Can Boards Rely On AI Remuneration Advice?

With targeted instructions and defined source documentation, AI can rapidly extract, process and transform market data, producing insights at speed and scale. But can AI be trusted to consistently provide factual responses to remuneration questions?

This is not a hypothetical. Internal rewards staff are using AI (sometimes at home, not on their work computer) to short cut their way to the CEO’s or Remco chair’s query. Even some board directors are using AI to “remedy” a thorny executive pay issue. Some have cited back to us the answer that AI has generated with the source link being historical research undertaken by our firm a decade ago.

The short answer is no. Do not rely on it (even if the answer was based on our firm’s decades old research). At least, not yet.

One of AI’s biggest strengths is extracting and processing information at scale. Key remuneration details of company executives are often buried in 300-page annual reports. It might take an experienced and well trained analyst 15 minutes to just accurately discern one company’s CEO remuneration package. With AI, the same information can be surfaced in seconds through a simple query. More importantly, AI can repeat this across hundreds of reports, delivering insights at a speed and volume no human team could match. However, how reliable and valid is the data from AI?

Despite AI’s efficiency, for executive pay it is not accurate, no matter how advanced an AI model is. Our research based on real life executive pay issues managed by highly qualified and trained professionals, shows the data from AI could be incomplete, outdated, or biased due to unclear queries or query logic and a lack of contextual awareness. More importantly, the errors are sometimes present in ways that are not immediately obvious. For example, there is an ~12% difference between an STI that is 100% of fixed remuneration and 100% of base salary. Compound the error over multiple records and there is a material difference of some magnitude on total remuneration. This emphasises the importance of always having a human in the loop to ensure consistently high data accuracy, at least for now.

When operated without carefully defined instructions and based on AI built-in sources the opportunity for errors is boundless. Even when operated with carefully designed instructions and pointing to specific data sources, AI still exhibits the following common issues:

• Misinterpreting remuneration concepts, e.g. confusing elements such as fixed pay vs base salary.
• Fabricating information, e.g. returning “usual practice” when querying a specific company’s STI approach. Beware that AI may hallucinate this as fact.
• Overlooking nuance, e.g. returning broad categories of STI measures rather than individual KPIs within the category.
• Struggling to make logical connections beyond surface-level wording, e.g. if asked about “David Smith’s remuneration”, AI may overlook information reported under the title “CEO’s salary”, because it does not always deduce that “David Smith” and “the CEO” are the same person.
• Providing a different answer when asked the same question a second time.

Correcting these mistakes requires careful review by a human. This also needs scale, in terms of the number and frequency of reward professionals using AI and having a peer human check their AI work.

Taking a step back from data accuracy, without judgement skills, conclusions drawn by AI from data may appear reasonable but may ultimately conflict with a company’s strategy, culture, governance principles and stakeholder dynamics. AI does not know all of that yet. And if it does, the time taken to develop the query set by a human may be better applied directly by the human doing most of the work.

Even assuming the AI model is a secure repository for company information, we have seen cases where AI has provided an answer to a remuneration related question with absolute conviction, only to state the exact opposite when queried on the rationale and source data for an assertion. This lack of consistency does not inspire confidence.

AI cannot anticipate shareholder sentiment, boardroom sensitivities or the reputational risks of certain pay practices. In particular, decisions on how much weight to give diverse viewpoints is difficult for AI.

To summarise, rather than replacing the Remco chair or the company’s internal rewards team, our experience is AI is more likely to augment work. We know, as we use it every day. With utilisation of AI-powered tools, less time can be spent on collecting data and more time analysing it, developing and testing alternatives (sometimes using AI again) and advising on strategy and engagement (definitely not AI). At its core, effective executive reward is about trust, strategy, context, and judgement – qualities that remain deeply human. And one last note, it is suggested Remco Chairs seek assurance from the author of papers submitted to the committee that data, analyses, conclusions and recommendations had suitable controls to ensure integrity of this data, analysis, conclusions and recommendations. Better that than being embarrassed in front of an investor or proxy advisor when confronted with errors.