How to use AI safely - tips and tricks

How can AI support customer loyalty? And can it minimise theft in stores? What are good prompts? How safe is AI and what are its limitations? These and other questions were discussed at a dedicated day of talks around artificial intelligence (AI) hosted by the German digital centre for SMEs with a focus retail on 11th February 2025. FashionUnited has summarised the key takeaways.

Philipp Hübner, project manager for e-commerce research, presented the EHI Retail Institute's white paper “AI and Customer Loyalty in Retail,” which is based on a survey of 232 decision-makers. They were asked, for example, whether it is easier to retain existing customers or acquire new ones (63.6 percent affirmed the former), and which factors influence customer loyalty. They also commented on areas where they see the greatest potential for AI.

AI and customer loyalty in retail

The respondents came from twelve different sectors, with 15.5 percent from the fashion and accessories sector and 1.3 percent from department stores or shopping centres (0.4 percent). Of the 232 decision-makers, 105 were board members and managing directors of 179 retail companies with an estimated net turnover of 111.6 billion euros.

A significant part of the study focused on the use of AI to increase customer loyalty. While more than half (55.2 percent) of the respondents considers AI “very useful,” only about one-fifth (19.8 percent) confirmed that their company already uses AI to increase customer loyalty. The rest is either planning to use it (54.2 percent) or not (26 percent).

“Managing existing customers is a central task for decision-makers in retail. Most companies rely on a loyalty program, and the retail sector sees great potential in the use of AI to increase customer loyalty,” Hübner summarised. The best uses mentioned were personalised direct marketing, pricing, special offers and discounts, loyalty programs, and analyses and forecasts of customer behaviour.

How to write good AI prompts

Klaus Kaufmann, AI trainer at the digital centre’s value networks division, addressed the topic of “prompting,” i.e., how to formulate specific and precise instructions or queries to achieve a particular result. He referred to the acronym SPRICH (“speak” in German) to get the most out of AI prompts, because “the more general the question, the more general the answer.” SPRICH stands for:

• Situation, i.e., the context
• Presentation - in which format the result should be
• Role, i.e., the persona
• Instruction - the task
• Character - the tonality
• Hints, e.g., a sample input

A successful prompt should therefore be much more than one sentence, describing a role and giving a concrete idea of what the end result should look like.

Using AI to reduce inventory shrinkage

An interesting contribution by Frank Rehme addressed a still-relevant topic: shoplifting. The managing director of the digital center has been attending the National Retail Federation (NRF) trade show in New York annually for many years. Eight years ago, he noticed the innovation of a startup that simply put many hours of footage from supermarket surveillance cameras online. Interested parties could register, view the recordings and report thefts. In return, they received points in a virtual account.

“It worked well; many people participated and wanted to play store detective,” Rehme recalled. Certain conspicuous behaviour patterns were recorded – “for example, when someone hid something in their clothing, held a product close to their body, put something in a backpack, stroller or helmet, hid the shopping cart, bought many of the same items at once, or consumed food in the store itself.”

These manually identified behaviour patterns were then used to train AI. AI was then tasked with detecting thefts from a multitude of surveillance camera windows on a screen, which are difficult for humans to evaluate all at once. “AI divides the image into colours, and the colour changes when it detects something conspicuous. This gets recorded and ends up in an employee's app, who can then check in the store,” explained Rehme.

This involves video alerts and a dashboard, “small” hardware that can be integrated into an existing camera infrastructure. “This is a very efficient solution. Surveillance cameras have been around for decades, and equipping them with AI is feasible in Germany and Europe as well,” said Rehme.

The expert concluded by mentioning the case study of a large retail chain with more than 1,000 stores that tested this system without its own security personnel at airport stores, relying solely on cameras with AI and airport security. “The result after six months was astonishing: 84 cases were detected and 12 repeat offenders identified. Two trafficking networks for stolen goods were broken up and a total of 150,000 US dollars in inventory was recovered. The company had a fiftyfold return on investments.”

AI is no miracle cure

Martin Talmeier, the digital centre’s project manager and lead coach, warned that AI alone would not save companies. “AI alone is completely useless. It absolutely needs the magic ingredient, the fuel, to achieve great things for you. The magic fuel is data,” said the expert. And if this data is chaotic or unclear, AI will not be the saviour.

“There is great hope among companies that AI can solve what has been left undone, but there are many limitations. One is that the knowledge of LLMs (Large Language Models, the technology that powers AI tools such as ChatGPT) comes from the internet. “And as we know, a lot of useless information comes from there as well,” warned Talmeier. Internal, company-specific knowledge, for example, is not on the internet and should not be.

However, if you simply put AI into a company's internal data set, it has no idea what to do with it. “There is the hope: AI will clean up, but it won't work without doing your homework. You have to get involved yourself first,” explains Talmeier. The expert recommends establishing a data strategy, i.e., a plan for how data should be collected, stored, and analysed, or an inventory: “What do we have, what do we want?”

Only 36 percent of German SMEs currently have a data strategy in place, and only 29 percent use structured data management, i.e., the organisation and management of data in a clearly structured format. However, this is the basis for making data accessible, analysable, and usable, according to Talmeier. “82 percent of SME managing directors affirm the importance of data. There is a large discrepancy between what one does and what one wants.”

Next, Talmeier talked about the data economy, i.e., the economic use of data to create value. “The value creation potential of the data economy in Germany alone is estimated at 425 billion euros in the next two to three years.”

This is based on the consideration that 80 percent of data is unstructured. “However, this is not a problem for AI. The data we have neglected until now is the best thing for AI. Once we have sorted it into folders, categories, etc., we may have overlooked something that would have been important in the future,” cautions the expert, adding that “a machine can sort unstructured data, but only 18 percent of SMEs are already using this data treasure.”

What is holding companies back from using their data?

One hurdle for many are data protection laws. “However, they are not intended to prohibit business transactions, but rather function like guardrails,” Talmeier pointed out. In addition, there are compliance costs, internal guidelines, uncertainty in interpretation, data protection officers, and legal regulations when it comes to the use of personal data, all hurdles for companies.

“The principles of purpose limitation (what is the data needed for), data minimisation (data may not be used indiscriminately), and transparency apply,” stated Talmeier. However, AI must be controlled, for which he proposes a three-stage plan “to mine the data treasure.” This includes a strategy phase in which the goal is defined. The second step is data analysis, i.e., an inventory: which data is available where and how, how current is it, and of what quality? The third stage should be a pilot project on a small scale that is “manageable but beneficial,” such as an analysis of customer complaints, developing shopping carts, and the like.

Limitations of AI

IT security specialist and self-proclaimed white hacker Robert Boehme addressed the limitations of AI. For example, it cannot perform calculations and must be trained to use a calculator even for simple arithmetic tasks. What it can do is derive calculation methods and formulas. Likewise, all AI models have “memory gaps,” meaning something is missing. AI itself does not notice this; there is simply a gap, but at least the gaps are stable, meaning always in the same place, so that something can be done about them.

It should also be noted that AI has biases. For example, if one simply enters “Herbert or Michael” or “Michael or Herbert” into an AI tool, one will get one of the two names back, with the first name appearing more frequently if you do this several times. So this is a bias. If you enter the prompt thousands of times, “Michael” will come back more often because the name is more common in English-speaking countries than “Herbert,” and AI is fed with much English-language training data.

According to Boehme, one should therefore keep three things in mind when using AI:

• It is brilliant and yet has no clue.
• AI is extremely self-confident while being totally clueless.
• AI is extremely powerful.