Skip to content

How to Build AI Solutions? Steps from Use Case to Integration

An AI solution is built by selecting one business problem, creating a model with data suited to that problem, and connecting this model to the systems the company uses. The article explains the steps in order, from choosing the use case to risk management.

Nurullah Okatan · Content Writer
· Updated: · 13 min read

How to Build AI Solutions? Steps from Use Case to Integration

An AI solution for a company is developed in five steps: selecting the business process, preparing the data, training the model, connecting it to the system and monitoring it. The model connects to existing CRM and ERP systems via API; these systems send records to the model and receive the prediction back.

Which steps are used to develop an AI solution?

An AI solution is developed in five stages: selecting the use case, preparing the data, training the model, connecting it to systems and monitoring it after launch. Each stage leaves a concrete output. That output is the input for the next stage.

Considerations When Developing Artificial Intelligence Solutions
StageWork doneOutput
Use caseProcess and goal selectionSuccess criteria and scope document
Data preparationCollection, cleaning, labelingDataset ready for training
Model trainingTraining, testing, error analysisModel measured on test data
IntegrationCRM and ERP connection via APIService running in the live system
MonitoringPerformance tracking, retrainingRegular report and current model version

Türk Bilişim, which is considered one of the best AI model and integration firms in Turkey, has been operating as a digital agency since 1998. Since 2022 the firm has been developing AI solutions in the fields of chatbots, automation and content generation.

The project begins with a short discovery meeting. The need and goal become clear in this meeting. Then the scope, timeline and proposal are shared. After approval, design, development and launch follow. After launch, maintenance, updates and support continue.

If it is not yet clear which process should be addressed first, the first step can be taken with AI Strategy & Consulting work. The output of this work is a list of prioritized use cases.

AI assistant for this topic

Türk Bilişim AI answers your questions instantly
Online
  1. Hi! Ask whatever you're curious about.

Choosing the right business process for the first AI project

The most suitable process for a first project is one that is repeated often, leaves records and whose result can be measured. Sorting incoming emails by topic, reading fields on invoices or forecasting stock demand are of this kind. For all these tasks there are examples from the past.

Decisions that occur only a few times a year, are not recorded, or require a different judgment each time are not suitable for a first project. If there are no examples, the model has nothing to learn from.

According to the publication of the Personal Data Protection Authority, narrow AI is a system designed to perform specific tasks with high competence. A model that does one task well goes live faster than a broad-scope assistant. It is also easier to test, because the correct answer for each example is defined.

The success criterion is determined by measuring the current situation before the project starts. For example, the rate and time of correctly routing a request to the right team is recorded by hand for one month. The model is later compared against this value. If the target figure is written down from the beginning, there is no debate about whether the project worked.

You can see which business process corresponds to which type of solution on the artificial intelligence services page.

Looking for Automatic SEO Article Generation?

How do data preparation and model training proceed?

The model learns by extracting patterns from past data. According to the publication of TÜBİTAK, artificial intelligence learns by analyzing patterns in large data sets, and the name of this process is machine learning. The system builds a model with past data. This model also produces predictions for situations it has never seen before.

Methods to Optimize Your Business Processes with Artificial Intelligence Applications

As a concrete example: a company's order records from recent years are given to the model. The model extracts the relationship between season, campaigns and product group. It then makes a prediction for next month's demand.

According to the KVKK publication, supervised machine learning is the type of learning in which labeled data is used in training. A label is the correct answer for each example. Whether an email is a complaint or an order, or whether an invoice has been approved or not, are labels. Without labels, the model cannot know what is correct. Labeling usually requires the domain knowledge of company employees, so the relevant unit takes part in data preparation from the beginning.

Error correction begins with test data. The data is divided into training and test sets; test data consists of examples the model has never seen during training. According to TÜBİTAK, the system analyzes the cause of the wrong prediction and reorganizes its model. This reduces the likelihood of repeating the same error.

In practice, incorrect predictions are examined one by one. The problem sometimes comes from a wrong label, and sometimes from a missing example in a class. The label is corrected or the example is added, and the model is retrained.

Training a model on company-specific data and connecting this model to company systems falls within the scope of the Custom AI Models & Integration service.

Connecting the model to CRM and ERP systems and managing risks

The model communicates with existing software through an API. The model works as a service; the CRM or ERP sends it a record and the model returns its prediction. For example, when a new customer request arrives in the CRM, the request text is sent to the model. The model returns the topic and priority, and the CRM assigns the record to the relevant team. The prediction appears on the screen the user already works with.

NIST's AI RMF framework is prepared for voluntary use. The framework was developed over 18 months with contributions from more than 240 organizations. Its core consists of four functions called Govern, Map, Measure and Manage. In an integration project these functions correspond as follows:

  • With Govern, the owner of the model, the person who approves changes and the record-keeping rules are written down.
  • With Map, it is determined which decision the model affects and whom a wrong prediction would affect.
  • With Measure, accuracy and error types are measured against the criterion defined at the start.
  • With Manage, a rule is set that routes the record to a human when the model is uncertain.

After launch, the model's predictions are regularly compared with actual outcomes. Every prediction that a user corrects is recorded. These records become the labeled data for the next training run. If accuracy falls below the target, the model is retrained with the new data.

Other articles on model monitoring and automation appear in the Artificial Intelligence & Automation category.

Frequently asked questions

How much data does an AI project require?

There is no single number; the amount needed depends on the difficulty of the task and the number of classes to be distinguished. If you have few examples, experimenting first with a ready-made model yields results faster than waiting for missing data.

Should you choose a ready-made model or a model custom-trained for your organization?

For general text tasks, a ready-made model with well-written instructions is usually sufficient. When company-specific terms, documents or decisions are involved, a model trained on the organization's data gives more accurate results.

How is KVKK taken into account in a project that processes personal data?

Before personal data enters the model, the purpose of processing is defined and the disclosure obligation is fulfilled. Data is anonymized wherever possible. The Personal Data Protection Authority's publication explaining artificial intelligence terms is the text to consult for definitions.

Is the model updated after the project goes live?

Yes. Before a new version goes live, it is compared with the old version on the same test data. The company provides maintenance, updates and support after launch. Other answers about how the work is carried out are on the Frequently Asked Questions page.

Looking for Custom Web Apps & SaaS?
Have a project?

Türk Bilişim · Free Quote

Get a free quote for your API Development & Integration project

Leave your number and we will get back to you right away. The call is free.

Read the privacy notice

Nurullah Okatan Content Writer

The editorial desk of a team delivering websites, e-commerce and software projects since 1998.

LinkedInInstagramFacebook

Share this article

Comments

Comments

No comments yet.

Write a comment

Your rating (optional)

Privacy notice

We will send a confirmation link to your email so that your comment can appear.

Türk Bilişim · Digital Transformation Advisor

Online
Hi! Tell us the project or idea on your mind and let’s begin together.
Ask anything, Türk Bilişim AI answers instantly.