IBM identifies the top AI adoption challenges as concerns over data bias (45%), lack of proprietary data (42%), insufficient generative AI expertise (42%), unclear business value (42%), and data privacy risks (40%).1 These obstacles can hinder AI implementation, slow innovation, and reduce the return on investment for organizations adopting AI technologies.
To overcome these challenges, explore the top 10 steps to developing AI systems.
10 steps to developing AI systems
1. Defining objectives and requirements
This phase begins the AI action plan and sets the foundation for the entire AI development process.
1.1. Determine the scope
Begin by identifying the exact problem you want AI to solve. For example, you might want to automate how your team handles customer questions or analyze text data to uncover market trends. Keep the goal focused and practical. As you shape the scope, consider principles such as fairness, openness, and risk management to ensure the system is used responsibly.
1.2. Resource allocation
Outline what the project needs to succeed. This may include skilled people, computing power, development tools, and clean, well-organized data.
Plan not only for building the system but also for maintaining it later on. Decide early whether you’ll rely on your internal team, work with external experts, or use both.
2. Gathering data
Data is the foundation of every AI project. A model’s performance depends less on how advanced the algorithm is and more on the quality of the data it learns from.
2.1. Understanding data types
AI systems work with two broad kinds of data:
- Structured data: Information stored in organized formats such as spreadsheets or databases.
- Unstructured data: Content like text, images, audio, or videos that doesn’t fit into fixed tables.
2.2. Data sources
Your data can come from many places, such as internal records, public repositories, web-scraped information, or third-party providers. Crowdsourcing and synthetic data can also help, especially in areas where privacy limits data collection, such as healthcare or finance.
2.3. Protecting data privacy
Use privacy-focused approaches that let your models learn without exposing sensitive details. Techniques such as federated learning and differential privacy enable training AI systems on distributed datasets while preserving user privacy.
3. Data preparation and manipulation
Once data is collected, it needs to be cleaned and shaped before training the model. This step ensures the AI system learns from reliable, meaningful information.
3.1. Ensuring data quality
Accurate input leads to better results. Review your data to spot and fix errors, fill in missing details, and ensure all formats meet your system’s requirements. Clean data helps prevent errors during model training.
3.2. Turning raw data into a proper form
Use basic statistical methods or manual adjustments to turn raw information into variables your model can understand. Some tools can automate this process, but it’s still essential to confirm that the transformations make sense for your goal.
3.3. Feature selection
Focus on the variables that truly affect outcomes. Removing unnecessary data points reduces noise and improves the model’s performance consistency.
3.4. Data annotation
Use large language models (LLMs)-assisted tools and human-in-the-loop systems to annotate unstructured data. This step is crucial for supervised learning tasks, such as computer vision or natural language processing.
4. Model selection and development
Choosing the right model architecture is central to developing AI effectively.
4.1. Choosing the right algorithms
Select your algorithm based on the task (classification, clustering, regression), available training data, and hardware constraints. Deep learning models remain effective for unstructured data, but transformers and foundation models now dominate tasks in vision and text.
Popular models include:
- Vision Transformers (ViTs) for image tasks
- BERT/GPT for language
- SAM for segmentation
Read large language model training to learn more.
4.2. Using pre-trained models
Pre-trained models, such as ResNet, CLIP, and GPT, can reduce the time required to create AI. Check out the large vision models to learn more about.
Fine-tune them with your training data for domain-specific performance. Use transfer learning or low-rank adaptation (LoRA) for resource efficiency.
4.3. Programming languages and tools
Python and R remain dominant programming languages for data science. Tools such as TensorFlow, PyTorch, and JAX support advanced model training.
Use LangChain, LlamaIndex, and other orchestration frameworks for building LLM-based applications.
5. Training the model
At this stage, the model starts learning from the data and develops the ability to perform its intended task.
5.1. The training process
Begin by feeding the prepared data into your model. As it processes the information, the system learns to recognize patterns and relationships that guide its predictions.
For larger models, methods such as transfer learning or low-rank adaptation (LoRA) can help reduce training time and computational costs without sacrificing performance.
5.2. Continuous learning
AI systems can evolve through online learning pipelines. Use retrieval-augmented generation (RAG) to inject real-time information into model responses. This ensures the AI stays current and effective in a real-world environment.
6. Validation and testing
This phase evaluates how well your AI model performs on unseen data.
6.1. Assessing model performance
Split your data into training and validation sets to see how consistently the model performs. Evaluate metrics such as accuracy, recall, and F1 score.
Go beyond basic testing by reviewing fairness and bias, running explainability tools such as SHAP or LIME, and performing stress tests to identify weak points. These steps help ensure the model works reliably across different scenarios.
6.2. Fine-tuning
If the results are below expectations, improve the model by using additional training data or by switching to alternative algorithms. For efficiency, consider applying parameter-efficient fine-tuning (PEFT) techniques, such as LoRA or QLoRA.
7. Deployment and maintenance
Deployment integrates the AI model into existing systems, while maintenance ensures long-term viability.
7.1. Deploying the AI model
Deploy AI using tools like Google Cloud Platform, Microsoft Azure Machine Learning, or Amazon SageMaker.
Consider serverless AI and edge AI for low-latency tasks and scalable infrastructure. Model-as-a-Service (MaaS) options help deploy without managing infrastructure.
7.2. Long-term maintenance
Utilize tools such as Arize AI, Fiddler, or WhyLabs for monitoring.
Implement drift detection and set up automated retraining. Ethical considerations, transparency logs, and user feedback loops help limit AI misuse.
8. Responsible AI and governance
Incorporate governance frameworks to guide the AI development process responsibly.
- Fairness and transparency: Adopt FATE principles (Fairness, Accountability, Transparency, Ethics).
- Compliance: Align with AI regulations, such as the EU AI Act and U.S. Executive Orders.
- Documentation: Use model cards and data sheets to document model behavior and data sources for reproducibility.
9. LLM integration and orchestration
Large Language Models (LLMs) now power many AI applications. Tools like LangChain and Semantic Kernel help create AI agents that can interact with external tools or documents.
- Use agents like Auto-GPT for task automation.
- Adopt orchestration frameworks for scalable large language model (LLM) pipelines.
10. Model evaluation beyond accuracy
Performance metrics are no longer enough. Expand evaluation to include:
- Trustworthiness: Bias detection.
- Explainability: For high-stakes use cases.
- User experience: Essential for AI copilots and chatbots.
Deploying AI systems across various industries
Healthcare
In October 2025, UPMC announced plans to deploy Abridge’s AI-powered documentation tool throughout its hospitals and clinics. The decision follows a successful pilot in Pittsburgh and will extend to more than 12,000 clinicians across 40 hospitals and 800 outpatient locations in the U.S. and abroad by 2026.
The system uses generative AI to turn spoken dialogue into clinical documentation and is now in use at over 200 health systems nationwide. At UPMC, it supports 44 specialties, including oncology, cardiology, and pediatrics, and has helped reduce administrative tasks while fitting easily into existing workflows.2
Cancer Center.AI developed a platform on Microsoft Azure that enables physicians to digitize pathology scans and utilize AI models for analysis. This system has improved diagnostic accuracy and reduced errors in initial pilot studies.3
Robotics and security
General Motors (GM) is expanding the use of artificial intelligence and robotics in its assembly plants. The company’s new initiative focuses on collaborative robots, or “cobots,” that can safely work beside human employees. The project is part of GM’s larger modernization plan, introduced at the GM Forward event in New York.
At the heart of the effort is the Autonomous Robotics Center (ARC) in Warren, Michigan, supported by a partner lab in Mountain View, California. Together, more than 100 engineers and researchers are developing robots that can learn from production data and adapt to real-world factory conditions.
Key developments include:
- AI-based learning: Robots are trained on years of production data to improve accuracy, predict maintenance needs, and increase efficiency.
- Collaborative robots: GM’s cobots work alongside humans on repetitive or physically demanding tasks, improving safety and flexibility.
- Predictive maintenance and quality control: Robots equipped with sensors and cameras can detect defects in welding and painting before vehicles leave the line.4
Boston Dynamics’ Spot robot dogs have been deployed in industrial settings by companies such as GSK and AB InBev for tasks including safety inspections and efficiency enhancements.
Figure 2: Boston Dynamics’ Spot robot dog example.5
Maritime operations
The Port of Corpus Christi has implemented an AI-powered digital twin system, OPTICS, to enhance real-time tracking and safety. The system uses machine learning to predict ship positions and supports emergency response training with generative AI.
Figure 3: An example of OPTICS, showing ship information.6
Workplace productivity
Access Holdings Plc has adopted Microsoft 365 Copilot, integrating generative AI into its daily tools, resulting in significant time savings on tasks such as coding and presentation preparation.
Agriculture
KissanAI released Dhenu 1.0, the world’s first agriculture-specific large language model designed for Indian farmers.
It provides voice-based, bilingual assistance, helping farmers access information and improve practices.7
Music and entertainment
Imogen Heap partnered with the generative AI music company Jen to launch the StyleFilter program, allowing users to generate songs with the same “vibe” as licensed tracks. This initiative represents a fusion of AI and creative expression.8
Reference Links
Cem's work has been cited by leading global publications including Business Insider, Forbes, Washington Post, global firms like Deloitte, HPE and NGOs like World Economic Forum and supranational organizations like European Commission. You can see more reputable companies and resources that referenced AIMultiple.
Throughout his career, Cem served as a tech consultant, tech buyer and tech entrepreneur. He advised enterprises on their technology decisions at McKinsey & Company and Altman Solon for more than a decade. He also published a McKinsey report on digitalization.
He led technology strategy and procurement of a telco while reporting to the CEO. He has also led commercial growth of deep tech company Hypatos that reached a 7 digit annual recurring revenue and a 9 digit valuation from 0 within 2 years. Cem's work in Hypatos was covered by leading technology publications like TechCrunch and Business Insider.
Cem regularly speaks at international technology conferences. He graduated from Bogazici University as a computer engineer and holds an MBA from Columbia Business School.
Be the first to comment
Your email address will not be published. All fields are required.