Introduction
The premise: I have a time machine and I were to go back to a time before I started researching implementing AI Chatbot solutions, what are 5 things I should tell myself to make my journey so much easier?
I would still encourage myself to do extensive research because most of it will have practical relevance, however, learning some things earlier on will enable quicker hands-on experience with less ambiguity. That is what I hope to pass on to you, 5 important things that will help you get down and dirty with AI quickly.
So exactly what would I tell myself?
A little background – setting the scene…
A common question worqinc receives from customers that we have onboarded on partner solutions or developed custom software for is: “Is it possible to integrate AI into this solution”?
While we are always happy to elaborate on how we can help with this, either directly or via one of our partners, showing in real-time how and what we can do pulls more weight.
1. What is AI vs What is not AI?
I would start with a demonstration on what is actually “AI” versus what looks like AI, but is not quite.
| Qualities of an actual AI solution | Qualities of an “AI” solution |
| Learns & Adapts – Improves over time based on data and patterns. | Rule-Based Automation – Follows predefined instructions without learning |
| Processes Unstructured Data (Text, images, audio, and video) | Basic Data Processing |
| Makes Predictions & Decisions | Predefined Responses – Static chatbots that follow decision trees. |
| Understands Natural Language | IF-THEN Logic – Hardcoded logic without pattern recognition. |
| Uses Machine Learning & Neural Networks as core techniques | Simple Software Features – Optical Character Recognition (OCR) or search functions without learning mechanisms. |
2. What important qualities of AI should we keep an eye out for?
Artificial Intelligence has taken most industries by storm, and companies advertise a plethora of features and qualities for their AI solutions… how do you separate meaningful qualities from those that are ambiguous, generic, or outright buzzwords?
While it is circumstantial, and we cannot generalise because we have to understand the context said quality is used in, notable terms include:
• Grounding ✅ – Responses are based on reliable, verifiable data sources (vs. hallucinations). One of the most common issues that AI sceptics rightfully reference, is the issue of hallucinations. There are countless stories and anecdotes of individuals relying blindly on what an AI product confidently returned as fact, only to realise later that the content had no basis in any actual data. Grounded responses address this problem by explicitly ensuring that outputs are backed by real, verifiable sources.
• Explainable AI (XAI) ✅ – Model offers justification for generated answers.
• Smart assistant ⚡– Very much a buzzword, you can explain the same sentiment in a much more meaningful way.
• Human-in-the-Loop (HITL) ✅ – AI solution design approach to actively train, tune, test and validate outputs to optimise accuracy and reliability.
• Next-gen AI ❔ – Often misused and does not actually refer to upcoming AI models or technologies.
3. Where and how do I start experimenting with building AI components?
Your mileage may vary, however, as both a Microsoft Tech Stack aficionado and an individual who thrives in understanding how something works by getting their hands dirty – you would be hard-pressed to find me recommending any other starting point than Azure AI Foundry (previously Azure AI Studio).
Azure AI Foundry provides an intuitive low-code environment for AI components which are part of the Azure AI Services (formerly Azure Cognitive Services) and Azure OpenAI resources. These resources include implementations for a variety of cloud-based AI tools in a “ready-to-use-capacity”, that allow you to start without building the functionalities from scratch. What’s even better is that virtually every page and wizard in this portal is accompanied by links to Microsoft documentation which you can follow:
- Find step-by-step guides on how to deploy individual components supported by the portal;
- Learn more about the components you are working on, and how to extend them further and refine their performance.
On top of providing you with an environment where you can both learn and deploy functional AI components, Azure AI Foundry integrates with other Microsoft Products relevant to the AI suite, including Azure Machine Learning.
Now, just to be realistic and reel in expectations just a little bit… when you get onto full-blown production-ready deployments, you will need to switch over to interacting with the underlying APIs with your own code. Moreover, Azure AI Foundry is built with Microsoft’s AI services in mind, however, when you transition to code level, the majority of components and the SDKs offered by Microsoft themselves have significant overlap with community-favourite & industry-standard AI SDKs.
4. How do we optimise our AI models?
Immediately after deploying your first AI component, you will probably say “This would be really cool if I can somehow get it to reference my data”, and you would be right – because it adds value that is specific to you (or more likely, your business case). These types of optimisations can be achieved using a number of techniques, and the good folks at OpenAI neatly summarise this into two dimensions:
- Context optimisation – Efforts that address what the model needs to know, in other words: content or data it can refer to be able to answer a question.
- Model optimisation – Efforts that address how the model needs to respond innately to the business domain it is operating on.
You can research best practices on prompt engineering and fine-tuning for days (in fact – they will be the subject of discussion another day) because these are very rich areas. The same can be said for Retrieval Augmented Generation (RAG), however, I learned the long way that I should definitely have started experimenting with RAG first – an idea that I would later come to learn recommended by Microsoft themselves recommend.
5. How do we go about this in a safe and responsible manner?
AI tools are incredibly powerful and offer unparalleled convenience compared to most software solutions. As per the old adage “With great power, comes great responsibility”. As solution providers, we must ensure that we deliver a system that takes the safety and consideration of the user and the resources it is consuming.
Safety for our users
Two words: Content Safety… this refers to practices (as well as actual utilities that can be embedded within the AI Solution, like Azure AI Content Safety) that help ensure that generated content is not harmful, offensive, or misleading content.
Areas addressed by content safety efforts include:
- Toxicity and Harm Prevention: Filters content for hate speech, violence, and self-harm risks.
- Bias Mitigation: Reduces the risk of reinforcing stereotypes or unfair outcomes (for further reference of how this fits in an AI adoption strategy, see the Google AI Principles spec).
- Mitigating IP Infringement in Generated Content: While not a complete solution, content safety tools can detect and block clear misuse (e.g., copyrighted logos or well-known lyrics) and flag content that resembles copyrighted material. For full IP compliance, additional review and tooling are required. (see Protected material detection in Azure AI Content Safety – Azure AI services | Microsoft Learn)
Safety for our resources
When we connect AI solutions to our resources using RAG, we significantly amplify how these tools can help address our business case because they become context-specific. While it is virtually essential to proceed with optimising an AI solution’s context, we need to do so in a careful manner.
We don’t necessarily have to look too far out from traditional methods for protecting business resources, as typical efforts come in handy here:
- Data Source Level Protection
- Access Control – Restrict data access using RBAC
- Encryption – Encrypt data at rest (stored) and in transit (transferred) with the help of utilities like AKV.
- Data Masking – Hide sensitive data while allowing AI to process it.
- Authorized Request Protection
- Authentication & Authorization – Ensure AI systems authenticate via OAuth, API keys, or Managed Identities.
- Zero Trust Security – Verify every access request before granting permissions.
- Rate Limiting & API Security – Prevent unauthorized or excessive requests by using Azure API Management.
- Tools Like Microsoft Purview
- Data Governance – Track and classify sensitive data to ensure compliance.
- Audit & Monitor – Log access and changes to AI-related data.
- Compliance Standards – Helps enforce GDPR, HIPAA, and other regulatory requirements.
Closing words
To close things off, working with AI is less about chasing the latest hype and more about understanding how the pieces fit together, technically, ethically, and practically. You don’t need to be a data scientist to start experimenting, but you do need a clear sense of what makes AI useful, what makes it safe, and where it fits in your existing solutions.
Stay tuned to this place for more blog posts on AI and Software Development in general. If you or your business requires technical consultation or services in implementing such services, reach out to us!