Basics to setting up an AI Transformation Program - High level

Overview: Many large organisations have begun their Artificial Intelligence (AI) journeys, but some have taken unusual directions, while others are adopting a wait-and-see approach.  As a general rule, I believe most organisations should attempt to identify the most important use cases for AI and, using a basic scoring system, prioritise the easy, high-value use cases first. 

A good option is to hire an AI-focused team to implement the AI program within your business.  Hire people from the company, as they possess key domain knowledge, and pair them with IT experts, preferably those with strong AI skills.  

Tip: All Digital Transformation Projects must cover People, Process and Technology in that order of importance.

Thought: The two biggest mistakes I have observed as of July 2025 in clients are a lack of clearly defined benefits of AI projects and poor-quality, insecure data.

1. AI Idea Generation: Collate a detailed list of possible ideas.  I like to use SharePoint, it's a good idea to open ideas up to the business or make gamification of idea generation to get a good set of ideas.  You'll get lots of overlap, distil it into unique ideas, and bring the team/stakeholders together.

Figure 1. AI idea filter funnel

2. Business Impact: How does this help our customers or generate revenue, or does this improve our operations?  Can this be quantified?

3. Technological Feasibility: Can technology meet the requirement, and at what cost?  A high-level technical design is strongly recommended, as illustrated in Figure 2 below.

4. Implement: If the idea holds up to Business impact, and is possible, select the highest value to lowest effort ideas.  It is a good idea to start with the simpler ideas to get the ball rolling.

Figure 2. HLD example

Note:  This post is looking at building AI to improve your business. AI also affects all internal staff, so train them.  Allow them to use AI tools like ChatGPT, and train them not to expose internal proprietary information.  Give your users the correct tools to be the most effective.

AI Digitalisation Transformation Checklist:

1. Data Governance Framework: Do you know what data you have, can you access it, and is it secure?
2. Enterprise Data Model: Define the rules for quality and integration between the EDM areas (these may be functional or geographic), ensuring we share data and make it available.
3. Master Data Management: Ensure the data and applications are working off the same sheet of music.  
4. AI Framework: What are the limitations on what you can do within business governance? How will you identify projects and prioritise them? How do you ensure alignment with the tools?  AI is expensive; don't just start picking our individual solutions.  I'd also call this 'measure twice, cut once.'
5. Don't eat the whole Elephant: Pick small projects to start with, clearly defined goals that show the wins from AI.
6. One Process: When the AI solution/project is built, will you still use the old way? It's expensive to create, document, and support multiple processes over the lifetime of solutions that have duplicate steps to accomplish the same task. Make your process simple and the preferred or only approach.
7. People: Have they checked the process? Is it optimised and not merely converted from the old paper process to digital?  Have you engaged key stakeholders and experts?  We should have asked if this is the best process, but also verified it with non-technical stakeholders.
8. Cost Optimisation and Value Realisation: AI-based transformations like Software development or engineering projects are about whole-life cost and value, ensure you are not just looking at building something.
9. Measurements and KPI's: How do we measure value? Are you monitoring usage and costs? They tend to be the easier measurements.  Security is also essential to measure, as well as what constitutes success.  Many benefits are intangible, so list them out. Can a value be guesstimated for intangibles?

I'm sure there are plenty more items, but this would be a good check to run repeatedly.

Microsoft Azure Artificial Intelligence (AI) - AZ-900 Notes

1. Artificial Intelligence (AI) 

• AI making PC behave like human intelligence.  
• Teach PC to do task for us.  
• PC predicts using patterns and can act.  And good at looking for anomalies.
• PC uses camera/photos to look for patterns.
• Engage in useful conversations, use multiple sources of knowledge.

2. Machine Learning (ML)

• Train PC's to see patterns and see patterns, and look for anomalies.
• Example. predict stock prices by looking at factors that affect stock price.
• Anomaly Detection - Detects unusual patterns e.g. CC used in Asia when normally in Europe, but transactions 10 min apart.  Therefore likely to be fraudulent.  Sort rubbish.
• Predictive models by finding relationships.  Give model data and train the model.  
• Example: flowers have features/characteristics e.g. colour, size, no petals, ...
• Using data to teach machine
• Supervise ML - need quality data including labels.  Avg humidity, hrs sunshine, rainfall, temp, month of year (features), ice creams sold(label/class), so we feed in temp is Regression ML.  Patient has features (weight, sex, age, bmi,...) give value btwn 0-1 of the person developing diabetes.  is Classification ML
• Unsupervised ML - data is not labelled.  Just feature provided, will group into clusters.  Pulls data out and figures out it's own criteria is Clustering ML.  Useful for fraud detection.
• Training - good data based on a training set and a validation set.  Train model, with most data, then check with remaining - allows us to see how close to what happened.  Service tries to figure out relationships.  Model is used by test data - see how close/useful model is.

3. Compute Vision 

• Self driving cars, sorting. Sort rubbish.
• Facial recognition, object recognition,..
• How do computers see?  picture is cut up into pixels, data is pulled and used to find possible ans. 
• Some types on Azure:  object detection i.e. car, bike, car, bus.  Image classification i.e horse, car.  Semantic segmentation i.e. Teams blur background.   Image analysis contect by bring various tougher.  Face detection.  OCR - read image and converts to text.

4. Natural Language Processing (NLP) 

•     interpret e.g grammerly, spam check, Alexia, 
•     Knowledge Mining - Extract info from knowledge and gain insights e.g. social media marketing.

Principals:

1. Fairness - ensure bias is excluded e.g based on gender.
2. Reliable & Safety - need high confidence and in certain systems cannot fail e.g. health systems, autonomous cars.
3. Privacy & Security - Ensure data is protected and not giving away sensitive data.
4. Inclusiveness - should be fair i.e. VI users
5. Transparency - what is the model based on, what could be an issue.
6. Accountability - who is liable for AI decisions

Azure:

1. Scalability & Reliability
2. AI Resources: sit in an Azure Resource Group

AI Services in Azure:

1. Azure Machine Learning - Developers to train, test and deploy ML models.  Within a subscription, create a Azure ML Workspace (consists of: compute, data, jobs, models) can then publish as a service.  Azure ML Designer, used for creating ML pipeline, data in to train model.  Automated Machine Learning user only needs to provide the data and select the model to use, service figures it out. 
2. Cognitive Services - vision, speak, language, decision.  Rest API endpoints - have already been trained, choose the model.  Can deploy multiple parts individually or together.

    3. Azure Bot Service - develop & managing intelligent bots like chat-bots
    4. Azure Cognitive Search - Data extraction, & enrichment for indexing.  Makes data searchable.

Anomaly detector resource - wizard to setup - Add Keys and endpoints to allow access.

Create a new Azure Machine Learning Service, will create a Workspace.  Use multiple azure services such as key-vault, AI, storage accounts.   

• Launch Studio
• Add Compute Cluster
• Add Data (csv, spreadsheets, nearly any form,...)
• AutomateML (figure it out without me) and run job
• Will show trends
• Deploy the model (i.e. to a web service)
• Shows "Endpoints" - get url and a test rig.