With any major new advances in technology there is the excitement of the new, the fear of missing out (FOMO), and the wrestling with reality.  We hope that the new capability will allow us to revolutionise our businesses and lead to the sunlit uplands of prosperity.  And it could do, if we ignore tech’s carnival barkers offering grandiose temptations and employ the common sense needed. 

When it comes to Artificial Intelligence it’s no different, we must work out if it represents a solution to a challenge we face, or rather is it a solution in search of a problem.  So how do we do that?  By going back to 1st principles. 

In some way all initiatives can be boiled down to some combination of the following:

Your task is to work out where your focus needs to be to deliver the optimal combination and therefore the most value for your business. 

Adopting any innovative tool or technology represents a substantial investment in time and effort so bear in mind that there is only so much return on investment that can be achieved for process control and cost reduction.  That said, if your business is a low margin, high volume then the value to your business of cost savings could be significant. 

For this you need a thorough understanding of your organisation, how it works, how it communicates and interacts with customers (external or internal).  Value Stream Mapping is a useful business tool that can help tease out potential opportunities for improvement whether those opportunities are for the implementation of AI or not. 

Value Stream Mapping provides an end-to-end view of your chosen function or process, specifically:    

This map then helps identify and quantify the tasks within a process that:

• add value 
• incur cost but are necessary 
• have no benefit but incur costs 

With this insight you can then identify the improvement opportunities, and more importantly the ones that will deliver the most value back to the business.  

The output also enables you to understand the implication of any change from a user adoption perspective, which when it comes to implementing a new solution / way of working will be vital. A clear understanding of the user impact will enable you to better manage the change, and therefore the chances of success.  

Having identified improvement opportunities, now you need to ask the question is AI the right solution?  Answer this question honestly. Can the challenge you need to meet be achieved more simply, at lower cost and more quickly using the technology you have today, or a potentially simpler solution?  

If you are trying to find a product by describing it, then this is a good use case for AI. Conversely, retrieval by product code is a traditional database retrieval operation where using AI would be overkill. 

Don’t try to solve a y = mx +c problem using AI (unless of course you need to extract the variables from a text block or an image). 

Is your challenge growing, staying the same or shrinking?  What is the consequence of failing to address the challenge.  We want AI to be a long-lived solution, constantly demonstrating value, rather than hastening a naturally evaporating issue. 

The idea of a chatbot works well when proposed on a PowerPoint slide.  However, doing a chatbot well is far from trivial.  If it cannot be done well then implementing it will likely do more harm than good.  We would get a higher return on investment by focussing on our business processes and seizing the opportunity presented by AI to follow a more radical evolutionary path.  This is where the highest value opportunity lies.  If we get our AI augmented processes correct, then a Chatbot should be relegated to just one of a far broader palette of possible interfaces to our business processes.   

One problem where Generative AI is particularly useful (with suitable guardrails) is where the answer is spread across multiple internal documents.  For a human to answer the question they would have to evaluate all the documents to find the information, evaluate its relevance and value and presented it back as a concise answer.  As a manual process this is time-consuming and expensive and requires experienced people to do well. 

Have my high value candidate for an AI initiative, now what? 

Now you need to evaluate the likelihood of success, and often this comes down to three things: 

1. Usability of the solution – will end users adopt it, does it make their lives easier? 
2. Data quality / condition – is the data in good enough shape, and is there enough of it to power your AI solution? 
3. Signal – are you able to generate meaningful outputs / insight from the data using the AI solution to justify productionising it? 

Usability

Think how your intended audience will use your AI product.  It may be so seamless as to be invisible such as improving an existing search facility.  However, some of us are unable to forget Microsoft Office Clippy digital assistant.  Conceptually brilliant but ticked all the boxes for an appalling user experience. 

Borrow from Apple design principles, especially those of simplicity and human centricity. 

If your AI product is visible to the person using it then they should feel bereft at its removal, not joyful. 

Understand the constraints that exist and will affect your AI initiative.  These may be any of the examples below or something else entirely.  

The constraints you face may be specific to your industry or organisation, then of course there is data quality. 

Signal 

Now comes the moment of truth, can you generate meaningful and reliable insight / output from your intended AI solution. Having confirmed that the condition of the data is ok, you now need to build a Minimum Viable Product of your AI solution and ensure that is can address the challenge you have i.e. can it predict better, recommend better, generate meaningfully etc. This is the moment when you evidence your theoretical business case through a working solution. 

When it comes to AI, building trust is important. This plays in tointo the user adoption challenge mentioned above. If people don’t trust something, then they tend not to use it. 

The real focus here is assessing the risk of AI giving the “wrong” answer. 

For a product recommendation engine, the “wrong” product recommendation may be serendipitous.  For business functions requiring factual advice a wrong answer can be disastrous. 

For GenAI, platforms like Databricks contain “Judges” which act as guardrails for output.  These allow you configure how GenAI behaves to suit your use case.  Think of it as a spectrum ranging from content collation and curation based on well-grounded facts at one end and content creation at the other.  Where on that spectrum does your use case lie? 

The Databricks judges do this by setting bounds for the GenAI “Fact” 

• The fact must be present in source material fed into the AI model. 
• It must occur in a certain number of documents 
• When a document is split into chunks the fact must occur in a certain number of chunks. 

Other tools exist to ensure that outputs do not do any of the following: - 

• Contain offensive language, symbols or emojis 
• Breach legal and regulatory mandates such as not exposing PII (Personally Identifiable Information) and not giving financial advice. 
• Recommend competitors and/or their products 

One of the criticisms of AI is its ‘black box’ nature which makes the application of a “Trust but Verify” principle challenging.  

The AI/BI Genie tool within the Databricks stack does allow us to “Trust but Verify”.  This tool allows someone to ask a question such as “What is the value of sales for clothing product lines broken down by quarter”.  In addition to the answer the tool will also reveal the SQL query it used to come up with the answer.  As SQL is the lingua franca of databases AI/BI Genie shows its working in terms that can be read (and verified) by a broad technical audience. 

In general trust is built through experience. We often refer to this as AI literacy. The more we use it the more we understand how to use it to ensure we get out of it what we need. The same goes for designing and building AI solutions. The pitfalls are discovered through experience but also are often consistent across similar solutions. So having team members or a partner with experience really accelerates things.  

What are your measures of success or failure?  Ultimately the benefit of any solution must be visible on the bottom line but is this as a direct or indirect such as using AI to support a push for brand awareness. 

How long will it take to determine whether those measures indicate one way or another? 

If your AI solution produces undesirable results, can you switch it off with minimal impact?  Is there a plan B?  Take the example of nervousness about the implementation of an AI based product recommendation engine.  If the engine gives nonsensical results or cannot cope with the traffic it receives then it could have a major negative impact on sales.  It would be wise to consider a few options 

• Activating a kill switch, fully reverting to the original mechanism 
• A graceful service degradation approach where an overloaded AI system offloads the excess load to the original mechanism 
• Rather than an all or nothing approach, selective routing of traffic to your AI solution stepwise starting with the area of highest confidence or lowest risk to upsell/cross sell. This is often referred to as a canary deployment. If the canary passes out, get out of the mine.  

• Work out what would deliver most value to your organisation 
• Make sure the right people are involved throughout 
• Don’t forget the “customer” 
• Be honest about whether an AI solution is appropriate to the problem you are trying to address. 
• Make sure the data you need to fuel your AI solution is of suitable quality and flows sufficiently well.