We continue the series of articles about digital advisors and artificial intelligence. The first article is available here , and the third here.
Content
How do digital advisors work? Of course, there can be a huge number of architectures. But, conditionally, they can be divided into several components.
Data acquisition system
These can be corporate IT systems, IoT devices and process control systems. It can also come from external sources, including the Internet. Or maybe it’s just a user interface in which he answers test questions and formulates a request, such as in a chatbot.
Data storage and processing system, including a model system.
This could be a data warehouse or lake with tools for processing and visualization. If we are talking about a large and serious DSS, then it will almost always be based on Big Data.
Also, any analysis is based on a certain model. It can be of varying degrees of simplification and based on different components. And all analysis will be carried out on the basis of this model.
Data processing methods may be different, but the latest trends are the use of machine learning methods (neural networks, decision trees, support vector machines, k-nearest neighbors) and in particular deep learning (convolutional neural network, recurrent neural network, deep belief network, generative -adversarial network, transfer learning and deep reinforcement learning).
The second direction in the form of rule-based expert systems has a key limitation - lack of flexibility and high labor costs for creating mathematical models. That is, you must first manually describe the mathematical model, create rules, and then, if something changes (data structure, external environment), it will be extremely difficult to adjust such systems.
Data output system
This is the interface in which the user interacts, where he receives analytics and recommendations, and where graphs are visualized.
Below is a visualization of a typical system.
In short:
data is supplied to the system from the outside or taken from the database;
they are analyzed either according to established rules or based on the work of a neural network; it is also possible to access statistical data;
reports or recommendations are being prepared.
If we talk about AI models, then here we are faced with all the limitations of AI, which we discussed in detail here and here . Let's look at them briefly below.
Data dependency and lack of data
To train AI models, you need highly labeled and structured data. And at this level of technology development there is a gap between what large companies want to see and what can be done. Companies simply don't have the amount of labeled data they need to create quality solutions. At the same time, the companies themselves treat their data jealously and do not want to share it with the market and developers. Although, let's be honest, there is little value there.
As a result, we want a lot, but there is either no data, or it is a swamp that is unsuitable for training AI.
Cost and terms of creation, high cost of support
Since corporations want complex solutions right away and on their servers, we are faced with the fact that any AI projects have price tags in the hundreds of millions, and it will be extremely problematic to recoup such a solution.
Let's look at several factors that influence pricing.
Placement in internal data centers
The economics of IT solutions are quite simple, and we will analyze it using a practical example from our experience. If we want to use a SaaS solution , then implementation requires 2 million for equipment and 20 thousand per month for using the IT platform. But if we decide to implement it in our data center with the purchase of licenses, then the implementation will already cost 2 million for equipment and 7 million for a software license. Plus another 20% of 7 million annually for technical support and installation of updates. Simple arithmetic says that you need 9 million to start and another 116.6 thousand per month for technical support and updates. This is how the IT solutions market works.
Development of integrated solutions
Let's look at the example of our digital advisor and why a B2C strategy is relevant for him.
During product research, we were faced with the fact that any company, even without a project culture at all, is not ready to take a ready-made solution based on international methodology. Everyone wants the product to be modified to suit their processes and regulations. As a result, instead of 1000 rubles per month per user, we get the need to form a team of business analysts, developers and methodologists for the customer. As a result, the implementation cycle will be 3-12 months and the budget will be from 5 million rubles.
Well, then, after a multimillion-dollar implementation, you need to ensure the operation of the IT infrastructure. And any AI solutions now require powerful and expensive equipment, which is also demanding on the climatic conditions of the premises where it works (ventilation and other engineering systems are a separate pain point for any data center).
Tendency to produce low-quality content, hallucinations
One of the key problems with generative AI (preparing recommendations based on fact is generative AI) is that it can easily hallucinate, that is, create false or incorrect content.
For example, experts from New York University's Tandon School of Engineering decided to test Microsoft's Copilot AI assistant from a security perspective. Ultimately, they found that about 40% of the time, the code generated by the assistant contained errors or vulnerabilities. A detailed article is available here .
Another example of using Chat GPT was given by a user on Habré . Instead of 10 minutes and a simple task, it turned out to be a quest for 2 hours.
Also, for example, try any chatbot to describe the pAei profile manager according to Adizes . You will be very surprised by what is written. And we are not talking about this, for example, about the case of the pAeI profile.
That is, any AI model needs to be double-checked. And again we come to the point that you need to be an expert in the topic in order to evaluate the correctness of the content and use it. Do managers need such a tool? And here we move on to the next problem.
Responsibility and the black box problem
The key question that any advisors must overcome is who is responsible for the final decisions? After all, any corporate manager wants to relieve himself of responsibility and minimize risks. He needs a tool that guarantees results. They are not satisfied with recommendations that cannot be referred to in case of failure.
In addition, the peculiarity of AI models based on neural networks is that even the creators do not understand its algorithms. Yes, unlike expert systems, AI solutions based on machine and deep learning are capable of self-learning, but a lack of understanding of the logic of work does not inspire confidence in corporate managers. And the developers themselves cannot do anything about it. As a result, the issue of responsibility for recommendations becomes cornerstone.
Vulnerability
Solutions based on neural networks and machine learning are vulnerable like any IT solution. Only in the case of complex AI systems, these vulnerabilities are not so much technological in nature (conditional holes through which one can enter), but logical. That is, we need to look for logical inconsistencies in decisions. And if they want to attack the company, then such a system will be one of the priority targets for attacks. After all, with its help you can not only steal data, but also mislead management into making wrong decisions, or provoke the management system to stop production.
An example is games with images. Researchers can change individual pixels and the neural network sees a helicopter instead of a dog, and a bear instead of a car. The same can happen with recommendation systems - attackers will make adjustments to the data, and instead of increasing capacity, you will decide to sell the business. The example is, of course, exaggerated, but the meaning is clear.
As a result, we get a number of factors:
we don't understand how it works;
we need to cross-check references;
it costs a lot.
As a result, what TOP manager would agree to invest millions in such a solution?
Complexity and high cost of development
Yes, DSS without AI is also expensive to develop. All because it requires the creation of complex mathematical models, or as they are now commonly called - digital twins . And taking into account the fact that such a model must imitate a real object / process with an error of no more than 5%, then complex algorithms are needed. Accordingly, this requires time and a team of professionals, which are very, very expensive.
Equipment requirements
Yes, highly accurate rule-based models are no less demanding on hardware and computing power. A striking example is MES systems , which must process huge amounts of data in real time and prepare recommendations for optimal production utilization.
Lack of flexibility and complexity/cost of customization
We all understand that business processes and equipment parameters cannot be fixed. Processes change, equipment upgrades/breaks down, and overall we live in a VUCA/BANI world. This means that the requirements may change. Moreover, any refinement and adjustment of such a system will be a separate project.
Question of responsibility
For rule-based systems, the issue of responsibility for recommendations is also relevant. Who will be responsible for incorrect recommendations? An expert who doesn't even work for the company? A developer who can’t describe everything? An analyst who can't figure out any possible scenario?
Dependence on developers and lack of training
The key limitation of expert systems is that they work only based on the rules that you put into it. Now the question arises: was the model accurate enough? And based on whose rules does it work? Let's look at the example of our digital advisor , and why we make it based on AI?
There are a huge number of project management methodologies and approaches to their implementation in the world. And each expert has his own vision. Accordingly, if we make a model based on one methodology, then who guarantees that it is correct?
At the same time, we can do it based on neural networks. And let’s say we add a project evaluation module (as in the diagram below). For example, through collecting feedback from stakeholders and data on timing and budget from accounting systems. In this case, the system will learn and over time the system will learn to choose the optimal approach and set of recommendations for implementing the project. No rule-based expert system can do this.
General disadvantages
Limited solutions by data
Any DSS is data-driven. That is, this is the implementation of a data-driven approach.
We examined in detail different strategies for making decisions using data in the article Data-driven or data-informed: why won’t numbers replace humans? Here we recall the key disadvantages of data-driven:
Lack of creativity
Data-Driven gives no room for creativity and different points of view.
Narrow Focus
Data can only answer qualitatively those questions that were initially raised at the design stage.
Elimination of Expertise
Data-driven approach does not take into account expert knowledge and experience
Bias
It is almost impossible to create an absolutely objective set of metrics; in any case, the manager’s view will be reflected in the set of metrics
The difficulty of digitizing people's reactions
Data-driven does not take the human factor into account: emotions and feelings, subjective opinion
Demanding on data quality
Data-driven is demanding on the quality of source data and the development of regulatory and reference information.
And although all modern management science is trying to get away from the human factor, this is still a utopia. At the heart of any successful companies and projects is a combination of a person and a system. And data-driven is not suitable for high-level decisions - making decisions about launching products, their development, developing strategies. And even among technology giants there is a division into data-driven, data-informed and data-inspired approaches:
- Data-driven (data-based solution) for solving everyday/operational problems or in conditions of stability, i.e. for making 80% of all decisions. It is an operational management and planning/control tool.
- Data-informed (data-informed decisions) when creating new products, working with people and planning for the near future. It is a tactical management and planning tool.
- Data-inspired (searching for ideas in data) to create strategies and find new niches. It is a strategic management and planning tool.
Below are examples of using different strategies.
Data-driven | Data-informed | Data-inspired |
AB tests | Development roadmap | Development strategy |
Assessing the stability and performance of IT systems | Prioritize the feature and functionality backlog | Search for new niches |
Degradation of personnel/managers and loss of competencies
An attempt to shift everything to advisers in the long term will lead to the disappearance of management competencies in the company. Everyone will blindly trust advisors. And in the future, when it is necessary to make decisions in conditions of uncertainty and crisis, to redesign the adviser, the company will no longer be able to do this. The company will be vulnerable and dependent on external consultants and developers who will be deeply indifferent to one individual company.
Reduced efficiency and defocus
The development of IT services and digital advisors will increase the amount of information that a manager will need to process. And now the information flow exceeds the ability of our brain and consciousness to process it effectively. This will ultimately lead to lack of focus and can make the decision more difficult, especially considering the factors above.
Employee mistrust and people's fears
With all the possibilities of digital tools, people will still resist their adoption. We discussed the implementation of changes in detail here, but in short:
- people don’t understand how it works, and many don’t understand how to use it;
- people are afraid of losing their jobs, power and influence, managers will especially ask this question: “Why am I here if a robot evaluates everything and makes decisions?”
- digital tools are already many years old, and many solutions have not lived up to expectations, and this creates a culture of mistrust.
Mutually exclusive recommendations
Different digital advisors may make mutually exclusive recommendations. For example, a financial advisor will tell you to cut operating costs and abandon investments, and a digitalization advisor will give recommendations to invest in artificial intelligence and the Internet of things. What decision should you make in the end?
Summary
With all the potential capabilities of DSS or digital advisors, they face complex and fundamental technological and organizational challenges.
Moreover, it is difficult to understand which challenges are more difficult to solve. And, most likely, breakthrough development in this area should not be expected in the segment of large corporations. DSS will arrive there already mature and proven, when they will already become a standard in our society. But we will discuss what to do and where to go in the next article.