AI in HR: How to write a prompt for artificial intelligence so that it immediately gives you the desired result?

You are the only real intelligence in working with AI

Oleksandr Lozhkin

HRIS Inplementation Consultant, SMART business.

Probably, almost every specialist in any field has recently had the question: “What if artificial intelligence takes my job?”. Indeed, AI is now actively integrating into all areas of our lives – both private and work ones. This process can be unusual, changes can be frightening and seem unnecessary. However, the best strategy is to make AI a tool for implementing your ideas and use progress to achieve the success of your company. In fact, this is easier to do than it may seem at first glance.

First, you should understand that artificial intelligence is a human assistant, not a competitor. It provides opportunities for development and creativity, taking over routine tasks. For example, in the HR sphere, AI proves its effectiveness in automating technical processes. Artificial intelligence can already review multiple resumes in a short time and select the most relevant candidates for further interviews, write and send business emails to potential employees (with invitations to interviews or, conversely, polite refusals). AI “skills” also include searching for specific data in documentation upon request and systematizing new information. Automation of administrative work and optimization of daily tasks leaves more time for HR managers to focus on their main area of ​​work – working with people.

AI in HR is comparable to a detective assistant in a classic investigative series. It takes on organizational issues and documenting leads so that the detective can fully focus on solving the case: in the case of an HR manager, this is talent search and strategy development.

However, Generative AI is not yet an ideal robot employee. It needs to be given clear tasks, otherwise it will be as useless for the company as a laptop without an operating system.

How to “communicate” with artificial intelligence correctly? Let’s look at the basic rules for writing prompts and learn a few life hacks that make the work of HR specialists with AI easier.

“Who am I?”, or How to write prompts for AI so that it understands exactly what you want from it

Imagine that artificial intelligence is a very intelligent intern with a ton of theoretical knowledge who has just joined your company. Their knowledge is a powerful tool, but you need to know how to use it. Your main task in working with artificial intelligence is to anticipate clarifying questions that it could ask. Let’s go back to the comparison with the intern: imagine you tell them “Bring some water, please”. Most likely, the intern will have questions: “Water from a cooler or in a bottle? What temperature should the water be? Pour it into a glass or a paper cup?”. The only difference is that the intern will ask you these questions right away. But artificial intelligence will not. In all likelihood, it will “bring” you tap water, and it will be hard to blame it for this.

To get exactly what you want, you should master the skill of writing prompts. These are the tasks for AI, designed as small requests-scenarios. Let’s look into the basic rules for writing prompts:

• Define the role of AI: Immediately identify the role of artificial intelligence for your task. Indicate from the point of view of which specialist it should generate content. For example, “assign” your position to the AI.
• Set the context: Explain to artificial intelligence what information it will receive and how it relates to its further task. Describe the vision of your company, indicate technical details, etc.
• Clearly articulate goals and requests: When setting a task for artificial intelligence, keep the wording simple and clear. Do not be afraid to repeat keywords and describe tasks step by step.
• Provide instructions for processing and formatting information: Describe the format in which you want to see the generated result. Determine the amount of information needed, specify in what form it should be presented – a table, a numbered list, etc. Suggest sorting the result by the request match coefficient or in ascending order of a specific indicator.
• Ask questions iteratively: If after completing the prompt you still have not received the desired result, analyze what details the AI ​​was missing to complete the task. Clarify why it could not complete the request as you needed and provide it with additional information.

Let’s look at a practical example of a ready-made prompt written according to the above rules. This template can be used to work with a GPT, Gemini or Microsoft Copilot chat:

You are an HR manager in a business company with 25 years of experience (Defining the AI ​​role). Now you will receive three resumes from different candidates for the position of “Content Marketing Manager” (Providing context). Review all three resumes and determine the best candidate for the position, with the preference for the following skills: 2+ years of experience in a similar position; awareness of the modern market; the ability to write marketing articles and ads. (a clearly formulated request for AI). Display the results of your analysis in the form of a table with candidate match rates – from best to worst (Providing instructions on processing and formatting information).

Interacting with Generative AI may seem difficult and unclear at first, but in fact, the rules for working with it are quite simple: if you want to get a clear result, you must have a clear request. Let’s say you want to get a human-like text – tell your AI> assistant about it in the request. Artificial intelligence is primarily a mathematical code, so remember your school tasks and apply the “given-find-answer” scheme. And remember: the best way to master a new skill is to constantly apply it. Do not be afraid to “communicate” with artificial intelligence, use it first for small tasks, explore its patterns and answers. The more you interact with AI, the easier it will become to work with.

“AI biases”, or How to avoid getting into social trouble due to a prompt error

Sometimes artificial intelligence immediately tells you that you have asked a socially sensitive question, and sometimes it does not. It should be taken into account that AI does not have a sense of tact and does not always contain the latest information about recent events in its database.

We suggest analyzing the case of the Amazon – one of the largest online stores in the world. The company began using artificial intelligence to automatically select potential candidates back in 2014. The idea was truly progressive: the corporate GPT selected the top five resumes from a hundred, which were then passed on to the HR department to invite candidates for an interview. However, as a Reuters investigation showed, in 2018, the company had an unpleasant situation: HR managers realized that their AI recruitment tool was biased against women. Almost all the resumes selected for further consideration were from men. How did this happen?

Artificial intelligence, of course, cannot have any sexist biases. The situation with hiring exclusively men occurred because Amazon’s computer models were trained to screen candidates by observing patterns in resumes that have been sent to the company over the past 10 years. Most of them came from men, a result of former gender stereotypes in tech industries. The AI ​​did not update the data or monitor the current hiring situation, so it did not know that women had long ago won the fight to work on a par with men in such organizations. Amazon was accused of sexism and closed its previous hiring program, developing a new AI structure based on its mistakes. Currently, the company successfully uses AI to automate most of its processes. However, the 2018 case remained in the articles – simply because of a mistaken prompt.

So, to avoid such situations, when working with AI, the following rules are to be observed:

• Always check the social context of your request.
• If possible, have a person who is well versed in current social, cultural and political trends check the results of the request.
• After generating the result, ask the AI ​​to separately check the information from social networks on this issue and find bits that can cause an undesirable reaction from society.
• If you use AI for hiring, immediately indicate that you want to see gender equality and diversity when analyzing resumes.

However, do not let the situation with Amazon scare you. One unpleasant incident does not mean that artificial intelligence poses a threat to your company. Generative AI continuously evolves and learns, allowing it to more quickly assess current societal trends over time. When used correctly, AI can act as a compass to help you navigate social contexts and avoid controversial situations.

“We are all human” – AI can also hallucinate

The information that artificial intelligence provides you should be checked not only in a social context. Anyone who has ever used generative AI knows that sometimes it gives a very strange result that is completely unrelated to the request. It also often happens that it “comes up” with information about people, companies or phenomena that simply do not exist. Such cases are called “hallucinations” of Generative AI.

Why does this happen? There is no exact answer to this question yet, but we decided to ask artificial intelligence directly what it thinks about its own “hallucinations”. The AI ​​gave the following possible reasons for their occurrence:

1. Incredible generality: Large language models do not understand the world in the traditional sense but work on the principle of a possible selection of words and phrases. They do not have a genuine understanding of the context or factual knowledge, which can lead to false, but plausible statements.
2. No knowledge of the world outside of training: The AI ​​has no access to real-time or factual knowledge other than the one it was trained on. It can’t independently verify new facts or events, so its answers may be outdated or inaccurate.
3. Context and consistency issues: Generative AI may struggle with long or complex contexts. It may lose consistency or fail to understand the exact gist of the question, leading to logical errors or “hallucinations” in its answers.
4. Incorrect or incomplete data: If the model was exposed to incorrect information during training, it may inadvertently incorporate it into its answers. The model may not have a built-in mechanism for fact-checking or separating reliable sources from unreliable ones.

A clearly worded prompt will handle most of these issues. However, the last one, “Incorrect or incomplete data,” can really complicate things. Free versions of the AI ​​learn from publicly available data over a period of time. Therefore, it happens that at the time of the request they do not have relevant information focusing on outdated data, or simply generate completely new, incredible information, giving it a completely realistic appearance.

The best protection against “hallucinations” is a corporate AI assistant, more precisely collecting and using probable, specific and narrow-profile data precisely for your field of activity. Thanks to an isolated information environment and effective training, it significantly reduces the chance of obtaining an irrelevant result. SMART business, knowing the advantages of artificial intelligence, is now actively developing a universal corporate AI assistant.

Information security: it should be remembered that the AI ​​did not sign an NDA with you

It’s time to talk about security. If you ask the artificial intelligence the question: “Can you keep my next message completely confidential?”, it will answer the following: “Unfortunately, I am not able to guarantee complete confidentiality of messages, since all our interactions can be saved and processed to improve the system. If you have confidential information, it is better not to share it here”.

It is very important to check that you are not providing publicly available AI tools with sensitive data for your organization. Free AIs use your prompts to further generate results and immediately warn you that they are not responsible for your information. If sensitive data is leaked, unfortunately, you will have no one to blame but yourself.

To minimize such risks, follow these rules when using AI:

1. Ensure data input accuracy – carefully review the data provided to the AI.
2. Ensure data output accuracy – always edit and check all Gen AI responses.
3. Anonymize personal data – do not use sensitive personal information in your requests or disguise it by removing important numbers or using code names.
4. Be transparent – whenever possible, let others know when you are using AI in a conversation; always ask permission to provide AI with data obtained from other people in a conversation.
5. Follow the law – understand and comply with local data use laws.

However, there is an easier way to protect your company from the risks associated with sensitive data. This is done with personal Generative AI> tools – such as Micrsoft Copilot with an enterprise license. It ensures that user and company data is protected and does not leak outside the organization thanks to Entra ID (a data protection system that provides adaptive access and unified identity management). In addition, Microsoft will not have access to your information, and it will not be used to train models.

“Have we met before?” – AI’s short memory and how to improve it

Anyone who has ever worked with free AI knows what a “short memory” it has. It can forget the result generated two answers earlier or lose all your progress after a session refresh due to the set limits on the number of requests. This is extremely inconvenient for long-term work with artificial intelligence, when it is important for you that it generates results based on previously provided information (for example, compare three new resumes with three others that it analyzed earlier in the day). How to deal with this?

Naturally, one of the options for solving this problem is to write several prompt templates in a separate document. Returning to the example above, when we asked the AI ​​to select candidates based on their resumes, we can imagine that the prompt template would look something like this:

You are /position in a company with X years of experience/. Now you will receive /input data/ for the position /position in the company/. Review all /data/ and identify the best candidate for the position, with preference for the following skills: /list of skills/. Display the results of your analysis in the form of /information presentation form/ with /information presentation conditions/ — from /information sorting conditions/.

This is a good option, but you must admit that filling in information manually each time can take a lot of time. Microsoft Copilot is much more effective in this regard, because with its implementation, the company receives a personal AI assistant that remembers previous requests and saves the history of interactions. This approach helps users automate routine tasks, increase productivity and facilitate work with interdependent prompts, because you can set a task for AI using any template. For example, “Analyze these resumes according to the pattern you used in this message yesterday.”

According to a McKinsey survey for May 2024, 65% of respondents report that their organizations regularly use Generative AI, which is almost double the figure for a similar survey just ten months ago.

Automation of routine processes is an integral part of the future of business. It is worth looking to this future with inspiration and hope, because, no matter what, a person is irreplaceable in working with artificial intelligence.

To confirm our words, we offer another short text from AI, generated upon our request: “You are a marketer at an AI company. Describe in three lines the ideal day of an HR manager if he or she works using artificial intelligence to perform routine tasks”:

In the morning, AI has already sorted the resumes, so HR starts the day with interesting interviews with candidates.

Lunch is spent discussing the development of corporate culture with the department, while AI analyzes the results of the meetings.

By the evening, HR completes all tasks, relieved: ‘Finally, with AI, I can focus on people, not numbers!’”.

You described the task in detail, but AI still gives you a different result than you expected? Submit a request, and we will help you save time on creating prompts: