Beginning Your Innovation Journey With AI

Advances in artificial intelligence dominate headlines in the technology sector, with many heralding the coming of a new industrial revolution that will transform productivity and efficiency.

Despite the many benefits of the technology, navigating a successful path with AI innovation can be challenging.

In this article, we’ll look at some of the potential pitfalls and challenges when planning a successful AI project.

AI projects often fail to have their intended impact due to a lack of clearly defined objectives. Gartner predicted that 85% of AI projects over 2022 would deliver erroneous outcomes due to bias in data, algorithms or the teams responsible for managing them.

Setting yourself up for AI success requires clearly defined KPIs and outcomes that you can measure with metrics to determine success.Examples couldinclude increased customer conversions, mission accomplished analytics goals,increased customer referrals or reduced customer service interventions.

2. Start small

Unless you are an enterprise organisation or software business, you’re probably not going to be building your own in-house AI team. Similarly, when defining your AI use case it’s best to start small in scope. Good examples could include:

•  Improvements to business processes that save time or reduce overheads
• Creating enhanced customer satisfaction and product/brand loyalty
• Better conversions when signposting customers to relevant products and services.

AI systems often have limited understanding of context in language versus humans, so consider how you could provide context to recommendations. While Large Language Model (LLM) AIs can interpret a user’s requirements or goals from their chat requests, consider how adding more context in your training data could deliver better results.

When working with LLMs, this can be achieved in a number of ways:

Prompt Engineering – this is the process of crafting clear instructions using specific keywords and phrases to fine-tune the AI’s response.

Adding context to instructions – this could be background information that helps explain your request, giving the model a better understanding of your needs.

Providing the AI model with examples – providing the AI with examples of your intended output can help to refine responses during the training process.

Bias can also be a risk factor in AI, so it’s best to implement restrictions on what topics your AI can and cannot engage with, if your solution is customer facing. A human in the loop during testing and after rollout is essential.

3. Ensure data quality

The quality of your data dictates whether your AI solution is viable or not. Consider for example using an AI model to identify trends in climate change. If the data is not cleansed to account for missing values over years or underlying inconsistencies, your results will be skewed.

Consider creating a yardstick for data quality and scoring all data input, to ensure good results.

4. Understand AI’s limitations

Let’s take the example of applying AI to recommendations in a video-streaming application such as Netflix or YouTube.

AI could be readily employed as an alternative to algorithms to serve the user content recommendations based on their past viewing history and similar content.

However, this can result in a self-fulfilling prophesy as the user consumes similar video within a bubble, unaware of the content that sits outside. Much like trending or popular articles on a website feed, without a degree of editorial curation AI recommendations can trap your audience in a cul-de-sac of similar content.

That’s fine for engagement when you’re capturing impressions for advertising revenue, but not so great when using AI to signpost training materials or marketing resources.

5. Harness existing AI models

Identifying the right AI platform or model to use can feel daunting, with a plethora of choices including Google AI, AWS, Microsoft Azure AI, Bot Framework and Copilot, and Open AI ChatGPT. Each offer different benefits depending on your chosen tech stack and potential integration points with other platforms.

Training an AI model from scratch also represents a significant time and cost investment.

A less costly and time-consuming approach to training your own model could be leveraging a pre-existing one. AWS offers a marketplace of pretrained AI models for Amazon SageMaker for tasks like object recognition and classification and natural language processing.

Harnessing an existing model can accelerate your project and reduce your time to market. Similarly, Google has a range of developer-friendly AI services and Microsoft offers ready-made AI services and models as part of its tool chain.

6. Invest in AI skills for your staff

The biggest barrier to AI success is likely to be a lack of staff skills. While not everyone is going to contribute to an AI project, data and AI literacy within your teams will be important success factors.

Awareness of foundational concepts such as prompt engineering and data quality can enable more effective communication and collaboration over AI initiatives.

Also consider workspace tools that have AI baked-in, such as the benefits of Microsoft Copilot which works across Microsoft 365 and Microsoft Teams.

Developing a dedicated Microsoft Teams app could enable access to the Azure bot framework, which can be used to automate manual tasks, signpost resources and enhance workflows. Consider the value of an AI assistant that can access Microsoft Active Directory to understand who your staff member is, their job role and what they need to be successful.

7. Be economic with AI in your tool chain

Be strategic about where you employ AI in your solution. AI is frequently used as a search tool for the signposting of content, and it can be tempting to point a bot at your site to index everything and have it take the load off your busy customer services teams.

This can help, but might result in higher operational costs, as your solution scales with your audience demand.

A better solution might be to break down the problem, using traditional code to solve some of the challenges. For example, you could implement a scraping tool to help manage the indexing of your site content.

You could also look at a software tool to help you manage and tag all uploaded content, reducing the need to endlessly re-index your site. You could audit your site to ensure alt tagging for imagery can be understood by your bot. Again, consider the value of giving your AI proper examples of your content to help ensure better results for your audience.

Consider also that if you are using a single LLM in its entirety to solve the problem in its entirety, your AI solution will be similar to your competitors following a similar approach. Consider investments in tooling, processes and refinements that will give your AI solution the edge.