TheCorporateCounsel.net

If we are honest with ourselves, we recognize that drafting SEC-mandated disclosure for public companies rarely involves sitting down with the proverbial blank sheet of paper and writing out a complete CD&A, for example. Rather, as my old friend Marty Dunn used to say: “Life is a markup of last year.” If we were to draft disclosure from scratch, it might yield a more insightful end product, but doing so is typically cost prohibitive from a company’s perspective, both in terms of time spent by company employees and hours billed by outside counsel. But what if that drafting could be done by a computer, based on careful prompting by an employee or outside counsel, whereby the costs of producing the disclosure moves toward zero? Such is the world that we have now entered with the advent of highly sophisticated agentic artificial intelligence capabilities.

We have now entered the age of agentic AI systems, which utilize machine learning models that mimic human decision-making to solve problems and carry out specific tasks. These models have “agency’ because, once they are created and launched, they can act independently and with a purpose, without the need for human oversight. Based on this description, one can no doubt see how agentic AI models can be particularly helpful when drafting disclosure in a proxy statement on an annual report. Based on the specific instructions provided to the model by a human, the agent could be tasked with drafting some or all of the proxy statement after reviewing proxy statements filed by the company and selected peer companies, analyzing pertinent information regarding the proposals provided by the company, reviewing applicable SEC rules and interpretations, analyzing proxy voting guidelines of institutional investors and proxy advisory firms, and reviewing any other inputs that the agent is directed to consider. In many ways, the process that the AI agent follows is analogous to the process that a human drafter might undertake when tasked with updating or drafting a proxy statement, except that the AI agent has a much more capacity to review and synthesis large amounts of data in much shorter periods of time than could be done by its human counterpart. In the cost-conscious corporate world that we operate in today, one can see how this agentic AI approach represents a viable drafting alternative.

As we all know, AI is not without its flaws. We are all familiar with the “GIGO” principle, with those letters standing for “Garbage In, Garbage Out.” Coined in the early days of computer programming, the principle contemplates that the output from a computer system is entirely dependent on the quality of the inputs, such that flaws in those initial inputs will always lead to flawed outputs. This concept that has been a staple of the programming world equally applies in the realm of AI. While agentic AI models have capabilities for reinforcement learning and self-supervised learning to improve their knowledge and performance, they are generally only able to operate as effectively as their human users permit them to operate through the prompts that direct their actions. Further, a particularly vexing concern with the use of any AI model is the possibility of hallucinations. The term “hallucinations” in the context of AI refers to situations where an AI model produces inaccurate, false and/or misleading information for a variety of reasons. The problem of hallucinations is particularly evident in large language models, where the model will insert factually incorrect information in an effort to predict what should be said (or what the human user wants the model to say), particularly in a situation where the model faces some uncertainty as to the outcome.

If you are planning to jump into the deep end of the pool and ask your AI agent to draft your proxy statement for the 2027 proxy season, here are a few suggestions to consider:

1. The use of AI is often an iterative process. You cannot expect a sophisticated AI model to spit out a fully-formed proxy statement without first spending a lot of time teaching the model about proxy statements and the materials that it should draw on to draft that proxy statement. Veteran users of AI models often note that developing prompts requires a great deal of advanced work, drawing on sophisticated human knowledge and experience.

2. Companies should be careful to avoid the GIGO principle. AI-generated materials will not be useful unless sufficient information is provided to the AI-model, and, even then, the information and instructions provided to the AI model may not be sufficient to avoid hallucinations and other data quality problems with the AI output.

3. One of the often-cited ways for trying to limit the risk of hallucinations is through more effective prompts. By making these prompts less open-ended and more specific, a user can seek to better control the way in which the AI model carries out its task. Another approach would be to instruct the AI model to base its response of specific, verifiable sources, rather than giving the model free reign to go out and find sources from the vast expanse of the Internet. Yet another approach is supplying the relevant information directly to the model, rather than allowing the model to rely on its own data that is has used to train itself. In this instance, because the model’s response is limited to the retrieved data, it reduces the need to fill in missing pieces of the puzzle with false or misleading information. Finally, it may be possible to limit the risk of hallucinations by strictly limiting the format and structure of the information that the model provides based on the user’s prompts, thereby discouraging the model from trying to fill in blanks with made-up content.

4. The human touch remains essential (at least for now), especially when making sure that the output from the AI-model is “ready for prime time.” Any AI-drafted proxy statement or other shareholder communication requires very close review and verification by humans, so as to avoid embarrassing missteps or non-compliant materials. Human reviewers should always be focused on the tone of the document, the continuity and accuracy of the information presented, and the possibility of hallucinations.

5. When writing disclosure for analysis by AI models, the company should take steps to communicate all important information in a machine-readable format, while emphasizing consistency and clarity throughout the document. Terminology should also be standardized and the disclosure should be scrubbed for contradictory and duplicative statements. Companies should consider retaining their stylized proxy statement formats for those investors who are not utilizing AI to analyze the proxy statement, particularly in situations where a company has a large retail shareholder base.

– Dave Lynn