Getting Business Value from AI: Map the Solution to the Business.
We now turn our attention to the process of understanding practically how the solution will fit in with – and usefully provide value to – the processes and operations of the business. Although it sounds relatively straightforward, it is vitally important in order for the solution to be adopted within the business. Machine learning (ML) models can be trained to answer specific questions or generative AI (Gen AI) models can respond to (curated) prompts, however this input/output interaction with a model may not map directly or neatly on to the question emerging from the business requirements. Therefore, how would an AI expert approach such a business-driven project? To consider this, for the purpose of this article, we consider a typical situation. For example, it might be that the senior stakeholder has the ambitious goal of using an AI, or ML, solution to automate marketing.
In such a case it is likely that multiple models and components need to be combined to arrive at a useful solution for the user. In isolation, the problem (and client expectations) may not neatly fit into the ‘data in/inference out’ format of traditional AI or the ‘prompt/response’ of Gen AI. Therefore, understanding of the business problem here is critical, breaking down the gap (which is currently a mysterious black box) into something that can be understood and implemented.
There is no single solution for this, and each case is different, but a useful approach for AI projects in general might be:
- Understand and document the functionality of the desired tool or system and any user interface.
- Document the data available.
- Break down into elements that can be solved by AI models (ML, natural language processing, computer vision or generative AI).
- For the remaining gaps, can multiple approaches be combined to solve it?
- Can rules be created that would address this successfully?
- If a gap still remains, can this be filled by a human who would perform a task within the solution, perhaps assisted by the options from the AI system? To deal with such a gap, the ideal situation would be to be able to learn/codify the behaviours of the human who is integrated into the system (a similar approach can be used to augment a poorly performing model).
Throughout the journey of business logic mapping, we need to be aware of the value that different elements bring, and how much money and time is available. It is usually worthwhile to create a value map: what are the tasks that are valuable to the business and hard to automate versus those which are time-consuming and could be easily automated using AI approaches.
Using the example of an AI solution to automate marketing, after selecting valuable use cases, we have identified that a model could be built to predict customer churn. This fits firmly into traditional AI (ML), but the same approach could be applied if the technology happened to be generative AI. Leaving aside the details of building such a model, it appears to work well, in that it can successfully predict if a customer is likely to churn (even evaluating a model’s performance depends on the business application, and this should be agreed in advance with the relevant stakeholders). Assuming that the engineering considerations have been successfully addressed, such as integrating it with the data and existing infrastructure (the effort required should never be underestimated in a project), we still need to ensure that such a model really adds value to the business.
Going back to the theme of our previous article, we also want to ensure that we do not work in a vacuum. Therefore, to ensure that the end result delivers the impact desired and anticipated, rather than being seen as a solution looking for a problem, it is necessary to continue to work with the senior stakeholders and end users to make sure that this actually happens. In this case, after working with the relevant stakeholders, the way to bring value to the business might be to make the outputs of the model available to marketers within the company so that they can specifically target individuals who are likely to churn.
One key consideration for users interacting with the output from ML models is to ensure that it is useful to them. For example, it is likely to be integrated into some sort of marketing UI, but then what would be a useful presentation of the data? Probabilities from 0.0-1.0 are unlikely to make sense and be valuable by themselves, but how about high/medium/low, or even top N or top N% most likely to churn. Here, as you can see, decisions need to be taken throughout the process, in close collaboration with relevant stakeholders, to ensure that the project achieves a useful result.
Surfacing the model output to marketers in a way that is useful to them would give us a way to demonstrate the value of the model and to get buy in from the user (which could then lead to further automation or building it into an end-to-end solution). This neatly illustrates the twofold benefit of working closely with users and senior stakeholders to not only validate the value of the solution, but also to serve as a checkpoint which can lead to increased project scope or greater investment.
Using the example of a marketing department wanting to incorporate AI to automate its processes, we have explored how an AI – in this case ML – model can be considered for incorporation into the existing business processes. Specifically, so it will be useful to, and usable by, the end users, and ultimately so that it can bring business value. In the next, and final, article we consider some of the practicalities of taking our AI project from a design into a working solution.
