Creating personas with AI: The check

 

Understanding target groups is the be-all and end-all of successful marketing, product development, sales, customer service and good employer branding. Long-lasting, satisfied customer and employment relationships can only succeed if you know who you are dealing with and what they want, need, feel, fear and what puts them off.

data-driven personas provide precisely these insights. However, you need exactly that to create them: data. You can collect it yourself, buy it, or you can order data-driven personas from qualified providers.

Think for yourself despite the revolution

In order to save work and/or costs, it is tempting these days to "prompt" personas - in other words, to ask an AI to create personas for your own company according to certain specifications. In fact, AI has recently revolutionized customer segmentation in marketing. However, you still have to think for yourself. Large language models (LLMs) such as ChatGPT can do a lot, but they are not the solution to all problems. If you want to get data- and therefore evidence-based personas, you have to feed them with data, which brings us back to purchasing or collecting the data, which doesn't save any work - anything else means: shit in, shit out. What's more, nobody knows exactly what actually happens inside the LLM with the data entered, i.e. how the AI processes and weights it. The result is therefore not transparent.

Let's take a closer look at the method and the advantages and disadvantages of AI-generated personas.

AI-generated buyer personas are not personas that Chat GPT & Co spit out, but those that are created with the help of AI: AI can help to analyze extensive data sets, for example online behavior, preferences and interaction histories. AI can also help to derive operationalized recommendations for action from data. AI can also be used, e.g. in the case of perplexity, to find data (representative studies, data from statistical offices, ...). AI can also create personas from data - but AI needs valid data and a clear sequence of prompts as well as human control that intervenes to ensure that the derivations are also tenable and counter-steer if necessary. AI can also help to make personas tangible - for example, by transforming them into chatbots that respond in the role of the respective persona.

AI is therefore a great thing - if you use it correctly: knowledge of the individual models (what can they do, what can't they do?), experience in prompt engineering and, above all, the availability of valid data are crucial.

Create personas correctly with AI: This is how it works

Without data, all you can do is guess - and that is fatal. So start by collecting and collating data from various sources, e.g. using social media monitoring tools, CRM data, publicly available or purchased studies or statistics. Online interactions, transaction histories and demographic information could round off the data package by way of example.

Before this data can be used, it must be carefully prepared. This includes identifying inconsistencies and outliers, removing duplicate entries and standardizing formats. Missing or irrelevant values should either be added or excluded so that a consistent and meaningful data set is created in the end. Normalization and scaling ensure that all variables are in a comparable relationship to each other - an important prerequisite if AI models are to deliver reliable results later on.

Machine Learning (Training)

Machine learning algorithms, such as clustering algorithms for segmentation, classification algorithms for pattern recognition and association rules for recognizing relationships are the next step. Train previously selected algorithms with your prepared data set so that they can learn and recognize patterns, correlations and dependencies within the data. This step is crucial to ensure that the algorithms can later process new, previously unknown data in a meaningful way. An algorithm can only recognize complex patterns well if relevant features are present and the complexity of the data set is sufficiently reduced.

Historical data can be used to predict the future behavior of customers or applicants (predictive modeling). Time series analysis, regression models or classification algorithms help to develop models that predict potential behavioral changes and thus enable proactive strategy adjustments. Use validation and cross-validation techniques to assess the predictive accuracy of the models.

Machine learning not only enables predictions and pattern recognition, but also the recognition of subtle and complex correlations that can easily "slip through the cracks" of human analysts. Understanding which influences cause the customer or applicant to react in this or that way is very important for a deeper understanding of target group motivation and can give your own company an enormous competitive advantage.

Persona creation and adaptation Once all the data is available, prepared and the algorithms have been trained, you can create detailed and differentiated personas. Here it is important that you have previously developed a framework that contains the relevant features of the persona. In short: that shows in detail what is important to you (e.g. media usage behavior) and leaves out what you may know but is not relevant (e.g. shoe size). As a final result, you want a persona that tells the story of your target group and reflects the behavior and underlying motivation of your target group, including preferences, weaknesses, desires and goals.

Once done does not mean "done forever" for personas, not even for AI-created ones. Machine learning algorithms must also continuously learn and refine themselves in order not to fall by the wayside. Such an iterative learning process involves constantly updating the models with new data, adjusting parameters and fine-tuning the algorithms. Adapt your personas to new data points with automated updates in real time. This ensures ongoing relevance and accuracy. This approach enables a dynamic and adaptive persona creation process that adapts to changing customer behavior, trends and changing or new markets.

To ensure that your personas accurately represent the target group and to avoid overfitting the training data, you can compare the personas with real customer data using cross-validation or holdout validation. Here it is important to test the effectiveness of the personas in different scenarios and situations.

Whether developed in-house or purchased: AI-generated personas are revolutionizing the understanding of job applicants or customers by creating dynamic, data-based profiles that evolve in real time. Through machine learning, they uncover patterns and correlations and can make predictions that remain undetected with traditional methods. They provide practical and accurate depictions of customer and applicant behavior, enabling companies to make more beneficial decisions.

However, creating personas with AI is much more complex than writing a few prompts - even if they are good ones.

What if we do try LLMs?

In view of this, it is understandably tempting to save yourself all the effort and try an LLM. We were curious ourselves and gave it a try. This is what happened to us:

Our prompt:

Create a complete candidate persona profile for IT employees in a utility company. The persona profile should include the following sections: Demographics, Political Attitude, a biography, a section on the persona's needs, vulnerabilities and emotions, interests, personality, communication strategies and ad types. In addition, media usage behavior, attitude towards innovation, attitude towards advertising, travel, nutrition and online shopping. I also need digital and analog advertising touchpoints. As this is a candidate persona, information on job search (where is the persona looking), approach, expectations of the job, expectations of the application process, pain points in the job and application, plus points for employers and things that would prevent the persona from applying or have already led to an application being abandoned are important. Please also include one section for each item mentioned.

 

The result:

Candidate Persona: "Martin Weber" - IT systems engineer in the utilities sector

🔹 Demographic data

• Age: 35 years
• Gender: male
• Place of residence: Medium-sized city in southern Germany (e.g. Ulm)
• Marital status: married, two children (6 and 9 years)
• Education: Bachelor/Master in computer science or business informatics
• Profession: IT systems engineer / IT security specialist at an energy supplier
• Income: 70,000 - 85,000 € gross per year

🔹 Political attitude

• Green-liberal tendency: Supports sustainability, energy transition and digitalization in the public sector
• Values data protection, civil rights and technological independence

🔹 Biography

Martin grew up in the suburbs, was tech-savvy from an early age and tinkered with computers as a teenager. After studying computer science, he initially worked for a software company before switching to the public/semi-public sector in order to do "something meaningful" - e.g. working on stable, sustainable supply systems. Today, he works for a municipal energy supplier on IT security projects and the digitalization of network infrastructure.

🔹 Needs

• Meaningful work with social added value
• Stable employment with a long-term perspective
• Modern technology and working environment
• Further development opportunities (specialist and management career)
• Flexible working hours and home office arrangements
• Transparent communication and appreciation within the team

🔹 Vulnerabilities & emotions

• Skeptical of bureaucracy and slow decision-making processes
• Frustrated when innovations fail due to hierarchies
• Sometimes feels "caught between two stools": wants to be innovative, but must ensure stable systems
• Fear of skills shortages and overwork

🔹 Interests

• IT security, cloud technologies, smart grids
• Sustainability & renewable energies
• Gaming, e-mobility, home automation (smart home)
• Cycling, technology podcasts, soccer with the kids

🔹 Personality (Big Five)

Feature	Characteristic
Openness	High (keen to experiment, interested in new technologies)
Conscientiousness	High (reliable, structured)
Extraversion	Medium (communicative at work, but not a "small talker")
Compatibility	High (team-oriented, helpful)
Neuroticism	Low (calm, analytical, objective)

 

Communication strategies

• Prefers clear, factual and transparent communication
• Responds well to honest, authentic employer messages
• Does not like marketing platitudes or exaggerated emotionalization
• Channel preference: E-mail, LinkedIn, specialist forums, personal discussions

📣 Display types that work

• Technology-oriented with a sense of purpose: "Design the IT that secures our energy supply."
• Structured, informative, fact-based - No colorful glossy campaigns
• With a real insight component: Interviews with colleagues, technology stacks, real projects

🧠n Media usage behavior

• Reads heise.de, Golem, t3n, ComputerBase every day
• LinkedIn active 2-3x per week, privately he uses YouTube and Reddit
• Hardly any TV, streaming (Netflix, Prime Video) in the evening with family
• Print: occasionally trade magazines or municipal utility news

⚙️ Attitude towards innovation

• Positive, as long as it brings practical benefits
• Considers "innovation for innovation's sake" a waste of time
• Supports digitalization when it simplifies processes

📺 A ttitude towards advertising

• Skeptical of traditional advertising
• Responds to authentic and informative content (e.g. case studies, employee videos)
• Does not like aggressive recruiting campaigns

✈️ Attitude towards travel

• Business trips only if necessary
• Private: prefers family trips within Europe, by car or train

🥗 Attitude towards nutrition

• Conscious, but not dogmatic
• Pay attention to quality and regionality
• Canteen food is okay as long as it's balanced

🛒 Attitude towards online shopping

• Regularly buys technology products online (Amazon, notebooksbilliger.de)
• Clothes in store, technology online
• Data protection is important to him

📍 Digital advertising touchpoints

• LinkedIn Sponsored Posts
• Google Ads for IT training or utility IT
• YouTube pre-roll for tech videos
• Job portals: StepStone, heise Jobs, Indeed, LinkedIn Jobs
• Specialist forums & IT security communities

📇 Analog advertising touchpoints

• Job fairs (e.g. Connecticum, bonding, regional career days)
• Trade journals (IT Administrator, LANline)
• Information events at suppliers
• Recommendations from colleagues (Employee Referral)

🔍 Job search

• Active on StepStone, heise Jobs, LinkedIn
• Observes municipal career portals
• Rarely uses recruitment agencies
• Likes to ask around about colleagues and acquaintances ("quiet" searching)

🗣️ Address

• Technically competent, but without exaggeration
• Credibility before coolness
• Keywords such as security, stability, social benefits, modern IT infrastructure work well

💼 Expectations of the job

• Secure jobpermanent
• Meaningful tasks with social added value
• Further training opportunitiesnew technologies
• Flexible working hours, home office, family friendliness

🧾 Expectations of the application process

• Short response times
• Transparent processno black box
• Clear contact persons
• Technical discussion with IT colleagues, not just HR
• Feedback - even if you cancel

⚠️ Job and application pain points

• Too much bureaucracy
• Slow decisions
• Unclear task description
• Lack of communication after receipt of application

✅ Plus points for employers

• Municipal or sustainable mandate
• Stability and modern technologies
• Transparent leadership and respectful culture
• Training budget and home office regulation

🚫 What stops him from applying

• Impersonal mass communication ("Dear applicant")
• Unclear or generic job advertisements
• Non-transparent salary structures
• Application platforms with a complicated upload process

⛔ What has already led to the abandonment of an application

• No feedback after weeks
• HR could not answer technical questions
• Bad reviews on Kununu
• Too many mandatory fields and password registration in the application portal

Sounds convincing, doesn't it? We thought so too. And that's exactly the problem. Because one day later, we entered the same prompt again. No, not with the competition, but again with Chat GPT. And the result was:

 

Candidate Persona: Tobias Keller - IT System Administrator at a municipal utility company

📊 Demographic data

• Name: Tobias Keller
• Age: 36 years
• Gender: Male
• Marital status: Married, 1 child (5 years)
• Place of residence: Medium-sized city in southern Germany
• Education: Bachelor in Computer Science
• Professional experience: 12 years in IT (focus on system administration, IT security)
• Industry sector: Energy supply / infrastructure
• Position: IT system administrator, team leader
• Salary: approx. 72,000 € gross per year

🗳️ Political attitude

• Rather green-socially minded
• Environmentally conscious, advocates digitization in public hands
• Critical of large corporations, in favor of data protection & open source

📖 Biography

Tobias grew up in a small town and trained as an IT specialist after leaving school, later completing a degree in IT while working. He has been working in IT for a municipal energy supplier for over a decade. He appreciates the security and predictability of his job, but often wishes for more modern structures and processes. In his spare time, he is involved in the local sports club and programs small open source projects.

💡 Needs, vulnerabilities & emotions

• Needs:
• Stability & security in the job
• Clear structures, but also development opportunities
• Compatibility of career and family
• Meaningful, sustainable work
• Weak points:
• Resistance to change with many years of IT work in the public sector
• Frustration with poor technical equipment or decision blockages
• Emotions:
• Frustration with legacy systems
• Proud of its contribution to public services
• Security gives him a good feeling, but occasionally boredom too

🎯 Interests

• IT security, automation, Linux
• Energy policy & sustainable technologies
• DIY, e-bike tours
• Podcasts & tech YouTube channels
• Open source projects

🧠 Personality

• Introverted to ambivert
• Pragmatic, detail-oriented
• Patient, but impatient with unnecessary bureaucracy
• Team player, likes to take on responsibility, but doesn't like being in the spotlight

📣 Communication strategies

• Tone of voice: Clear, direct, without any marketing buzz
• Channels: Email, LinkedIn, specialist forums, tech blogs
• Appreciation & transparency are central to communication
• Authenticity instead of glossy campaigns

📺 Display types

• Text ads with a clear job description
• Technically detailed Job offers (e.g. which systems are used)
• Video testimonials from employees with a similar profile
• Green IT & sustainability as a hook

📱 Media usage behavior

• Professional: Heise, Golem, Stack Overflow, LinkedIn
• Private: Reddit (de & tech), YouTube (tech channels), podcast platforms
• Little social media except LinkedIn & YouTube
• Uses RSS feeds, tech newsletters

💡 A ttitude towards innovation

• Open to innovation, if sensible and well thought out
• Skepticism towards hype topics without real added value
• Values open source and IT security
• Would like to work with modern tools, but with healthy pragmatism

📢 Attitude towards advertising

• Rejecting intrusive advertising
• Accepts advertising if it offers informative added value
• Prefers technically relevant, factual content

✈️ Attitude towards travel

• Business trips only if necessary, preferably remote work
• Privately like short vacations with family, no long-distance travel
• Sustainability-conscious: prefers train to plane

🥗 Attitude towards nutrition

• Flexitarians with a tendency towards vegetarianism
• Places value on regional & sustainable products
• Canteen offers important when choosing an employer

🛒 Attitude towards online shopping

• Uses online shopping regularly (Amazon, notebooksbilliger.de, etc.)
• Appreciates price comparison portals & customer reviews
• Also supports local dealers, if possible

📍 Digital and analog advertising touchpoints

Digital:

• LinkedIn Ads (job offers with techn. details)
• Newsletter on Heise, Golem
• YouTube pre-roll ads for tech channels
• Stack Overflow & GitHub Sponsored Jobs
• XING (limited)
• Reddit campaigns in r/en & r/sysadmin

Analogs:

• Trade journals (e.g.c't, iX)
• Posters/advertisements in canteens of utility companies
• University cooperation (dual study programs, evening events)
• Meetups, barcamps, local IT get-togethers

🔍 Job search

📌 Where does the persona search?

• LinkedIn & Stack Overflow Jobs
• Heise Jobs, Golem Career World
• Career pages of municipal utilities and suppliers
• Recommendation from colleagues / network
• Career events (IT trade fairs, barcamps)

📣 Address

• Technically sound, but understandable
• No HR buzzwords (e.g ."rock star", "nerd")
• Emphasis on stability, but also modernization
• References to meaningful work & sustainability

✅ Expectations of the job

• Stable, long-term employment
• Modern IT systems or the will to renew
• Influence on technical decisions
• Further training opportunities
• Family-friendly working hours / home office

⏳ Expectations of the application process

• Clear, transparent communication
• Technically oriented discussions with realistic tasks
• No excessive assessment centers
• Fast feedback, ideally within 2 weeks
• Opportunity to get to know the team informally

🚧 Pain points for job & application

• Outdated technology & systems
• No listening ear for suggestions for improvement
• Unstructured application processes
• Slow feedback
• Inflexible working time models
• Decision-makers with little technical understanding

💚 Plus points for employers

• Modern technical equipment
• Transparent corporate culture
• Flexible working time models / home office
• Further training & certifications
• Sustainability & social contribution of the company
• Family friendliness

🚫 Things that keep you from applying

• Unclear task descriptions
• "We are looking for an IT ninja" rhetoric
• Non-transparent salary information
• No home office possible
• No real technical contact person in the application process
• Lengthy or unprofessional application tools
• Arbeitgeberbewertungen auf Kununu < 3,0

😤 Reasons for abandoning an application

• Application system with mandatory registration or errors
• No feedback after 2 weeks
• Unprofessional communication
• Interlocutor without technical understanding
• Incorrect or embellished presentation of technology in the company

Although Martin and Tobias have undeniable similarities, they are two different personas. The Big Five analysis we received with Martin was withheld from us with Tobias. And even worse: when we enter the same prompt a third time, we get to know Markus:

💼 CANDIDATE PERSONAL PROFILE: "Markus - The security-oriented IT specialist"

📊 1. demographic data

• Name: Markus Schröder
• Age: 38 years old
• Gender: Male
• Marital status: Married, 2 children (7 and 10 years old)
• Place of residence: Outskirts of a medium-sized German conurbation (e.g.Mannheim)
• Education: Graduate computer scientist (FH)
• Professional experience: 12 years in IT, including 5 years in IT infrastructure at a utility company
• Employment status: Permanently employed in a non-terminated position

The other results for Markus go beyond the scope of this article, but differ significantly from the other two. You understand what this is all about: In a fourth inquiry, we meet Thomas from Stuttgart...

This prompted us to ask Chat-GPT what the personas are based on.

ChatGPT praised us for the critical demand and named an astonishing number of studies and sources for Martin - some of which actually exist, such as the ARD/ZDF Online Study 2024 or the bottleneck analysis by the Federal Employment Agency. With other sources, ChatGPT remained rather vague and only mentioned "PwC & EY studies on digital transformation in the energy and utilities industry" or "VDI/VDE studies on IT in the energy sector (smart grids, OT/IT security)", without saying exactly which ones - if they exist. We were told that one of the sources mentioned, the "Universum Talent Survey 2024", did not exist:

"I used "Universum Talent Survey 2024" as a collective term, but in reality such surveys are usually region-specific, not necessarily a survey called exactly "International Talent Survey 2024" and published nationwide for all target groups. However, there are comparable data and rankings from Universum from 2023/2024, e.g. specific to individual countries or student groups. My persona is based on a mixture of such sources - i.e. partially extrapolated."

Furthermore, Martin should be based on "Typical feedback from applicant interviews and employee value proposition projects in the public and semi-public sector". Asked how ChatGPT knows these, it tells us:

"I do not have access to real, personal application data or confidential interviews, and of course I do not collect or store any of it.

When I mention "typical feedback from applicant interviews and EVP projects" in the sentence, I mean something else:

I rely on public, aggregated or secondary published results from such projects - not original interviews, but published findings and patterns that come from many such projects.

I have learned these typical patterns in my training and knowledge - not through specific personal data, but through:

• Published trend analyses
• Technical literature and study evaluations
• Generally accessible industry examples"

In other words: I read all possible available studies and articles in which someone has written something about IT employees in the utilities sector, throw everything into a pot, stir once and then make up what Martin probably thinks about applications."

ChatGPT also explains the Big Five: "I "calculate" with probabilities, mutatis mutandis:

If someone works in a safety-related profession, is interested in sustainability and has an affinity for technology, then it is plausible that they would be located in models X and Y in such and such a way."

The results are nice clues, but in no case a real persona.

When asked the same question about Tobias, ChatGPT admits after naming generic, partly non-existent, but for the uninitiated very convincing sounding sources: "This persona is a synthetically generated ideal typology that is based on realistic and generalized patterns - not a real person, but a fictitious target group representation, as used in employer branding or recruiting for segmentation and targeted addressing."

If there is one thing data-driven personas do not do, it is "generalize". Generalization is the exact opposite of data-driven and therefore not helpful for targeting the right audience in either recruiting or marketing. Personas are much more about moving from generalization to personalization in order to achieve tailored experiences for candidates and customers.

LLMs are wonderful for many things - but only to a limited extent for data-driven personas

Chat GPT - and its "colleagues" Gemini, Claude and Co. can therefore "conjure up" personas that sound convincing at first glance. On closer inspection, however, inaccuracies, contradictions and overgeneralizations quickly become apparent without self-provided data, which can lead to misguided strategies. A major problem here is that LLMs not only create false personas, but also cite false but very convincing-sounding sources. Furthermore, they generalize what they have learned or simply hallucinate. Why? Because LLMs are language models: They calculate the probability of one letter being next to another and create texts based on that - extremely simplified.  

LLMs are wonderful tools for inspiration, brainstorming, text optimization, even programming, or for bringing personas to life or recognizing patterns in data.

However, those who blindly rely on AI end up with personas that don't work or, in the worst case, end up in a situation like the management consultancy Deloitte, which recently relied on Chat GPT and its plausible-sounding sources for a major government contract in Australia and then had to admit mistakes and pay a fine - not to mention the loss of reputation.

If you want to check all the sources provided by the AI, you often have to ask the LLM for them and then check one by one whether they exist and are relevant. This saves almost zero work compared to "doing it right". Especially if, in addition, the LLM has not been "fed" with the self-collected data relevant to the desired persona. This means that even if all sources exist and are related to the topic, the resulting persona is data-based but still not 100% evidence-based, because here and there you will still find generalizations from secondary sources, which is better than pure gut feeling, but not as evidence-based as using first-party data.  

Challenges of AI-based persona creation

Used correctly, AI helps us to create better, deeper and more vivid personas than we could ever do "by hand" or with human effort, thought and analysis. But even if we do everything right, feed our AI with enough and the right data and train it according to the textbook, there are still some pitfalls that we need to be aware of:

 

• We need to train our AI with data that is as up-to-date as possible, otherwise our result personas could soon lag behind reality. Even large LLMs face this problem. Iterative learning during persona creation and constant adaptation in real time counteract this.
• Historical data used for AI training has another problem: it can reflect existing social prejudices. This could reinforce stereotypes or fail to accurately reflect social diversity. Identifying such biases in large amounts of data so that the personas ultimately represent the target group fairly and accurately is not easy.
• Quantitative and qualitative data should flow equally into the persona. AI is particularly good at evaluating large amounts of quantitative data, but the qualitative part must not be neglected, as this is where the subtle nuances of emotions, desires and needs that make up the depth of a good persona become apparent.
• Data protection laws also and especially apply to big data. Work on personas must be data protection compliant at all times.
• The cultural characteristics of personas are just as important. What is a veritable outburst of anger in one culture is just a clear statement in another that does not frighten anyone. To make personas culturally sensitive, the data should come from the region in which the persona is used. It is also important to see the persona in the right context: A bank employee ticks differently to a physiotherapist and differently to a student, even if there are many similarities at first glance.