Using Generative AI in Business: Best Practices, Risks, and Limitations

This text was written by a human assisted by an genAI to standardize the tone and spelling of the content. All information has been validated by human experts in the field. The header image was generated by an genAI from the author's prompt.

How do I use a genAI?

Let's face it, a well-used genAI can save a lot of time. It's important to understand, however, that the user must already have a good understanding of the subject and be able to «challenge» the genAI's answer. Unlike AIs, genAI are much more likely to make mistakes, given their nature. Knowing your business and the expected output is an advantage that the best users will have. If it's not possible for the user to verify the generator's output, he or she may introduce a logic bug that will induce artificial hallucination.

Know what you're trying to achieve before questioning the genAI

This may seem counter-intuitive, but it's the key to success with these tools. It's not about knowing the exact answer (because then you probably wouldn't need to query a genAI), but at least knowing what it is. You should also know how to check the facts given by genAI. Some will give their sources. Take the trouble to check that these sources actually exist, as it can happen that the genAI invents sources for himself, and that the genAI has indeed summarized the source mentioned. If it doesn't give its sources, think about validating its answer by carrying out a classic search on the answer given. Don't rely on an AI to check an AI's answer. Learn to validate manually.

The best term we can use is «Assistant» or «Co-pilot». It's not a question of replacing humans, but of making them more efficient. Since genAI don't «think» in the strict sense, they need to be guided. Properly guided, they will deliver results that meet your expectations.

GenAI are generally better for checking than for creating

It's a paradoxical observation, but the strength of genAI is not in generating, but in verifying, validating and modifying. They are not ideal for creating. This is one of the reasons behind the debate on copyright and intellectual property rights. If you want to change the tone of an e-mail, the genAI are brilliant at it. You want to optimize code, you've come to the right place. If you want to write a newspaper article, expect the worst if all you had to start with was a blank page.

Users who already use genAI to speed up their work have found this to be true. The first draft of the (non-specialized) genAI will generally produce questionable output, but you ask it to correct the output it has generated itself, and it will already be much better.

Precision is better than vagueness

There's a scale for everything, but the more precise a prompt is, the more precise the answer will be. Providing examples, detailed specifications, giving the agent a role, specifying specific sources, all these tricks will enable you to have a response where the risk of hallucinations is greatly reduced, but also where the quality of the response received will be superior.

What are the risks?

Risks are often summed up in three points: leakage, privacy and intellectual property. But there are many more. I present here a slightly more elaborate list, which is by no means exhaustive. Its purpose is to raise awareness and suggest other avenues to explore.

Information leakage

Make sure you always check the conditions of use. Most AI tools use input data and generated output for training purposes. The training data then becomes part of the system and may form part of the responses that are generated by the genAI later on. Putting sensitive information, or giving access to sensitive data, to genAI is tantamount to having leaked this sensitive information to the public.

Privacy policy

There are two levels of privacy. Yours directly, and that of others, sometimes indirectly. Once again, check the terms of data use. Alexa, Gemini, Siri, they're all handy, but do they send your personal data and that of your contacts when you query it? Remember that you're responsible for the data you share directly, but also think about the fact that you may be sharing information from third parties and that these people haven't usually given their consent to share this data. I invite you to visit the Commission d'accès à l'information du Québec for more information.

Intellectual property

This can be divided into two sub-points. The first is the copyright of training sources. Billions of texts, images, sounds and videos. All this data, used «without permission», despite being in the public domain. As we mentioned earlier, genAI don't create, they modify. Everything that is generated is in fact the product of training on billions of existing examples. It's even possible to copy a style we like by adding «in the style of...» to the prompt. So the whole debate comes down to the fact that the machine can simulate an artist, a writer or whatever.

The other point is the intellectual property itself. Legally speaking, the prompt belongs to the person who wrote it, but genAI's output, at the time of writing this article, in the majority of current jurisdictions, is not protected by copyright. Not only that, but its output is generally used to self-train afterwards. So for the time being, the output generated by an AI alone is not protected by copyright. (However, if your artists alter this output by using it as «inspiration», that result will be). Make sure you understand these legal nuances before using AI-assisted material.

Overconfidence

I've said it many times in this series, but genAI are exceptional in their ability to produce content that is believable on the surface, but whose content is false, incomplete, or outright made-up. Always assume that the output of a generator is a canvas, but that you must add your color. In some cases, the answer may turn out to be correct, but assume that there is always an error. Stay critical and alert to what the genAI is producing and you should be fine.

Cognitive dependency and loss of in-house expertise

Another more subtle risk is cognitive dependency: when AI becomes the main source of reasoning, human expertise can erode if not actively nurtured. In other words, by using genAI to write, analyze, summarize and propose, users can gradually lose control over business reasoning, and even unwittingly develop a certain form of intellectual laziness where it becomes easier to make a query than to really analyze the situation. There is also a long-term risk of losing the ability to challenge the answer given by the genAI.

Best practices generally call for validating with more than one source. My personal recommendation: using more than one AI is good, but validating with expert human beings is even better. Another trick is to ask the genAI to challenge the user. (Beware, complacency bias could cause the AI to systematically generate contradictions to «challenge» ideas. Use with caution). The important thing is to challenge yourself and keep validating your output.

Implicit bias and normalization

GenAI have no malicious intent, in fact, they have no intent at all. However, they have been trained on a certain dataset. They have therefore «learned» to reproduce sometimes reductive models. They will normalize certain worldviews, favoring formulations that are more «present» in the media, and attenuating divergent or minority opinions. The result will be a loss of diversity. And as the genAI return to their own outlets to continue training, this further widens the gap between these different opinions. We saw an extreme case in 2016 with Microsoft Tay. By training on user data from the Twitter platform, Tay developed some... let's say disturbing behaviors. It's important to understand that on the internet, it's not always the truth that is the most present, especially on social networks. If in its training data, the AI sees ten thousand articles that advocate false propaganda against one hundred articles that denounce this lie to show the truth, unfortunately, the AI doesn't understand the context and doesn't criticize the text, it only learns to reproduce these texts, and will therefore learn a lie caused by too strong a presence of it in its training data. I'll say it again: make it a habit to ask the AI for its sources, especially when it comes to text generation. You can also ask it if it has any sources of opinion that differ from the suggested answer. Make up your own mind, and stay vigilant and critical.

Responsibility and traceability

Here's an example of what not to do: «Why did we choose this option? → Because the AI suggested it». Even if the decision turned out to be a good one, this is a very bad argument. There was a recent article where a person persisted in telling an underground parking employee that his vehicle could pass, because ChatGPT had confirmed this, even though the employee knew full well that this was not the case. Basing decisions on text generated «pseudo-randomly» by a machine that doesn't know the reality in which we evolve is always a bad decision, whether the answer is right or wrong. In an example like this, you have to ask yourself «who is responsible for the error?» and «can it be reproduced?». Not all AIs are able to provide clear traces of their reasoning process. For the first question, my personal answer is always the same: the person responsible for the error is the user who asked an AI for help without validating the answer generated for him. However, this is not an opinion that everyone shares, blaming the process rather than the individual. I'll let you make up your own mind about that. What is clear, however, is that it's not healthy to absolve ourselves of responsibility for our own mistakes just because «the AI told me so».

Out-of-context uses (drift)

We suggest a recipe, a prompt that works in a certain context, you start applying it, you forget its original context, you start using a prompt that is not adapted to the situation, you get a result where hallucinations and errors abound. The expression «knowing how to use the right tool at the right time» fits well in this context. You wouldn't use a frying pan to unscrew a screw... Nor would you use a drawing program to do your tax return. Prompts are the tools at our disposal to interact with a tool that is both practical and fragile, because it depends on thousands of parameters. Always keep in mind what you want to do. Don't blindly apply a «pre-made recipe». As mentioned above, not all AIs are created equal. Learn to use the right AI for its strengths in a given situation, rather than a generic AI for all your needs. There are AI «aggregators» for these use cases, such as Perplexity, Aymo AI, Elvex, Eleven Labs and others. Always check their terms of use, but these aggregators are veritable toolboxes, often offered at a lower price than paying for each tool individually.

Gap between in-house and the «rest of the world»

Unless you've trained an AI explicitly on your own data, the AI doesn't know your internal rules. This will be the case for all public AIs. They will rely on public data only, but these are unlikely to apply to your specific context. It could therefore make non-compliant recommendations, ignore certain constraints, cause inconsistencies, fail to meet certain standards, and so on.

Examples of correct use

As mentioned in the intro to all the articles in the series, «this text was written by a human assisted by a genAI». In fact, I used several agents. First, I wrote my first draft in Word, knowing that Copilot M365 would come in handy later on. Once the first draft was built, I used a first prompt using the PIS concept and my hallucination reduction trick to reframe the answers Copilot had to give me.

« You're an expert in artificial intelligence in the broadest sense. You're a very factual person who relies only on verified and reliable sources. You've been given the task of checking the content of these four blog articles to verify the accuracy of the facts mentioned. The articles must remain accessible to both those with some knowledge of the field and curious customers. A first draft has been submitted and we're asking for your feedback. This is an article that is intended to be light and easy to understand for internal colleagues, but which could also be read by customers. We're trying to inform, but in a light-hearted tone. We ask that you always remain factual and provide your sources. If you're more than 10% in the dark, simply say you don't know, giving sources where the information might be available so we can make up our own minds. »

This prompt is not a recipe to be followed, but illustrates the concepts mentioned.

Once the basics were in place, I would ask him about various aspects: the density of the text, cases I might have forgotten, rewording for certain phrases that were too technical, etc... As he replied with sources, I was able to add some of the links he gave me that I found relevant in the «always curious» section. I also asked him to check some of the facts in the document (although this is not Copilot's main strength).

Once the writing was complete, I did a first proofreading pass with Antidote, then had it proofread by people representative of my target audience, then added their suggestions to the document.

Then, after a final check by Copilot, I passed the torch to Gemini via NotebookLM. I gave him a similar mandate to Copilot, but this time asked him to check for any factual errors that might have crept in. He found errors that Antidote had missed, but above all certain facts that were not unanimously accepted by the scientific community. Validating with more than one AI reduces the risk of bias (a bit like asking several experts for their opinion before making an important decision).

Then I finished with a final proofreading by (human) colleagues and someone representative of my target audience. The article was ready to be sent for publication.

For the English translation, an automatic system also suggested a first draft, which was then proofread to validate the terminology, especially the technical side. Again, this is not a recipe to be followed, but rather a suggestion for responsible use to be adapted to your needs and situations.

So by using the tool responsibly, we've arrived at this series of articles which I hope will have helped you demystify AI and genAI.

In brief

GenAI are excellent tools for assisting us with repetitive everyday tasks, but they are not there to replace a human being. We must remain vigilant and aware of the risks. Be alert, and learn how to use these tools, which can become a way of improving your performance. They are very practical, but they do have their limits.