Anti-patterns that cause problems for AI implementation

Artificial intelligence (AI) implementation is a complex process, too often masked by a reputation marked by hype and panacea promises. While AI democratization aims to make AI accessible to a broader audience, many people still lack an understanding of AI, thus limiting its potential benefits. Consequently, AI implementation has become complicated by the rise of anti-patterns, common yet ineffective solutions that arise during the implementation of AI projects.

To avoid anti-patterns and thus to implement AI models responsibly and effectively, businesses should be aware of them.

How does artificial intelligence work?

To start with, the artificial intelligence (AI) we see in sci-fi movies, machines solving creative problems requiring common sense, is called “wide AI.” If we picture AI solving problems even better than humans — this is “super AI.” We are far from wide or super AI; in fact, some even argue it can not be done. At present, the AI we have is “narrow AI.” It can solve one specific problem, but it lacks common sense requiring effort to adapt it to solve a similar but not quite the same task. Despite these limitations, narrow AI is extremely powerful, and, in the right hands and in the right place, it can make people believe that they are dealing with wide or even super AI.

AI models evolve via machine learning (ML). This is complicated, but let’s generalize and make an analogy with a production process. ML takes input (e.g., information from your CRM, databases, spreadsheets) and produces an output (e.g., finding fraudsters, handling claims, classifying what the customer asked). For the machine, producing the required output from input requires data to learn from. For most businesses, that’s it, and, from a technical stance, there is no need to delve deeper into how this happens.

AI is still extremely challenging, but a lot depends on how you define the initial problem to be solved and, with expert advice, determine whether it is even an AI-compatible issue. Unfortunately, many decision-makers go the wrong way, applying anti-patterns to AI implementation.

What are the anti-patterns of AI implementation?

According to the authors of Design Patterns, anti-patterns are commonly used processes or structures which, despite initially appearing to be an appropriate and effective response to a problem, result in more bad consequences than good ones.

1. We have tons of data. Let’s make use of it with AI!

Anti-pattern: Often enough, companies develop an action plan like this:

1. Get data;
2. Apply machine learning;
3. ????;
4. Profit.

However, such an AI strategy rarely works well. What is more likely to happen is that many smart people will form hypotheses and test them against their data. As a result, they may find few patterns relevant to the company’s business issues. However, such Data-First machine learning projects are far from optimal, failing because the questions are typically wrong or even irrelevant. Why? This approach leaves the question-asking job to data scientists or data analysts whose job doesn’t involve familiarity with how the business is run and where the main pain points are.

Solution: Don’t start from your data. It’s the problem that matters the most. The craze for artificial intelligence is creating a sense of FOMO (fear of missing out) among organizations large and small: “What if we all fail to use artificial intelligence and are left behind by the competition?” That’s the question an increasing number of companies are asking. Certainly, you should be asking that question because you can be left behind, but you must approach your AI projects from the right angle. People can find almost anything from data, but without clear business guidance, rarely are such results business-enriching. The business side needs to lead, offering clear goals. Don’t leave your AI team without business guidance.

There is no single way to determine whether an issue is a good fit for AI implementation.
However, there are some quick tips to recognize the signs that the challenge could be solved with AI models. The first step is to choose the problem-first approach, and read our AI life cycle guide and execution plan for an AI project.

2. AI solves “something” “somehow”

Anti-pattern: Often, businesses just want an AI model to improve customer retention or increase customer satisfaction. Such tasks are too vague for AI models.

Solution: Vague business goals should be turned into more specific ones as soon as possible, so you will need some exploration before AI implementation. For example, the problems solvable with AI technologies could be:
What triggers the churn of our customers and when?
Can we visually detect faulty products before they are sent to customers?

With the last example, going from a vague goal of “improve customer satisfaction” to a specific “reduce the number of faulty products shipped to customers” enables the problem to be solved via data collected and, thus, AI-relevant.
Please, check if your improvement ideas meet the following criteria:

• Processes that are mostly routine and standard
  Routine and standard processes are the bread and butter of artificial intelligence, even if those processes are complex. However, multiple edge cases can make application difficult.
• Processes where it is possible to create a learning loop
  The learning loop is an iterative process whereby the model can receive feedback on the differences between the predicted and real output, and, through this learning loop, refine and improve future predictions.
• Processes where interactions with the physical world are nonexistent or simple
  It is easier for AI systems to work with digital data, such as texts, pictures, or logs. Even with current technological advancements, tasks involving sensors are complicated as they inherently have noise and fuzziness. Moreover, interactions with the physical world are just plain difficult; Boston Dynamics’ videos are impressive now, but it took them years, if not decades, to get to the toddler-level performance. However, if the returns of a proof-of-concept model are good enough, implementing an AI system is worth a try. That’s why many companies are developing self-driving cars, despite the myriad challenges and immense difficulty of such tasks.

• Processes that are performed often and require a lot of people
  If a task is performed often, it should provide much data. Moreover, a task requiring many people will realize a significant gain after full or even partial AI automation.
• Processes that could be done better if there was more time for analysis before decision-making
  Some tasks, such as an audit, imply a vast workload. To manage, humans analyze the most critical information before making a decision, which mostly works fine. However, AI models can help people make higher-quality decisions by quickly examining all the data and highlighting what requires more attention.
• Processes that could be done better if we could see all the connections between the data
  An improved understanding of data connections would allow for more in-depth data analysis, providing insights into patterns, trends, and relationships that may not be apparent with siloed data. Yes, it is the other side of the coin of the previous point.

3. Let’s fix the data first and then start the AI project

Anti-pattern: Fixing the data is a common point in AI implementation checklists. Those checklists to assess your data quality aren’t bad, but they are full of “it depends” even without explicitly saying so. Still, not knowing how to interpret the checklist or what is truly important there can lead to a lot of “empty” work.

This happens because defining the specific detailed requirements for the data before profoundly going into the problem with an AI team is almost never possible. In addition, having good quality data for artificial intelligence (i.e., data you can use to train an AI) and for other purposes (e.g., simple analytics to look at trends) can be a very different game. Moreover, the requirements for data used for artificial intelligence can also change over time.

Solution: Fixing the situation with data is part of an artificial intelligence project; there will always be something that needs improvement. However, doing it beforehand is mostly a waste of time and money. Instead, before AI implementation, make sure access to extant data exists, but don’t go further than this on the first try.

4. Anyone can do it these days: I’ll give it to any interested person

Anti-pattern: This anti-pattern is a result of an effect of blogs, courses, and calls to democratize AI that makes it look too easy. Of course, a few lessons in AI could help, but the person wouldn’t know what they are doing exactly. To make it more clear: imagine you need surgery. Who would you prefer: an experienced surgeon or a physician who did a two-month online course?

Solution: While it is true that artificial intelligence has become more accessible in recent years, developing and implementing AI solutions effectively requires a combination of technical skills and domain knowledge. Simply giving AI tools to someone without proper training and experience is unlikely to produce optimal results. Yes, everyone has to start somewhere, and open-source ML libraries are part of the equation, but AI implementation is much more than that. That’s why it is studied for years in university, and practical experience remains invaluable.

AutoML or an AI platform doesn’t help, either, as such solutions still assume that the rest of the work (such as data collection and preparation, goals definition, etc., which is actually most of the work) is done correctly.

A good AI team should ideally follow three key principles when developing turnkey AI solutions from idea to production:

1. Delivery ownership. The machine learning team should fully analyze and understand the business objectives regarding development. For example, the MindTitan team always verifies that the clients get what they need, not just what they ask for. This means that once completed, the AI systems solve business problems, can improve business metrics, and fit into existing processes.
2. Oriented towards actual development at a fast pace. A team of AI and ML experts is not just a consultancy or a lab that does research. They know that the business needs to get AI projects done as efficiently as possible and do their best to achieve this. Also, often you just do not know if it will work well enough before you try.
3. With you from start to finish. Good AI experts and ML teams understand AI development concerns, not only the AI model but also everything around it: verifying the idea, working with the data, launching an application, and improving infrastructure. Accordingly, the team should be present throughout the entire process, from idea validation to pilot project to production.

5. Let’s hire a data scientist

Anti-pattern: Hiring a data scientist to add to an existing development team can be a cost-saving decision. However, such a move might result in even more time and money because of increased delivery time and poor quality outcomes. Also, the hired data scientist often doesn’t do any data science; instead, they just spend their time explaining to people what could theoretically be done with AI. This might not be completely wasted work, but it is definitely an inefficient use of this resource and could be done more efficiently with an AI partner.

Look at the picture below: one person cannot have all the necessary skills. Instead, results come from AI and ML teams, comprising not only data scientists but also machine learning engineers, project managers, data engineers, analysts, developers, and others.

Solution: To build a qualified team, outsource AI development.

6. Let’s hire a whole team if one is not enough

Anti-pattern: While an in-house AI team can theoretically provide easier control over data and processes as well as access to expertise, these are usually more expensive in the long run. Moreover, they limit expertise, lack flexibility, experience a lot of downtimes, and realize results more slowly.

Solution: There is no long-term need unless you are making a complex AI project, such as a self-driving car.
Thus, the solution can be machine learning outsourcing.

Team building is complex and time-consuming, and not every required role is a full-time job.

For example, AI analysts are required from the very beginning of a project for mapping: to determine what needs to be done from the business perspective and the technical demands of the project. Of course, while they could still be needed during application development, their involvement is lower after the initial mapping.
At the same time, developers are only required at the app development phase after the first usable model version is ready for testing.

AI projects, due to their nature, are cyclical. There can be moments when there are so many things to do that it might seem like they will never end. At other moments, team members may feel that they are in the eye of the storm, with a calm that they didn’t expect. In AI, such pauses may happen for various reasons: more data collection is needed, integration work needs to be done, the performance is good enough, data drift does not happen overnight, etc. However, the load usually decreases with every iteration, as shown in the graph below, unless some systematic change happens that requires more work. If the load seemingly doesn’t go down, the team might likely be doing things that don’t have an ROI.

Additionally, when the total team’s workload decreases after the artificial intelligence project is launched, what happens to the team when the AI solution is “done”? Of course, people can always come to sit at work and do something, but are the salary outlays in such a case justified?

8. After all, AI development is also just software development

Anti-pattern: If you have this approach, your AI project will likely fail since there are many significant differences compared to regular software development. For example, it is vital to have expert knowledge of when it makes sense to stop iterating the ML process: it shouldn’t stop too early nor go on for too long. This knowledge comes with experience only.

Moreover, unlike IT, which is based on cause-and-effect logic (if-then), AI uses a trial-and-error approach, whereby data and algorithms determine the outcome. IT is usually predictable, whereas AI is more akin to research and development. Neglecting these differences can lead to unrealistic expectations and disappointment in AI projects, which can result in project cancellations and budget cuts.

Solution: AI and software development share some similarities but are also distinct.

Like software development, AI development requires the creation of algorithms, testing, and debugging code, and using programming languages and tools. However, given that it also requires a deep understanding of ML concepts, mathematical algorithms, and statistical methods, an AI solution requires significant data pre-processing, cleaning, and feature engineering. Moreover, AI development is subject to certain ethical and legal considerations, such as data privacy, fairness, and explainability. These considerations are not typically relevant to traditional software development.

So, the best solution will be to outsource a machine learning team or find and hire an in-house expert (if you have a very difficult and long-lasting project, such as an autonomous vehicle).

10. If we are already developing the AI, let’s also develop the application on our (fixed!) budget

Anti-pattern: Although fixed budgets are more common in the public sector, they can happen in every industry. To clarify, let us remind you of the old story about the seller who wanted to make as many hats from one lamb skin as possible. One master told him that only one is possible – and made a good and beautiful hat. Another master said that as many hats as he wants are possible – and made twelve hats – but each the size of an apple.

With a fixed budget, the number of AI development cycles decreases because money, time, and attention are limited, potentially leading to a less effective AI solution performance or ending development too early, although it just needed more iterations. Also, rushed development may lead to an unusable application because the process of AI development reveals what needs to be altered bit by bit, which otherwise might make the developed application partially or even fully useless.

Solution: The focus should shift from the AI model to the application only when we know that the AI can work (well enough) and we have defined how the relevant business process needs to work for the AI model to operate.

11. We’ll figure out how to use the AI once we know if it works

Anti-pattern: This is more common in larger organizations, but it can happen anywhere. Although AI was proven to work, the process is kept the same because a basic check was not done before or the ways of fitting AI into the extant process were not determined during AI development.

Solution: Business leaders should at least make a basic reality check to understand how complicated it is to change a current process. While a strong data science team will bring the AI model to life from idea to execution, it is up to the business stakeholders to own the problem and be willing to implement the solution in real-life business processes. Yes, there can be new process requirements uncovered during AI development. However, these new requirements should not be a reason to forego a basic check. For example, if the AI is supposed to give recommendations to people, business leaders need to determine how the team should design the process so that people actually use it, decide on the exact benefit of the AI to people who use it, and ask what would happen if they do not use those recommendations?

Instead of conclusion: the best tips on successful AI implementation

These tips can help ensure the success of AI implementation and lead to better outcomes for your organization. By setting clear goals, focusing on specific business issues, ensuring data quality, hiring the right team, investing in infrastructure, planning for continuous improvement, and communicating effectively with stakeholders, you can maximize the business value of AI.

So, how to be successful?

1. Have a business problem/automation opportunity wherein you know that, although there is no easy solution, you intuitively know that the solution lies in the data;
2. Consult with AI experts about your business problem/automation opportunity to understand the feasibility;
3. Outsource the development of artificial intelligence unless you are dealing with a project with the complexity of a self-driving car; and
4. Ensure continuous and close collaboration between the AI team and the business side.