everything you know, but then you have to ask the questions as if you didn’t know that. I explain: in certain fields, like healthcare or government, you are not always allowed to access all the information, all the time.
You can only use part of it, and you have to ask the questions with only the data you know. Our approach can filter the database with only what you are allowed to know, before sending it to the AI. In the other scenario you will get answers with information that shouldn’t be there, as you trained the model with everything.
Another reason why private AI is better is that it is more auditable. If you ask a question and don’t like the results, you’d like to know why the answer was wrong. And now you can see exactly why. Instead, if you have a previously trained AI, it becomes a black box. You don’t understand why AI gave you the answer you got.
So, in most circumstances, private AI is better. There are certain cases when you want pre-train an AI, but most of us work with changing data, where the information is not the same as last year. We, as humans adapt to change all the time.
The things we believed yesterday are no longer valid. Maybe because they have become illegal, or the circumstances have changed, or they are socially unacceptable or simply we find out they are wrong. We discard past information all the time. And AI doesn’t know how to do that.
And because businesses are in a state of constant flux, this represents a problem when using pre-trained AI systems.
What kind of impact will AI have on the employment?
People are concerned about this. They are worried that AI will increase unemployment, by replacing people, and put them out of work. I believe that, instead, AI will increase the number of jobs.
We have a few years to digest what to do with AI and find its place in the workplace
That we will need more human labour in the AI world, than we had beforehand. That there will be more jobs using AI, writing AI, interacting with AI and the other jobs will simply become more powerful, because of the fact that they are being accompanied by AI.
I think we are going to be able to create far more. But our appetite for achievement will not stop. Every time you have given humanity a new tool, whether it is a wrench, a bicycle, a computer or a calculator…every tool you give us, we just decide we want to do more with it. Every time there is a new invention, we decide that we want to do more. And we grow the demand, so I think we will have room for full employment with AI.
Anyway, this past year has been exhilarating but I think we are going to slow down now. We are going to apply AI to specific jobs and practical things. We have a few years of digesting what to do with AI and find its place in the workplace.
Some technologies don’t have much substance, but I think this one does. There is real value in AI.