4 Steps to a Successful Artificial Intelligence Solution Implementation

By December 1, 2021December 17th, 2021Artificial Intelligence, Services
AI project implementation

For most people, artificial intelligence is something straight from the (imminent) future, in which – in extreme anti-utopian visions – intelligent robots take over the world. In fact, AI-based solutions are already a part of our lives. We use many of them every day, often without even realizing it. So is artificial intelligence just another technology – similar to all other IT systems and applications? And what exactly does the process of implementing such a solution look like?

What’s the difference between AI and other IT projects?

The primary differences between projects involving artificial intelligence and other software implementation projects are related to the degree of development of the field and our experience in creating and applying certain solutions, rather than differences in the nature of artificial intelligence itself.

This means that some solutions have been developed and improved for many years. In the IT world, we have had enough time to create processes and methods for them that are reliable, while others are still in the experimental phase – whether they use artificial intelligence or not.

A good example of this is OCR, or Optical Character Recognition, which reads text from images and converts the text into an editable format. OCR (later ICR – Intelligent Character Recognition) uses artificial intelligence mechanisms, yet the implementation of this technology is not very challenging, and probably not many people even think about what mechanism stands behind it. As an online service, OCR has been available since the early 2000s.

However, AI/ML solutions used on a large scale are still a relatively new field, so most such projects are of R&D nature. They can be compared to expeditions of discovery: very risky, full of the unknown, and constantly emerging problems waiting to be solved. Moreover, they are characterized by a very high probability of failure.

It is worth noting that data plays an extremely important role in AI/ML projects. In a typical IT project, relatively simple sets of data are analyzed by a human, who on this basis arranges appropriate algorithms. In an AI/ML project, we deal with problems in which there is so much data and it is so complex that we are not able to tackle it by applying ordinary algorithms. In this case, data also plays a much more important role in the solution-building process itself than it does in other implementations – it is the data that determines how artificial intelligence will use it with the various models we try.

It is this dependence on data and the relatively early stage of the AI/ML field that is critical to the fact that nearly 80% of projects fail. However, those implementations that do make it to the end and prove to be successful bring huge, measurable benefits.

AI methods

So how to balance the desire to use the most modern solutions, using AI/ML, with the obvious fear of engaging in projects that are highly likely to fail, and in addition will generate considerable costs?

The answer to this question is a properly designed process of implementing AI-based solutions, which is designed in a way that will most accurately determine the needs and expectations of customers and minimize the risk of project failure.

In the following part of the article, we present the main stages of the project of implementing an AI-based solution. For better illustration, we will use an example.

Step 1 – getting to know the problem and the customer’s expectations

At the very beginning, we create a business hypothesis (hypothetical business case) which will be verified at the next stage of the process. Based on an interview with the customer, we define the problem, determine the initial assumptions of the project and effects that would be satisfactory for the customer in terms of quality, as well as specific return on investment (ROI) – measurable profits and benefits. We determine with the customer what effect of the AI solution he would consider a success.

It is worth noting that the benefits do not have to end with the savings resulting from the use of an AI solution in one specific process (in our example it would be the process of reading documents).

The calculations should also take into account indirect benefits – resulting from the fact that, for example, people involved in the process, which will now be handled by the AI solution, can devote their time to other tasks. We can also look for profits in other places in the company, in other processes that will be indirectly improved by using artificial intelligence.

AI project

After defining the problem, the expected results of implementing a solution using artificial intelligence, and the budget available to the client, we present the available options of the solution. Not every problem can or should be solved using artificial intelligence.

As we have already mentioned, projects using AI are burdened with risk, so we always try to find alternative solutions that might work better in a given case. Just because a client approaches us with an idea to solve their problem with the help of artificial intelligence does not yet mean that it is the best option. Due to the time-consuming character, costs, and high uncertainty of AI projects, we recommend such implementations only in cases where they are actually likely to benefit the customer much more than other alternatives.

It is worth noting, however, that much depends on the strategy of the customer – one company may not decide to use AI because it will be able to achieve a similar return on investment by more conventional means, while another, despite similar conditions will decide to go forward with the AI/ML implementation as they will consider the prestige that comes with having such solution an important factor in their business strategy.

In step 1, we also take an initial look at the data that the customer has, which would later be used to feed the AI system. Among other things, we check:

  • how much data they have at their disposal,
  • for how long they have been collecting data,
  • at what rate and in what quantities it can be collected.

At the very beginning of the cooperation, in which we de facto only recognize the topic, the customer gains awareness of:

  • what the process of implementing an AI/ML solution looks like,
  • what are the possible outcomes at each stage of the process,
  • what potential risks are associated with each of them.

Automating the process of entering data contained in documents into the system – an example

 

Problem

The client has a large number of documents (legal acts) that need to be entered into the system. The task is to read the appropriate categories of data contained in legal acts, e.g. owner names or property addresses. Artificial intelligence will work based on this data. The scale of the task is so large that manual data entry would be very laborious and time-consuming.

Step 1 – creating the business hypothesis

In order to assess which solution is best for the customer and whether it is worth implementing an AI-based solution in this case, we examined what type and volumes of data we are dealing with:

  • How many pages of documents must be entered into the system per month?
  • How many attributes (data categories) must be completed for each document?
  • What has the document data entry process looked like so far?
  • How much time and work were involved?

We looked at the alternative solutions to the same problem and what would be their cost-effectiveness in comparison to an AI-based solution. For example, would it be more cost-effective to

  • continue with the current practice of entering data from documents into the system manually?
  • outsource data entry to a specialized external company?

After calculating the costs associated with each of these options and reviewing the budget, the client decided to go with the implementation of artificial intelligence.

Step 2 – Verification of the business hypothesis from step 1 and building a business case

Let’s assume that after verifying the initial assumptions, determining the budget and initial ROI, and considering alternative paths to solve the problem, the customer decides to implement an artificial intelligence-based solution.

At this point, the next stage begins (let’s call it Step 2). This is still an early stage of the project, where the business hypothesis and assumptions from step 1 are verified. We are still far from the actual construction of the AI solution.

In stage 2, we create a business case – a business justification for the implemented project that presents the assumptions and profitability of the implementation. We also verify the preliminary ROI calculated in the previous step – creating a business hypothesis. We analyze what the current costs of the project are and based on this information, we determine whether we are able to fit in the assumed budget.

The nature of AI projects makes it difficult to accurately define project costs. Such attempts are however necessary because the budget is one of the factors that greatly influence the decision to embark on an AI implementation. If the client doesn’t have adequate funds, moving on to subsequent stages of the project doesn’t make much sense.

If a limited budget is the only thing standing in the way of full implementation, another option is to use artificial intelligence only to a limited extent. This way, the customer stays within budget and gets a solution to some part of their problem. This way out can bring enough benefits to consider this option.

In this phase, the customer provides us with the data that we use to build the prototype AI model and that will be used by the AI in the production phase. It is at this point that we assess what value this data has in the context of the problem that we are trying to solve. We also define the means we will use to extract value from this data so it can be utilized properly by the AI solution. If at this stage, we are not so sure about these means, we conduct appropriate tests.

To increase the chances of the success of the project, it is very important that the customer takes care of not only providing a large amount of data but especially good quality data. It is a common misconception that having large amounts of data is enough to make AI models work well.

What is much more important is the quality of that data. It is so important that there are separate job positions for people whose main responsibility is to verify the quality of the data, which will be used in machine learning or AI models.

What exactly does good quality data mean? Among other things, it is important that the sample is representative of all the data collected, that there are no gaps in the data, that it comes from the same environment as the data used in the production environment, and that the data is not processed (processed data may carry less information than the original data).

business case AI

In the framework of the created business case, we define the parameters, e.g. qualitative factors concerning the precision and sensitivity of the solution and response time. Furthermore, we specify quantitative assumptions and design the architecture: the technologies used, the environment in which the solution will operate (cloud, customer infrastructure, hybrid environment). Finally, we create the quotation.

As a summary of step 2, the customer receives a document – a kind of report on the basis of which they decide to implement or not to implement an AI-based solution.

This stage involves several days of work and an estimated cost of 60 000 – 100 000 PLN. Although it may seem a significant amount, it is still just a part of the total cost of implementing an AI solution, so it is important to conduct it as reliably as possible so that before making a decision, the customer has the most in-depth knowledge of their problem, solution, risks, benefits, and costs, but also additional benefits resulting indirectly from the implementation of the solution.

This will help avoid misunderstandings at later stages of the project, as well as reduce risks and possible financial losses. As we all know, the earlier the problem is detected, the less it costs to fix it. According to the 1:10:100 rule, detecting problems at the design stage is 10 times cheaper than detecting them at the implementation stage.

At the end of step 2, the client:

  • deepens their understanding of what an AI-enabled project looks like: what are the steps and risks,
  • knows what an AI solution that would answer his specific problem would look like, and what the architecture of the solution would be (based on the built prototype model and the preliminary project of the solution),
  • knows what kind of costs would be involved in the implementation,
  • based on the report containing the results of the analysis of the customer’s data, conclusions, and recommendations for data collection in the future, the customer decides about implementing the solution or withdrawing from the project.

Automating the process of entering data contained in documents into the system – an example

 

Step 2 – process analysis, solution design, pricing

  • Based on the interview with the client, we wrote down the current workflow for the process of entering data from legal acts into the system and the workflow for the same process in the case of applying a solution using artificial intelligence.
  • We carried out data analysis – analyzed the documents in terms of the attributes they contain and the relationships that govern them (what categories of data do the documents contain).
  • We created a solution design – we defined how each category of data will be read from the documents (how we will obtain each attribute).
  • We conducted experiments to reduce the risk of failure – we taught the models and set up processes for selected attributes (categories of data that will be read from the documents) to evaluate the quality of the results. We ran several iterations of this process.
  • We developed the solution architecture.
  • We developed assumptions, including quality assumptions – defining e.g. the accuracy of results, minimum number of document pages processed per month, etc.
  • We created a quote for the work and a proposal ready to present to the client.

Step 3 – implementation

So we have reached the solution implementation stage! The business problem has been properly defined, the customer has decided to go towards an AI solution, the business case has been successfully validated. The customer has decided – let’s implement!

At this stage, we build the actual AI/ML system. To do this, we prepare learning and test data and create AI/ML models which we then test and tune. We also conduct performance and FAT tests.

After passing all the tests, the solution is made available to the customer and integrated with their business processes. Then we launch it in production – in the form in which it will be used by the end-users. All the time we also monitor its work and feed it with newly delivered data. In this phase, we also conduct training for the users of the solution.

Here it is once again worth emphasizing the specificity of systems that utilize artificial intelligence. Unlike standard IT solutions, whose effectiveness does not generally increase over time, training with AI models makes it possible to improve the quality of results and adapt to changing data. Somewhat like a brain, an AI model has the ability to learn from newly emerging information and flexibly adapt to changing conditions.

[More profits with reliable forecasts of energy production from renewables (globema.com)]

At the end of step 3 we have:

  • a ready to use, properly tuned AI/ML solution
  • a working system integrated with the customer’s business processes

Automating the process of entering data contained in documents into the system – an example

 

Step 3 – building and implementation of the solution

  • We prepared the teaching and testing data (we tagged categories and attributes on the full pool of documents intended for testing and teaching the AI models).
  • We performed the process of teaching the models – using the prepared documents. We also tested their quality.
  • We built, installed, deployed, and tested the functional solution.
  • We launched the solution in the production version – for the customer.

Step 4 – system maintenance

After the stage of implementation and launch of the production version of the solution, our work does not end. We make sure that the solution works properly, we provide advice and support, and, if need be, introduce changes.

A short summary of the subsequent steps of an AI/ML-based solution implementation project:

Step 1

Business hypothesis

  • Learning about the customer’s problem
  • Formulating a business hypothesis for verification
  • Calculating costs and ROI
  • Considering alternatives to AI/ML
  • Overview of the data available to the customer
Step 2

Creation & verification of business case

  • Customer data analysis
  • Building and testing a prototype AI solution model based on the customer’s data
  • Summary of the potential contained in the data
  • (Positive) result of business hypothesis verification
  • Initial solution design and offer
  • Initial decision on implementation
Step 3

Deployment

  • Building of the AI/ML solution
    • Gathering test and teaching data
    • Creating AI/ML models
    • Testing & tuning of the models
  • Deployment in the production version
    • Integrations and changes in the business processes
    • Training
    • Deployment on production
Step 4

Maintenance

  • Supervision of the proper functioning of the solution
  • Periodic model training and quality improvement

Possible paths of an AI solution implementation project

Above we have described the typical stages of conducting an AI solution implementation project with a positive scenario, i.e. when the customer decides to implement AI, the business case passes the verification, and at no stage do we encounter contraindications to carrying out the implementation.

However, many projects do not go through all these stages, in fact, only a small percent does!

In the diagram, we present alternative paths of an AI project:

In order not to succumb to the myths about artificial intelligence, it is worth learning about the course of such a project – its stages, the risk associated with each of them, the data we must have at our disposal, as well as the degree of involvement on the part of the client.

It is worth remembering that these are usually long-lasting projects involving large sums of money, and in addition, most of them are research and development projects – “expeditions into the unknown” of sorts, with a very high chance of failure. Luckily the described process of implementing an artificial intelligence project takes this specificity into account and is built in such a way that minimizes the risk of failure and incurring unnecessary (and not insignificant) costs by the customer at each stage. It is also conducted in such a way that enables the customer to get to know the process as thoroughly as possible.

AI system maintenance

Although of high risk, when implemented successfully, AI solutions greatly improve processes and bring enormous benefits. What’s more, through the process of learning, AI systems, somewhat like a brain, have the potential to become smarter and perform better over time.