Gain from Generative AI: Effective Prompting for Accountants

Artificial Intelligence (AI) continues to draw attention among trusted advisors, changing how we foresee our means to explain, interpret, and predict financial and other information.

The subset of AI currently making the most waves is generative AI, which leverages machine learning techniques to create content such as text, images, music, or even complex models from patterns in input data.

When paired with highly precise instructions, or semantic parameters, generative AI may produce more valuable outputs.

Choosing Generative AI Platforms

Choosing the right generative AI platforms for your needs can be like selecting the right artist to commission a painting. It requires a clear understanding of your requirements and careful evaluation of the resource's capabilities.

If you were commissioning an artist for a painting, you would consider the artist's style, technique, portfolio, and their ability to interpret and create based on your requirements. In much the same way, when selecting a generative AI platform, you may consider factors such as the model's capabilities, its training data, its flexibility to handle different prompts, and its ability to generate outputs that align with your specific needs.

The “style” of an AI model is dictated by its training data. Much like an artist studying under a certain mentor or being influenced by a particular art movement will reflect those styles in their work, a generative AI model will generate outputs that reflect the patterns and relationships it has learned from its training data as well as tuning and optimization that was applied to the base model.

Natural Language Programming and Prompting

Natural Language Programming (NLP) is a field of AI that gives machines the ability to understand, respond to, and generate human language. It is a bridge between humans and AI, enabling professionals to communicate with AI models in natural language, as if we are talking to another human.

In the context of generative AI, NLP can be used by an AI platform to interpret prompts (requests or instructions) and possibly to structure those prompts with detailed parameters. This structuring may allow a user to guide the AI more effectively, thereby enhancing the quality and relevance of the output.

Using Semantic Parameters

A generative AI platform's effectiveness usually depends on prompts used to guide the AI in generating relevant and accurate output. With some platforms, semantic parameters—specific instructions written using plain English—may help increase that relevancy and accuracy.

Semantic parameters are specific inputs included in a prompt that provide guidance to an AI model using terms that are meaningful and understandable to human users. They can help steer the output of a generative AI to align with certain topics, styles, tones, or other characteristics.

As an example scenario, when instructing a generative AI to draft various types of business emails, a user might include semantic parameters to indicate the type of email, the tone, and the desired level of detail. In the hypothetical prompts below, an advisor structures a prompt using the parameters “email_type”, “tone”, and “detail_level”.

The email_type parameter guides the AI in emphasizing that this is an update, perhaps with an introductory sentence or two to that effect. The tone parameter ensures that the language used is professional and appropriate for a client communication. The detail_level parameter instructs the AI to include a high level of detail, important in a tax provision update where specific estimates, explanations or timelines may be appropriate. Those details could be included in a details parameter.

Here, the email_type parameter steers the AI to draft an email with the structure of an intercompany request. The tone parameter sets a collaborative tone, suggesting collegial terminology or emphasizing teamwork (as opposed to phrasing the request using an authoritative or directive tone). The detail_level parameter indicates that the email should provide enough information to clearly elaborate where necessary, but doesn't need to go into high detail about the background or reason for the request.

In this case, the email_type parameter instructs the AI to create an email that summarizes information rather than going in-depth. The tone parameter instructs the AI to use language suitable for an executive audience. The detail_level parameter indicates that the email should provide a high-level summarization rather than a deep dive into the factors of causes of the numbers. A details parameter could be used to provide the free cash flow and KPIs data.

The availability and effectiveness of semantic parameters would depend on the capabilities of the specific AI platform and model being used. Some AI platforms may not support this level of parameterization, or may interpret the parameters differently. Many AI platforms provide specific documentation with instructions on using parameters.

Aiding Data Visualization with AI

Generative AI also has the potential to expand the way we visualize data in business applications. Today advisors can generate graphs, charts, or even more complex visual representations just by describing the request using plain English.

An example of this capability is seen in Power BI's Q&A feature. Analysts can ask Power BI a question in plain English, like "What were the total sales for the last quarter?", and the program will not only understand the question and retrieve the answer. This feature uses NLP to interpret the question and generative AI to produce the answer, thereby saving time and potentially reducing errors.

Conclusion

To get the most out of generative AI and semantic parameters, here are a few tips:

1. Identify which platforms can effectively interpret semantic parameters, such as ChatGPT and Midjourney.

2. Be clear and specific when structuring your prompts and parameters. The more specific you are, the more accurate the output is likely to be.

3. Experiment with different parameters and evaluate the output. This process may involve some trial and error, but the impact may be significant.

4. Don't shy away from seeking help. Products with features like Power BI’s Q&A often have resources available online to help learn these skills.

The integration of AI and NLP in accounting is advancing quickly. Generative AI platforms and features like Power BI Q&A are already making it easier to harness the power of these technologies.