Artificial Intelligence: Key implementation considerations and potential pitfalls

Artificial Intelligence: Key implementation considerations and potential pitfalls
05/08/2020 MHC


In the previous article, Artificial Intelligence: Opportunities to enhance the customer’s onboarding experience, we covered how artificial intelligence (AI) provides opportunities to enhance the customer’s onboarding experience.

This article discusses some of the key areas for consideration, and some potential challenges, when planning and preparing for an AI enabled technology implementation.

When talking about AI, we refer to machine learning (ML). ML is a subset of the larger field of AI, that focuses on teaching computers how to learn without the need to be programmed for specific tasks. It is possible to create algorithms that learn from, and make predictions on, large data sets. The algorithms also adapt in response to new data and experiences, to improve efficacy over time.

Machine Learning implementation steps and challenges

The first step is to assess which business objectives or problem can be best achieved or solved by using a ML solution.  It is recommended to begin with a deep-dive analysis of the current business objectives or pain-points. This will highlight the desired outputs and drive out which AI features may be beneficial to enhancing the customer and user experience, and potentially to improve internal controls. One should avoid focusing on the AI functionality until the problem statement has been clearly defined.

Once the area of focus has been agreed, the qualities of the existing data needs to be understood; its accuracy, how it is collected, and whether there is benefit in acquiring additional external data, to augment the firm’s own.

Successful implementation of ML requires careful preparation of the data; cleansed, organised in chronological order, made consistent, labelled etc., before it can be of use. The availability and quality of data is a key dependency, as attempting to train a ML model with insufficient or inaccurate data will lead to poor results. It relies upon clear patterns for the model to explore when making a prediction, and avoidance of clear but accidental patterns, leading to the model learning biases. To reduce conscious and unconscious bias, the data to be used for training the model should be diverse and reflective of the population you are attempting to model.

An additional challenge with creating datasets for ML is data privacy and security. It is important to understand data privacy rules, as defined by regulations such as the European General Data Protection Regulation (GDPR) and understand any potential constraints to building the required data set.

Once the business requirement has been quantified, and the data challenges understood, the firm must decide whether to buy or build.  The in-house build option requires significantly greater investment in specialist people and IT infrastructure, but may be the right long-term strategy for the organisation. However, for many organisations the availability of an ‘off-the-shelf solution may be more appealing as a ‘quick-to-market’, lower cost and lower risk entry option into AI.  Buying-in a model may also reduce the risk of model explainability, as the Vendor would most likely have already addressed this risk. Model explainability can be summarised as, the understanding of why an AI model makes the decisions that it does.

Irrespective of the buy or build decision, the organisation will need to train the model with the prepared data.  It is important that this step is business-led, and includes data scientists and data engineers, or external Vendor representatives, working together.  Once the model has been trained and its performance deemed fit for purpose, the model can be integrated into the production environment.  However, as with a child that grows and develops, ML requires regular evaluation, to ensure it continues to perform as expected and identify when re-training is required.

Building and deploying ML is a complex process. Organisations new to ML, should carry out one or more proof of concept pilots before embarking on an enterprise-scale deployment.  Such pilots are particularly suited to initiatives that have high potential business value, or allow the organization to test different technologies at the same time.

In Conclusion

This article has highlighted some of the key steps and challenges when implementing a machine learning solution, together with some of the major differences in approach versus more traditional technology change programmes.

In addition to extensive business and financial crime expertise, MHC consultants have an AI adoption and implementation capability to support our clients.

If you would like to speak to us, to discuss how MHC can support you in the delivery of a new or an existing AI Programme, please contact us on