‘Claude included it, so I kept it in there,’ a colleague said recently when I asked about a slightly odd sentence in a paper they were working on. It wasn’t the first time (and won’t be the last) that I have seen someone hastily rely on an AI tool for an output, whether for ease or because they assume the AI model knows something they don’t.
I’ve written before about why that’s not a good idea. Using AI is not a shortcut, and its output should only ever be a starting point, regardless of where in the process you’re using it.
Give it all the information you’d have if you were doing this on your own
My colleague’s belief in Claude has stuck with me uncomfortably. So I asked Claude whether it is wise to use its output just because ‘Claude said so’. Claude was emphatic (in my reading of its tone): ‘Inclusion is not endorsement. I generate text that fits the shape of the request; I have no way of knowing whether a company has the transition plan, the board committee or the supplier audit programme I have just described. If most reports in the training data mention one, one will tend to appear.’
As more people experiment with AI as part of the corporate reporting process, whether for writing text or creating images, what should they do – and avoid – to get the best out of it without introducing risk?
Intention and interrogation
Start by being intentional about what you are setting out to achieve. When developing content for an annual report or putting together an investor presentation, AI can support you best when it knows what you are aiming for and you articulate that clearly upfront. That vision will also help you know when to stop once you’ve got the output you are looking for.
That initial vision, though, does not always mean you will end up exactly where you thought you would. There is a need here for critical thinking combined with curiosity.
You need to be able to defend every aspect of the output
Although you can delegate the basic drafting or image creation tasks, you cannot delegate the thinking or the assessment of whether it is any good. It also means you may end up somewhere you had not anticipated because the AI model has brought something up you had not initially considered, which may well be something that turns your original idea into something better.
The ability to interrogate (not merely question) an AI model as you seek an answer is essential. Being curious about how it reached its conclusions and developed its output can help you assess whether it is plausible – and even more importantly – accurate.
When you use AI to generate ideas about what to include and how to present it, you will be the one who needs to decide if those ideas are any good, and your experience and expertise will help you support that conclusion. The more capable and impressive AI models become, the easier it will be to trust them, and that means that your role as critic and assessor becomes even more essential.
Supervision and ownership
Knowing how to supervise an AI model can determine its usefulness (a process that is not so very different from supervising people). One aspect of this is becoming an expert in effective prompting because, like people, an AI model needs clear direction and enough information to be able to create or edit a draft.
Do not rely on the AI’s training or make it guess what is correct – you need to give it all the information you would have at your fingertips if you were doing this on your own. Also make sure you take the time to answer its questions and thoroughly interrogate its output.
Become an expert in effective prompting because, like people, an AI model needs clear direction
For the more complex and sensitive sections of your reporting, AI can require hours of interaction with you, and many iterations, to get to a finished product. As Claude said in response to my query, it does not ‘know’ the answer. But a person does.
Taking ownership of the process and the output is often overlooked but is arguably what makes the final product any good. Using an AI model to do part of the work does not remove responsibility for what ultimately is produced. You will need to defend every aspect of the output, and that means not saying, ‘Claude included it, so I kept it in there.’
If you develop these necessary skills, taking ownership should come naturally because you will have had a heavy hand in the output and can be proud of what has been drafted, drawn or presented.
Intertwined input
While the use of AI for creating corporate communications is growing, it is also becoming more common, and indeed expected, that people disclose their use of AI. But what it means to ‘use’ AI is increasingly blurred as the models become more capable of helping us, particularly with writing nearly anything we give it with incredible fluency, and creating diagrams that adeptly turn a thousand words into a picture.
The more we direct, supervise and shape what it has written, and let it inspire what we write, the less we are able to distinguish between (or explain) what part was the AI model’s contribution and what part was our own – each party’s contribution to the output becomes intertwined.
The real question, then, is: what was made possible by the combination of your ideas, your direction and your interactions with the model that would not have existed otherwise? And that might be impossible to explain adequately, if at all.
