AI vs Automation

"Artificial Intelligence and Automation are two different concepts, but when used together, they can deliver powerful business outcomes.

Let's understand the difference with some simple examples.

First, Artificial Intelligence, or AI.

AI enables machines to learn from data, identify patterns, make predictions, and support intelligent decision-making.

For example, ChatGPT can understand our questions and generate relevant responses. Similarly, image recognition systems can identify objects in pictures, and recommendation systems can suggest products based on our interests.

The key characteristics of AI are learning from data, recognizing patterns, making predictions, and adapting to changing situations. AI can also work with complex and unstructured data, such as text, images, and speech.

Now, let's look at Automation.

Automation uses technology to perform repetitive tasks according to predefined rules and instructions, with minimal or no human intervention.

For example, an organization can automate email replies, invoice processing, payroll calculations, data entry, and report generation.

Traditional automation is particularly effective when the process is predictable, the inputs are structured, and the rules are clearly defined.

Unlike traditional automation, it does not independently learn or improve its rules unless additional learning or optimization mechanisms are introduced.

Let's compare AI and Automation.

The primary purpose of AI is to provide intelligence, such as recognizing patterns or making predictions. The primary purpose of traditional automation is to execute repetitive tasks efficiently.

AI can learn from data and adapt to new situations, whereas rule-based automation follows predefined instructions.

AI can handle both structured and unstructured data, while traditional automation generally works best with structured, predictable inputs.

For example, AI can analyze customer feedback to understand sentiment, while automation can route the feedback to the appropriate department based on predefined rules.

So, what happens when we combine AI and Automation?

Imagine a customer support system. AI analyzes an incoming customer message, identifies the customer's intent and sentiment, and suggests an appropriate response. Automation then creates a support ticket, assigns it to the relevant team, and sends an acknowledgment email.

Here, AI provides the intelligence, while automation executes the workflow.

To summarize, AI and Automation are different but complementary. Automation helps us perform routine tasks efficiently, while AI adds capabilities such as pattern recognition, prediction, and intelligent decision support.

When combined effectively, they can improve productivity, reduce manual effort, minimize errors, and deliver better business outcomes."

Types of AI — Narrow AI vs Broad AI vs AGI

"Artificial Intelligence can be classified into three categories based on the range of tasks it can perform and its level of intelligence: Narrow AI, Broad AI, and Artificial General Intelligence, or AGI.

Let's understand each type with simple examples.

First, Narrow AI, also known as Artificial Narrow Intelligence, or ANI.

Narrow AI is designed to perform a specific task or a limited range of tasks. It works within a defined scope and cannot automatically perform every other task that a human can do.

For example, a spam filter identifies unwanted emails, a search engine ranks results based on your query, and a recommendation system suggests movies based on your interests.

These systems can perform their designated tasks very effectively, but their capabilities are limited to what they have been designed or trained to do.

Next, let's look at Broad AI, or General-Purpose AI.

Broad AI can perform a wide variety of tasks across different domains.

For example, ChatGPT can answer questions, write content, explain concepts, analyze data, generate code, and support learning.

Unlike traditional Narrow AI systems, which focus on specific tasks, broad-capability AI systems can handle many different kinds of requests through a single interface.

However, performing many different tasks does not mean that the system possesses human-level intelligence in every domain. It can still make mistakes, struggle with unfamiliar situations, and require human oversight.

Now, let's discuss Artificial General Intelligence, or AGI.

AGI refers to a form of AI that would possess general intellectual capabilities comparable to human intelligence across a broad range of tasks.

Such a system would be able to learn new skills, apply knowledge to unfamiliar situations, reason through complex problems, and adapt across different domains without requiring separate task-specific training for every new challenge.

For example, imagine an AI system that could independently learn a completely unfamiliar profession, transfer knowledge from one field to another, and solve new problems with flexibility comparable to a human.

That is the kind of capability people envision when discussing AGI.

However, AGI remains an unachieved research goal. Today's AI systems have not demonstrated reliable, human-level general intelligence across all domains.

Let's summarize the key differences.

Narrow AI focuses on specific tasks, such as spam filtering or face recognition.

Broad AI can perform many different tasks, such as writing, coding, data analysis, and tutoring. ChatGPT is an example of this broad, general-purpose capability.

AGI, on the other hand, refers to a more advanced, general form of intelligence that could learn, reason, and adapt across domains at a human level or beyond.

Remember, the distinction is not simply about how many tasks an AI system can perform. It is also about the depth, flexibility, and generality of its intelligence.

The key takeaway is simple: Narrow AI performs specific tasks, Broad AI handles a wide variety of tasks, and AGI represents the goal of achieving human-level general intelligence across domains.

Understanding these categories helps us distinguish the AI capabilities available today from the more advanced capabilities researchers are still working toward."

Three Major AI Application Domains

"Artificial Intelligence has many applications, but three major domains form the foundation of many modern AI systems: Data-Driven AI, Computer Vision, and Natural Language Processing.

Let's understand each domain with real-world examples and see how they work together.

First, Data-Driven AI, also known as Data Science-based AI.

This domain uses data to identify patterns, make predictions, detect anomalies, and support decision-making.

For example, recommendation engines used by Netflix and Amazon analyze user preferences to suggest relevant content or products.

In manufacturing, predictive maintenance models analyze equipment data to identify potential failures before they occur. Banks also use data-driven AI to detect suspicious transactions and potential fraud.

Within the CPMAI methodology, this domain has a strong connection with the Data Understanding and Data Preparation phases. Data quality, data lineage, historical bias, and privacy must be carefully managed before building reliable AI solutions.

Second, Computer Vision.

Computer Vision enables machines to interpret and extract meaningful information from images, videos, and other visual inputs.

Think of it as giving computers the ability to see and interpret the visual world.

For example, autonomous vehicles use computer vision to detect pedestrians, traffic signs, and surrounding vehicles. Hospitals use it to assist medical professionals in analyzing medical images, while factories use vision systems to inspect products for defects.

However, computer vision systems must be designed and tested carefully. Facial recognition, for example, can introduce fairness and demographic bias concerns if the training data is not sufficiently representative.

This is why ethical AI, responsible AI, and rigorous testing are important when developing computer vision applications.

Third, Natural Language Processing, or NLP.

NLP focuses on enabling computers to work with human language, including reading, interpreting, generating, and translating text, as well as processing spoken language.

For example, ChatGPT uses large language models to generate responses, virtual assistants can respond to voice commands, and translation systems convert text from one language into another.

NLP is also used for sentiment analysis, where AI identifies whether customer feedback expresses a positive, negative, or neutral opinion.

In the CPMAI context, language-based AI solutions raise important considerations around explainability, transparency, hallucinations, alignment, and data privacy.

Now, here is where things become really interesting: these three domains can work together.

Consider a smart manufacturing facility.

Computer Vision examines images of products on a production line to identify defects. Data-Driven AI analyzes equipment performance and production trends to predict possible failures. NLP allows an operator to ask questions in natural language and receive an explanation of the production data.

When these capabilities are integrated with robotics and automation, the system can support inspection, maintenance planning, and operational decision-making.

Another example is a multimodal AI system. It can process an image, interpret a spoken question about that image, and generate a text-based answer. This combines visual understanding with language processing.

Finally, consider autonomous systems, such as self-driving vehicles. They perceive their surroundings using sensors and computer vision, use AI models and planning algorithms to select actions, and control the vehicle through software and automation. Human oversight and safety controls remain important.

To summarize, remember these three key ideas.

Data-Driven AI helps us discover patterns and make predictions.

Computer Vision helps machines interpret visual information.

Natural Language Processing helps machines work with human language.

Individually, these domains solve different problems. Together, they enable more capable AI applications, from intelligent customer support to smart factories and autonomous systems.

As we move forward in our AI journey, understanding these domains will help us identify the right AI capabilities for the right business problem."

Functional Types of AI

"Artificial Intelligence can perform different functions depending on how it is designed and deployed. In this slide, we will explore four important functional types of AI: Predictive AI, Generative AI, Agentic AI, and Physical or Industrial AI.

Let's understand each one with a simple explanation and real-world examples.

First, Predictive or Analytical AI.

Predictive AI analyzes historical data to identify patterns, detect anomalies, and forecast future outcomes.

Think of it as an analyst who studies past information to help us make better decisions.

For example, banks use predictive models to identify potentially fraudulent credit card transactions. Retailers use demand forecasting to estimate how many products customers may purchase in the future.

The key idea is that Predictive AI helps us understand what is happening, what might happen next, and where we should pay attention.

Second, Generative AI.

Generative AI learns patterns from training data and uses them to create new content, such as text, images, audio, video, and computer code.

Think of it as a creator or co-pilot who helps us turn ideas into useful outputs.

For example, ChatGPT can draft a project report, summarize a lengthy document, or help generate code. Image-generation tools such as Midjourney can create visual concepts from text prompts.

Unlike traditional predictive analytics, which primarily forecasts outcomes or identifies patterns, Generative AI produces new content based on the prompt and its learned patterns.

However, its outputs still need to be reviewed because they may contain errors or inaccurate information.

Third, Agentic AI.

Agentic AI refers to AI systems designed to pursue goals by planning steps, selecting actions, using tools, and adjusting their approach based on results.

Think of it as an assistant that does more than provide suggestions. It can work through a sequence of tasks to achieve a defined objective.

For example, imagine an AI agent assigned to investigate a cybersecurity alert. It might gather relevant logs, analyze suspicious activity, consult approved security tools, and recommend or initiate a predefined containment action.

Depending on its permissions and safeguards, it may also execute certain remediation steps, such as applying an approved configuration change.

The important distinction is that Agentic AI can coordinate multiple actions toward a goal rather than simply generating a response to a single prompt.

Human approval, access controls, and monitoring remain essential, especially when actions could affect business operations or security.

Fourth, Physical or Industrial AI.

Physical AI connects intelligent software with physical machines, sensors, and control systems so that AI-enabled systems can interact with the real world.

Think of it as an operator that can sense its environment and help machines perform physical tasks.

For example, AI-enabled industrial robots can inspect products and assist with manufacturing operations. Autonomous forklifts can use sensors and perception systems to navigate warehouses, while intelligent irrigation systems can use soil moisture and weather data to optimize water usage.

These applications combine AI with robotics, sensors, control systems, and automation to carry out physical activities.

Now, let's compare the four functional types.

Predictive AI helps us forecast outcomes and detect patterns.

Generative AI creates new content.

Agentic AI plans and coordinates actions to achieve goals.

Physical or Industrial AI enables intelligent systems to interact with the physical environment.

These categories are not mutually exclusive. A single solution may combine several of them.

For example, a smart factory might use Predictive AI to forecast equipment failures, Generative AI to summarize maintenance reports, Agentic AI to coordinate maintenance workflows, and Physical AI to control robotic equipment.

The key takeaway is this: Predictive AI helps us anticipate, Generative AI helps us create, Agentic AI helps us act toward goals, and Physical AI helps intelligent systems operate in the real world.

Understanding these differences helps us select the right AI capabilities for the right business problem."

AI vs. Automation vs. Traditional Software

"Traditional software, automation, and Artificial Intelligence are three important technologies used in modern businesses. Although they may work together, each serves a different purpose.

Let's understand the differences with simple examples.

First, Traditional Software.

Traditional software follows explicit instructions written by developers. It uses predefined logic to process inputs and produce expected outputs.

For example, a payroll system calculates employee salaries using predefined formulas, tax rules, and employee data. Similarly, a billing application generates invoices based on product prices and quantities.

Traditional software is generally predictable when given the same inputs and conditions. However, when business rules change, developers may need to modify the program.

Think of traditional software as a system that follows instructions exactly as programmed.

Second, Automation.

Automation uses technology to execute repetitive tasks according to predefined rules and workflows.

For example, imagine an invoice approval process. When an invoice is received, the system checks whether required fields are present, routes the invoice to the appropriate manager, and sends a notification when approval is completed.

Other examples include employee onboarding workflows, data entry, report generation, and automated email notifications.

The main benefit of automation is that it reduces repetitive manual work, improves consistency, and saves time.

The important point is that traditional rule-based automation does not independently learn from experience. If the rules or workflow need to change, they generally need to be updated.

Third, Artificial Intelligence, or AI.

AI systems use techniques such as machine learning to identify patterns in data, make predictions, generate content, and support decision-making.

For example, a business can use AI to predict which customers are likely to stop using its services. Retailers can forecast future demand, while computer vision systems can recognize objects in images.

Generative AI can also create reports, draft emails, and develop content based on user instructions.

Unlike conventional rule-based systems, machine learning models can learn statistical patterns from training data and generalize to new inputs. However, AI does not automatically become more accurate simply because it receives more data. Improvement depends on data quality, model design, evaluation, and sometimes retraining.

AI can also handle uncertainty, but its predictions and outputs may be incorrect and should be evaluated appropriately.

Now, let's understand the difference with one practical example: processing an invoice.

Traditional software calculates the invoice total using programmed formulas.

Automation receives the invoice, routes it for approval, and updates the relevant business system according to predefined rules.

AI can extract information from an invoice image, identify unusual charges, and flag potential discrepancies for review.

When these technologies are combined, an organization can build a more intelligent and efficient invoice-processing system.

Traditional software provides the programmed logic, automation coordinates the workflow, and AI adds capabilities such as recognition, prediction, and intelligent analysis.

So, remember these three key distinctions.

Traditional software follows programmed instructions.

Automation executes repetitive tasks and workflows.

AI learns patterns from data or uses other AI techniques to perform tasks such as prediction, recognition, reasoning support, and content generation.

These technologies are not competitors. They are complementary capabilities that organizations can combine to improve productivity, reduce errors, and achieve better business outcomes."

"In the CPMAI methodology, understanding how AI creates business value begins with three important capabilities: Perception, Prediction, and Planning.

These are commonly known as the three Ps of AI. They help us understand how AI systems interpret information, estimate what may happen next, and determine appropriate actions to achieve a goal.

Let's explore each capability with practical business examples.

First, Perception.

Perception is the ability of an AI system to interpret inputs from the world, such as images, audio, video, and sensor data.

Think of perception as the eyes and ears of an intelligent system. It helps the system understand what is happening in its environment.

For example, in a manufacturing facility, a computer vision system can examine product images to identify defects. A speech recognition system can convert spoken instructions into text. Similarly, an equipment monitoring system can analyze sensor readings to identify unusual operating conditions.

In simple terms, perception helps AI understand the available inputs and extract meaningful information from them.

Second, Prediction.

Prediction uses available data and learned patterns to estimate what is likely to happen in the future.

For example, a retailer can analyze historical sales data, seasonal trends, and customer demand to forecast how many products it may need next month.

A bank may use predictive models to estimate which customers are at risk of leaving, while a manufacturing organization can predict the likelihood of equipment failure.

Notice that prediction does not tell us exactly what will happen. Instead, it estimates possible outcomes based on the available data and the model's assumptions.

These predictions help organizations make more informed decisions, manage risks, and allocate resources effectively.

Third, Planning.

Planning involves selecting actions or developing a sequence of steps to achieve a defined goal while considering constraints and available resources.

For example, a logistics company needs to deliver packages to multiple destinations. An AI-based planning system can evaluate routes, delivery deadlines, traffic conditions, and vehicle capacity to identify an efficient delivery plan.

Similarly, workforce scheduling systems can allocate employees to shifts, and autonomous robots can plan movements through a warehouse.

Planning may involve evaluating different alternatives, considering trade-offs, and adjusting the plan when conditions change.

Importantly, planning does not always mean that the AI system executes the actions itself. Execution may be performed by another software component, an automated workflow, a robot, or a human operator.

Now, let's see how the three Ps can work together.

Imagine a smart manufacturing plant.

First, Perception analyzes images and sensor readings to detect an unusual condition in a machine.

Next, Prediction estimates the likelihood of equipment failure and identifies the potential operational impact.

Finally, Planning determines the best response, such as scheduling maintenance, allocating a technician, or adjusting the production schedule.

Together, these capabilities can help the organization reduce downtime, control costs, and improve operational efficiency.

This example also illustrates how AI can move beyond simply analyzing information to supporting goal-directed decisions and actions.

So, when should an organization use AI?

AI is particularly valuable when a business problem requires interpreting complex inputs, predicting uncertain outcomes, or planning actions across multiple alternatives and constraints.

However, if the task involves straightforward calculations or a simple, repetitive process governed by stable rules, traditional software or automation may be a more practical and cost-effective solution.

Remember the three Ps: Perception helps AI understand, Prediction helps AI anticipate, and Planning helps AI choose actions toward a goal.

When these capabilities address a genuine business need, AI can create measurable value for the organization."

The 3 Ps of AI — Uber Planning a Passenger's Trip

"Let's understand the three Ps of AI — Perception, Prediction, and Planning — through a familiar real-world example: booking an Uber ride from home to the airport.

How does the system understand the current situation, estimate journey times, and select a suitable route? Let's explore.

Step one: Perception — What is happening?

When a passenger books a ride, Uber's systems use information from multiple sources to understand the current situation.

This may include the passenger's pickup location, nearby available drivers, current traffic conditions, road restrictions, and relevant weather information.

For example, the system identifies the passenger's location at home, detects nearby drivers, and obtains information about congestion along the roads leading to the airport.

This is Perception. The system gathers and interprets information about the current environment.

Step two: Prediction — What is likely to happen?

Next, AI models and routing algorithms use historical and real-time information to estimate possible journey times.

Suppose the usual route, Route A, is expected to take 50 minutes because of heavy traffic.

An alternative route, Route B, is estimated to take just 38 minutes, while Route C may take 45 minutes.

These estimates help the system anticipate delays and compare possible outcomes before selecting a route.

Remember, these journey times are illustrative. Actual estimates depend on traffic, road conditions, and other factors, and they may change during the trip.

This is Prediction: estimating what is likely to happen based on available information.

Step three: Planning — What should we do?

Now the system evaluates the available routes and selects a suitable option based on estimated travel time, distance, traffic conditions, road constraints, and other relevant factors.

In our example, Route B has the shortest estimated journey time of 38 minutes.

The system can provide navigation instructions, while driver-assignment and dispatch systems help match the trip with a suitable available driver.

As the journey progresses, the system can use updated traffic information to recalculate the estimated arrival time or recommend a different route when appropriate.

This is Planning: selecting and updating actions to achieve the goal of reaching the airport efficiently.

Let's connect the three Ps.

Perception helps the system understand the current situation.

Prediction estimates possible outcomes, such as journey duration and traffic delays.

Planning selects an appropriate route and supports navigation toward the destination.

Together, these capabilities can contribute to faster journeys, more efficient routing, and a better passenger experience.

The key takeaway is simple: Perception tells us what is happening, Prediction estimates what may happen, and Planning determines what to do next.

This is how the three Ps of AI can work together to solve a practical, everyday business problem."

The Seven Patterns of AI Projects — PMI-CPMAI

"Not all AI projects solve the same type of business problem. Some personalize customer experiences, some recognize images, some predict future outcomes, and others take actions autonomously.

In the PMI-CPMAI context, understanding different AI patterns helps us identify the appropriate AI capabilities for a business problem and select a suitable solution.

Let's explore seven important patterns of AI projects.

Pattern number one: Hyperpersonalization.

Hyperpersonalization uses data and AI to tailor experiences, recommendations, and services to individual users.

For example, an e-commerce platform can recommend products based on a customer's browsing history, purchase behavior, and preferences. Streaming platforms can similarly recommend movies or shows suited to individual interests.

The objective is to make the experience more relevant to each user while respecting privacy and data protection requirements.

Pattern number two: Conversational and Human Interaction.

This pattern enables people to interact with AI through text, voice, images, or other natural forms of communication.

For example, a virtual assistant can answer customer questions, help employees find information, or translate conversations between different languages.

Generative AI and large language models have significantly expanded the capabilities of conversational systems. However, appropriate safeguards are necessary to handle inaccurate responses, sensitive information, and complex requests.

Pattern number three: Recognition.

Recognition systems identify or classify objects, speech, text, or other information from data.

For example, computer vision can inspect products on a manufacturing line to detect defects. Speech recognition can convert spoken words into text, while image recognition can identify objects in photographs.

The main objective is to recognize meaningful information from inputs that may be difficult to process using conventional rules alone.

Pattern number four: Pattern and Anomaly Detection.

This pattern identifies recurring structures, unusual behavior, or deviations from expected patterns in data.

For example, a bank may use anomaly detection to flag unusual credit card transactions. A cybersecurity system can identify suspicious network activity that may indicate a security incident.

The important distinction is that an anomaly is not necessarily fraud or an attack. It is a signal that may require further investigation.

Pattern number five: Predictive Analytics and Decision Support.

Predictive analytics uses historical data and statistical or machine learning models to estimate future outcomes and support decision-making.

For example, a retailer can forecast product demand to improve inventory planning. A manufacturing company can predict equipment failure and schedule preventive maintenance.

These predictions help managers evaluate risks, allocate resources, and make better-informed decisions.

However, predictions are estimates rather than guarantees, and their reliability depends on the quality of the data and the model.

Pattern number six: Goal-Driven Systems.

Goal-driven AI systems select or learn actions to achieve a defined objective. They may evaluate alternatives, use feedback, and improve their decisions over repeated interactions.

For example, a system might optimize energy consumption in a building while maintaining comfortable temperatures. Reinforcement learning can be used in some applications to learn a policy through rewards and penalties.

The key idea is that the system is not limited to identifying patterns or making predictions. It also considers which actions may help achieve a goal.

The objective, constraints, reward design, and safety controls must be carefully defined.

Pattern number seven: Autonomous Systems.

Autonomous systems perceive their environment, make decisions, and perform actions with varying degrees of human supervision.

Examples include warehouse robots that move goods, autonomous vehicles that navigate roads, and robotic systems that perform industrial tasks.

These systems may combine several AI capabilities, including perception, prediction, planning, recognition, and goal-directed decision-making.

Because their actions can affect the physical world, safety engineering, testing, monitoring, and appropriate human oversight are especially important.

Now, let's look at the bigger picture.

Imagine an intelligent warehouse.

Hyperpersonalization could tailor the experience of individual customers. A conversational assistant could answer questions about orders. Recognition systems could identify packages, and anomaly detection could flag suspicious activity.

Predictive analytics could forecast demand, goal-driven systems could optimize warehouse operations, and autonomous robots could transport goods.

Notice that one AI project can combine several patterns to solve a larger business problem.

The key takeaway is this: Choose the AI pattern based on the business need, not simply because a particular technology is popular.

Define the problem first, determine the value the solution should deliver, select the appropriate AI capabilities, and establish the necessary data, risk, governance, and human oversight controls.

That is how organizations can move from experimenting with AI to delivering meaningful business value."

Why AI Projects Fail and How AI Differs from Conventional Projects

"AI projects offer tremendous opportunities, but not every AI initiative delivers the expected business value. Some projects fail because organizations focus too much on technology and not enough on the business problem, data quality, and implementation challenges.

In this slide, we will explore seven common reasons why AI projects fail and then compare AI projects with conventional software projects.

Let's begin with the seven reasons for AI project failure.

Number one: Unclear Business Problem.

An AI project must begin with a clearly defined business problem and a measurable objective.

For example, an organization may want to implement AI simply because it is a popular technology. But if the team cannot explain what problem AI will solve or how success will be measured, the project may deliver little business value.

Number two: Poor Data Quality.

AI models depend on data. If the data is incomplete, inaccurate, inconsistent, or biased, the model may produce unreliable results.

For instance, a predictive maintenance model trained on poor-quality equipment data may fail to identify important warning signs.

This is why data understanding, data preparation, and data quality management are essential.

Number three: Wrong Use Case for AI.

Not every business problem requires AI.

Suppose an organization wants to calculate employee salaries using predefined formulas. Traditional software may perform this task more reliably and cost-effectively than an AI solution.

AI should be selected when its capabilities provide a meaningful advantage over simpler alternatives.

Number four: Lack of Domain Expertise.

AI specialists may understand algorithms and models, but they also need input from people who understand the business.

For example, a healthcare AI team needs appropriate clinical expertise to interpret medical data and validate the usefulness of its results.

Without domain knowledge, a technically accurate model may still solve the wrong problem.

Number five: Limited Stakeholder Engagement.

AI projects require collaboration among business users, data scientists, technical teams, managers, and other relevant stakeholders.

If users are not involved in testing and validating the solution, the final system may not fit their actual needs.

Change management and organizational support are equally important for successful adoption.

Number six: Unrealistic Expectations.

AI does not guarantee perfect predictions, immediate results, or complete autonomy.

Teams need realistic objectives, experimentation, validation, and continuous learning.

Setting achievable expectations helps organizations evaluate progress and improve the solution over time.

Number seven: Deployment and Operational Challenges.

A model that performs well in a development environment may not perform equally well in production.

Integration with existing systems, changing data patterns, model drift, security, scalability, and ongoing maintenance can all create challenges.

AI projects therefore require operational monitoring, appropriate retraining, and continuous performance evaluation.

Now, let's compare AI projects with conventional software projects.

The first difference is the objective and nature of the output.

Conventional software generally follows programmed logic to produce predictable results for defined inputs and conditions. AI systems may produce probabilistic outputs based on learned patterns, so their accuracy and reliability must be evaluated.

Second, consider requirements.

Traditional software projects often begin with relatively stable requirements and detailed specifications. AI projects may require more experimentation because the team needs to discover whether the available data and models can achieve the desired outcome.

Third, look at data and the development approach.

Conventional software can often be developed primarily through explicit rules and business logic. AI projects require additional attention to data collection, preparation, model training, validation, and tuning.

The AI development process is therefore often iterative, although conventional software projects can also use iterative methods.

Fourth, consider success criteria.

For conventional software, success may include meeting functional requirements, performance expectations, delivery dates, and budgets.

For AI projects, success also depends on model performance, business impact, user adoption, and the ability to maintain reliable results after deployment.

Fifth, consider maintenance.

Traditional software maintenance often involves fixing defects, updating features, and adapting to changes in requirements.

AI systems may require additional monitoring for model drift, changing data patterns, and declining prediction quality. Retraining or recalibration may be necessary when performance deteriorates.

Finally, consider skills and risks.

Conventional projects typically require software engineering and project management skills. AI projects may additionally require data science, machine learning engineering, domain expertise, and responsible AI capabilities.

They also introduce risks involving data bias, privacy, explainability, model performance, and regulatory compliance.

So, what is the key takeaway?

AI projects are not simply conventional software projects with an AI model added to them. They require a clear business problem, suitable data, appropriate model selection, iterative experimentation, stakeholder engagement, and continuous operational monitoring.

In the PMI-CPMAI context, the focus is on taking a business problem through a structured, iterative AI project lifecycle while validating both technical feasibility and business value.

Remember: successful AI projects begin with the business problem, depend on the quality of the data, and succeed when the solution delivers measurable value in real-world operations."

The Six CPMAI Methodology Phases

"Successful AI projects require more than selecting an algorithm or building a model. They need a structured approach that connects business needs, data, AI development, testing, and real-world business value.

The CPMAI methodology provides a six-phase approach to guide AI projects from identifying a business problem to deploying and operating an AI solution.

Let's walk through all six phases.

Phase one: Matching AI with Business Needs.

The first step is to understand the business problem and determine whether AI is an appropriate solution.

The team identifies business objectives, evaluates potential AI use cases, assesses feasibility, and defines measurable success criteria.

For example, a manufacturing company may want to reduce unexpected equipment breakdowns. The team evaluates whether predictive maintenance could help reduce downtime and maintenance costs.

The key output of this phase is a prioritized AI use case with clearly defined business value and success criteria.

Phase two: Identifying Data Needs for AI Projects.

Once the business use case has been selected, the next step is to determine what data is required to develop the AI solution.

The team identifies relevant internal and external data sources, assesses data availability and quality, and addresses data access, privacy, and governance requirements.

For our predictive maintenance example, the team may need historical machine-failure records, temperature readings, vibration measurements, and maintenance logs.

The key output is a data requirement specification and a plan for obtaining the necessary data.

Phase three: Managing Data Preparation Needs for AI Projects.

Having data available does not automatically make it suitable for AI.

The data must be collected, integrated, cleaned, transformed, and prepared for model development.

The team addresses missing values, inconsistent records, duplicate entries, and potential data bias. It also establishes appropriate data pipelines and quality controls.

In our example, sensor readings from different machines may need to be standardized and matched with historical maintenance records so that the model can identify meaningful patterns.

The key output is a prepared, high-quality dataset that is ready for model development.

Phase four: Iterating Development and Delivery of AI Projects.

Now the team develops the AI solution through an iterative process.

It selects appropriate algorithms and tools, trains models using prepared data, evaluates their performance, and refines them based on the results.

For example, the team may develop a predictive model that estimates the likelihood of equipment failure based on sensor readings and historical breakdowns.

If the initial model does not perform well enough, the team may revisit the features, data preparation, or model configuration and run further experiments.

The key output is a trained and validated AI model, together with the relevant technical documentation.

The important point is that AI development is rarely a simple, one-directional process. The team learns from experiments and improves the solution incrementally.

Phase five: Testing and Evaluating AI Systems.

Before deploying the solution, the team must determine whether it performs reliably and delivers the expected business value.

This involves testing model performance against predefined metrics, validating the solution against business requirements, and assessing security, fairness, reliability, and other relevant risks.

In our example, the team evaluates whether the predictive maintenance model can identify likely equipment failures with sufficient reliability and early warning time.

It must also consider false alarms and missed failures. A model that raises too many unnecessary alerts may waste maintenance resources, while a model that misses critical failures may create operational risks.

User acceptance testing and deployment-readiness assessments help determine whether the solution is ready for production.

The key output is a test and evaluation report with a go-or-no-go recommendation.

Phase six: Operationalizing AI.

The final phase involves deploying the AI solution into the production environment and ensuring that it continues to deliver business value.

The team integrates the solution with existing systems, monitors performance, manages data and model drift, and plans for maintenance and updates.

For predictive maintenance, the model might be integrated with a factory's equipment monitoring and maintenance management systems so that relevant alerts can reach the maintenance team.

After deployment, the organization continues to monitor whether the model remains accurate as equipment conditions and operating patterns change.

It also measures the actual business outcomes, such as reductions in unexpected breakdowns, downtime, and maintenance costs.

The key output is a production AI solution with an ongoing monitoring, maintenance, and value-realization plan.

Let's summarize the six phases.

Phase one matches AI with business needs.

Phase two identifies data requirements.

Phase three prepares and manages the data.

Phase four develops and refines the AI solution.

Phase five tests and evaluates the solution.

Phase six deploys and operates AI in the real world.

Remember, these phases provide a structured framework, but AI projects may revisit earlier activities when new information or evaluation results emerge.

The ultimate goal is not simply to build an AI model. It is to deliver a responsible, effective, and sustainable solution that creates measurable business value.

That is the essence of the CPMAI methodology: start with the business problem, work systematically with data, learn through iteration, validate the results, and continuously monitor the solution after deployment."

Trustworthy AI — PMI-CPMAI

"As Artificial Intelligence becomes increasingly integrated into business decisions, one important question arises: How can we ensure that AI creates value without causing unnecessary harm?

The answer is Trustworthy AI.

In the PMI-CPMAI context, trustworthy AI focuses on developing and using AI systems that are ethical, responsible, transparent, explainable, and properly governed.

Let's understand these five important principles with practical examples.

First, Ethical AI.

Ethical AI means ensuring that AI systems align with human values, respect human rights, promote fairness, and avoid causing harm.

For example, imagine an organization using AI to screen job applications. If the system unfairly disadvantages candidates based on gender or other irrelevant characteristics, it raises serious ethical concerns.

Organizations must identify and minimize bias, consider fairness, and assess the potential social and environmental impact of their AI solutions.

The key question is: Is the AI system doing what is fair and right?

Second, Responsible AI.

Responsible AI focuses on accountability and taking ownership of the design, development, deployment, and use of AI systems.

For example, if an AI system makes an incorrect recommendation that affects a business decision, the organization must have clear responsibilities for investigating the issue and taking corrective action.

This requires defined roles, appropriate human oversight, risk management, and continuous monitoring.

Remember, using AI does not remove human accountability.

Third, Transparent AI.

Transparency means being open about how an AI system is developed, what data it uses, what its purpose and limitations are, and how AI contributes to decisions.

For example, if a bank uses AI to support loan assessments, customers and relevant stakeholders should receive appropriate information about the system's role in the decision-making process.

Transparency does not necessarily mean publishing every line of source code. It means providing relevant information so that stakeholders can understand how the system is being used and what its limitations are.

The key question is: Are we being open about how AI is used?

Fourth, Explainable AI.

Explainable AI focuses on providing understandable reasons for an AI system's outputs or decisions.

For example, suppose an AI system flags a financial transaction as potentially fraudulent. An explanation might identify unusual transaction amounts, an unfamiliar location, or behavior that differs from the customer's normal pattern as contributing factors.

Such explanations can help analysts investigate the alert and decide what action to take.

The explanation should be appropriate for its audience. A technical team may need detailed model diagnostics, while a business user may need a concise explanation of the main factors influencing the result.

It is also important to remember that an explanation should accurately reflect the model's behavior rather than simply provide a convincing-sounding story.

Fifth, Governed AI.

AI governance establishes the policies, standards, processes, and oversight structures needed to ensure that AI is developed and used appropriately.

For example, an organization may establish an AI governance framework that defines who can approve AI projects, which risk assessments are required, how sensitive data must be protected, and how deployed systems are audited.

Governance also includes compliance with applicable laws and regulations, data governance, ethical review, and ongoing assurance.

The objective is to ensure that AI systems remain aligned with organizational policies and applicable requirements throughout their lifecycle.

Now, let's connect all five principles.

Imagine an organization deploying AI to support employee recruitment.

Ethical AI helps identify potential discrimination.

Responsible AI establishes accountability and human oversight.

Transparent AI communicates when and how AI is being used.

Explainable AI helps recruiters understand the factors influencing recommendations.

Governed AI establishes the policies, approvals, audits, and controls needed to manage the system.

Together, these principles help build confidence in the technology while reducing avoidable risks.

So, what is the ultimate outcome of Trustworthy AI?

It is not simply a model that performs well technically. It is an AI solution that creates business value while respecting people, managing risks, meeting applicable requirements, and maintaining stakeholder trust.

As AI project managers, we must consider these principles throughout the AI project lifecycle, from selecting the business use case to deploying and monitoring the solution.

Remember: Ethical AI focuses on fairness and values. Responsible AI focuses on accountability. Transparent AI focuses on openness. Explainable AI focuses on understandable reasons. Governed AI focuses on policies, controls, and oversight.

When these principles work together, AI can deliver more trustworthy, responsible, and sustainable business outcomes."

Transparency vs. Explainable AI (XAI)

"Transparency and Explainable AI are two important concepts in Trustworthy AI. Although they are closely related, they address different questions.

Let's understand the difference using a simple example of a bank rejecting a loan application.

First, Transparency.

Transparency answers the question: Can I understand how the AI system works, how it is used, and how it is managed?

Transparency provides visibility into the AI system, including its purpose, data sources, limitations, ownership, and governance.

For example, a bank may disclose that it uses AI to help assess loan applications. It may explain what types of data the system uses, what its limitations are, and how human employees oversee the process.

This information helps customers, regulators, and other stakeholders understand how AI is being used and who is accountable for its operation.

The key benefit of transparency is informed oversight, accountability, and trust.

Now, let's look at Explainable AI, or XAI.

Explainable AI answers a different question: Why did the AI system produce this particular result?

It focuses on explaining the factors that contributed to a specific prediction or decision.

Consider the same loan application. Suppose the bank's AI system rejects an application. An explainability technique might identify a high debt-to-income ratio and a short credit history as important factors contributing to the model's prediction.

Techniques such as feature importance, local explanations, and SHAP values can help show how different input features influence a model's output.

In simple terms, Explainable AI helps people investigate and understand why a particular result occurred.

However, an explanation does not automatically prove that the decision is correct or fair. The explanation must be interpreted carefully, and the underlying model and decision process still need to be evaluated.

Let's compare the two concepts directly.

Transparency focuses on the AI system as a whole.

It helps us understand what the system does, what data it uses, what its limitations are, and how it is governed.

Explainable AI focuses more specifically on an individual prediction or output.

It helps us understand which factors influenced that result and why the model produced it.

Transparency supports accountability and oversight, while Explainable AI supports understanding, scrutiny, and the ability to challenge particular predictions.

Let's take another example: AI in hiring.

Suppose a company uses AI to help screen résumés.

Transparency means informing applicants that AI is being used, describing its role in the screening process, providing appropriate information about the data involved, and explaining how human oversight works.

Explainable AI, on the other hand, might help a recruiter understand why a particular résumé received a low match score. For example, the system may not have detected certain skills required for the role.

The explanation can help the recruiter review the result, check whether relevant experience was overlooked, and decide whether further assessment is necessary.

Now, remember one important point.

Transparency and Explainable AI complement each other, but neither guarantees that an AI system is accurate, fair, or trustworthy.

A system can be transparent about its design and data sources yet still produce inaccurate predictions. Similarly, a model may provide an explanation for a prediction that is itself unreliable or affected by bias.

That is why Trustworthy AI also requires responsible development, appropriate testing, fairness assessments, governance, and ongoing monitoring.

**The key takeaway for PMI-CPMAI is this: Transparency helps us understand the AI system and how it is used, whereas Explainable AI helps us understand why a particular AI output occurred.

Both are essential capabilities for building AI solutions that can be evaluated, challenged, and used responsibly."

SHAP — Explainable AI

"Have you ever wondered why an AI model makes a particular prediction? For example, why does a bank's AI system approve one loan application but reject another?

To answer such questions, we use Explainable AI techniques. One of the most popular techniques is SHAP, which stands for Shapley Additive exPlanations.

Let's understand how SHAP works and why it is important.

First, what is SHAP?

SHAP is a game theory-based approach that explains a machine learning model's prediction by estimating how much each feature contributes to the model's output.

Think of a prediction as the result of several contributing factors. SHAP helps us understand which factors push the prediction higher and which push it lower.

For example, when assessing a loan application, relevant features might include income, credit score, loan amount, age, and employment history.

SHAP assigns a contribution value to each feature, helping us understand its influence on the prediction.

Now, let's understand the key idea behind SHAP.

SHAP uses a concept from cooperative game theory called Shapley values.

Imagine that several players work together to achieve a result. Game theory provides a way to fairly allocate the total contribution among those players.

SHAP applies a similar idea to machine learning. The features are treated like contributors, and their Shapley values estimate how much each feature contributes to a prediction relative to a baseline.

A positive SHAP value pushes the model output higher, while a negative SHAP value pushes it lower, assuming the direction is interpreted relative to the model's output scale.

Let's see how SHAP works in four steps.

Step one: Establish a baseline prediction. This might be the model's average output across a suitable reference dataset.

Step two: Evaluate how the model's prediction changes when different features are considered, using the SHAP method's approach to comparing feature contributions across possible combinations.

Step three: Calculate a SHAP value for each feature, representing its contribution relative to the baseline.

Step four: Combine the baseline with the feature contributions to reconstruct the model's output.

In simplified form, the model output equals the baseline value plus the sum of the SHAP values. For many models, this explanation is calculated on a particular output scale, such as the raw model output or log-odds, rather than directly on the probability scale.

Now, let's explore the three important SHAP visualizations shown on the slide.

First, the Global Explanation, or Feature Importance plot.

This plot helps us understand which features are most influential across a dataset.

In our example, age has the highest mean absolute SHAP value, followed by income, credit score, loan amount, and employment length.

Mean absolute SHAP values measure the average magnitude of each feature's contribution, regardless of whether that contribution is positive or negative.

This helps analysts identify which features tend to have the greatest influence on model predictions overall.

Second, the Local Explanation for a Single Prediction.

This explains how individual features contribute to one particular model output.

Suppose a loan model begins with a baseline output of 0.20 and produces a final output of 0.72. The SHAP explanation illustrates how the contributing features move the output away from the baseline.

Some features push the output higher, while others push it lower.

The important point is that these values describe the model's behavior. They do not automatically establish that a feature caused the real-world outcome or that the decision is fair.

Third, the SHAP Summary Plot.

This plot shows the distribution of SHAP values for individual features across many predictions.

Each dot represents a feature's contribution for a particular observation. The horizontal position indicates whether the contribution pushes the model output higher or lower, while the color represents the feature value, from low to high.

This visualization helps analysts examine patterns, variation, and potential relationships between feature values and model predictions.

Why is SHAP useful?

SHAP provides both global and local explanations. It can help business users understand model behavior, support model debugging, and identify potential issues that deserve further investigation.

It is used in areas such as credit risk assessment, healthcare, customer churn prediction, fraud detection, and HR analytics.

For example, in fraud detection, SHAP may help an analyst understand which transaction features contributed to a high-risk prediction.

However, SHAP is not a guarantee of fairness, accuracy, or regulatory compliance. Its explanations must be interpreted carefully, especially when features are correlated or the model is complex.

The key takeaway is this: SHAP helps us move beyond simply receiving an AI prediction to understanding how different features contribute to that prediction.

For AI project managers, this is an important tool for improving model interpretability, investigating potential bias, supporting informed decisions, and building more transparent AI solutions."