While it isn’t necessary to understand the intricacies of every generative AI model, it is important to broadly understand what the technology does, and how it generates outputs. Reading articles, watching videos, and attending webinars, courses or workshops can all help. 

Clarify your short and long-term business objectives and have a clear idea of how generative AI can help in achieving them. Consider unforeseen consequences in your evaluation of long-term consequences. For example, if the aim is to reduce junior staff costs, consider the long-term implications of not having junior staff being trained up to senior roles. Weigh up the risks and rewards and ensure you clearly define any expected benefits and KPIs to measure success.

As with any technology, generative AI should not be adopted for the sake of it. Identify where the opportunities lie, and focus on those. The key consideration when determining tasks to be completed by generative AI is to remember that the output of generative AI is not always trustworthy. Therefore, the best areas to start with are those that require little or no judgement, are low risk, time consuming and repetitive. Whether a level of human review is included will depend on the level of risk and volume of activity.

Generative AI is very well suited to tasks that follow a specific template and well-defined rules – for example, assigning accounts to categories. However, it is different from automation as it not only performs automated tasks, but can also augment the output with additional information eg summaries and highlighting of important points. Repeatable and scalable tasks will be more worth the investment in setup time, cost and forward planning to scale if initial use cases are successful. Generative AI is also good at providing answers to questions that can be verbalised. Effective prompt engineering is crucial to getting the right output and is covered in the “Using generative AI” section.  

You should try out tools in a range of scenarios, before wholeheartedly committing. All the major cloud providers (Google, Amazon, Microsoft) offer scalable resources and user-friendly interfaces and tools to help people develop and deploy generative AI models in their own environments. Pre-built models and services can be used with minimal coding or technical knowledge. Some are pre-trained for tasks such as anomaly detection, text generation and image synthesis, simply requiring a connection to the user’s data repository. Application programming interfaces (APIs) can also be leveraged to integrate generative AI capabilities with other software applications. 

There are some important distinctions in the types of AI models:

• Most generative AI models are ‘foundation models’, trained on a vast quantity of data and designed to be adaptable to a variety of tasks. They offer broad potential, but may be limited in terms of context.
• The alternative approach is a ‘narrower’ AI solution, which is trained on specific data for specific purposes. This would require more setup time and cost and may be limited in how users can interact with it, but can be much more contextually aware and provide more focused responses.
• A hybrid approach, where the foundational model is used to deliver interpretation and generation of text, voice, image or other mediums, in combination with a tailored model trained on business-specific information, can deliver huge benefits. This is where tools such as such as Microsoft 365 Copilot come into play. Copilot draws on OpenAI’s GPT foundation model, and overlays this with the context of organisational data, to assist with productivity, creativity, analytics and collaboration (for example, being able to take a detailed business case Word document and turn it into a PowerPoint presentation, or produce a budget from a few bullet points in an email).

Implementing generative AI can have a significant impact on an organisation. It is important for accountants to engage the right level of expertise early in the process, to get a good understanding of the opportunities, risks and impact on the organisation. Talk to legal, data protection, HR, ethics, and other experts within your organisation and get external expertise as necessary.  

There are lots of specialist providers, resources, and experts around that can assist non-tech organisations to explore, experiment, build and deploy Generative AI applications, and benefit from insights and industry best practices. Be aware of the available options before jumping to a particular solution and know that you don’t have to go it alone. Don’t be afraid to ask probing questions of potential suppliers around the training data used (if any), model operation, accountability, bias, privacy and any third-party assessments of their solutions. Establishing an ongoing vendor oversight process is also important to ensure that vendor supplied products and services continue to have the right governance and controls.

Whether you will use the public (open) version of a tool, or a segregated private version installed on your own environment, is an important choice, and one that may be dictated by the need to feed the model with potentially confidential data. In most cases, organisations will start with exploring generative AI tools in public environments, before transitioning to a private setup – this gives a good sense of the capability of the model, but not how it will behave on an internal platform. It should be tested in a separate non-production environment wherever possible, or at the very least loaded with sanitised data.

As with any new technology, it is important to treat the implementation of generative AI as a change project that requires careful management of the people and processes likely to be impacted. It should never be assumed that people will readily embrace new technology. Indeed, many will have valid concerns for the future of their roles, and others may be sceptical of the quality of the results, so these need to be addressed early.

Perfect data is a noble aim but rarely achievable in most organisations. Generative AI tools are generally good at taking in unstructured information, but feeding them with poor quality, inaccurate, or badly managed data will likely end in failure. Before training a generative AI tool on your own data, consider:

• cleaning and pre-processing data to deliver standardisation and improved quality;
• implementing a data management strategy including use of metadata (i.e. cataloguing/tagging of data);
• establishing appropriate retention policies to ensure that generative AI utilises only relevant, current information;
• integrating and consolidating data into a unified format;
• reviewing data to ensure that, before it is used, it is factually correct and free from bias; 
• adopting robust data governance practices, including consideration of data ethics, privacy and protection; and
• reflecting on data storage arrangements, considering that cloud-based solutions are likely to be far more effective when looking to utilise generative AI technology due to the computing resources required.

To use generative AI well requires a certain level of knowledge and as the technology evolves, there is a need to maintain and update that knowledge. The big generative AI technology providers are investing heavily in supporting beginners with free training, as well as documentation, tutorials and open source examples. More advanced users can take courses and gain certifications. Users should also keep up to date on the risks and regulations.

A responsible approach is important whether you are building and/or training generative AI models or using generative AI capabilities that are incorporated into third-party products and services. Subsequent sections of this guide explore some of the key risks and governance principles. 

• aistandardshub.org