As with any emerging technology, GenAI’s mishaps catch more media’s attention than its successes. You’ve probably heard of at least a few such instances.
We’re not just talking about six-fingered hands and other visual glitches that internet users eagerly share, but business implementation fails. For example, Chevrolet’s chatbot offering a customer the deal of a lifetime – a 2024 Chevy for $1. Or the discount promised by a Canadian Airlines chatbot that led to the airline losing a court case.
But have you heard that Klarna, after launching an AI assistant, reduced the average resolution time to 2 minutes (from 11 minutes) within a month, and their estimated profit increase in 2024 is $40 million? Their assistant performs the work equivalent to 700 full-time agents.
Common mistakes and risks of using Generative AI models
Klarna’s project is just a drop in the ocean of GenAI successes, as confirmed by studies cited in the Microsoft New Future of Work Report (December 2023). These studies clearly show that by using language models, employees needed 37% less time for typical text-related tasks and were twice as fast in solving decision-making problems.
What makes some GenAI projects enormously successful while others end in failure? Through the implementation of over 20 GenAI projects and conversations with dozens of organizations in recent months, we have gathered valuable insights that you can turn into your own success.
Dive into the 10 most common mistakes in implementing GenAI, with exemplary Generative AI business use cases. Read to the end – the summary contains a checklist that will help you navigate the risks of generative AI in business.
#1 Implementing (Gen)AI in projects that do not need it
The hype for a particular technology can make it look like a cure for everything – but it never is. Technology should be chosen based on the project’s needs.
Let’s say it out loud – AI, especially GenAI, is not always the best solution for all problems. If you can outline the exact steps for a task, there’s usually no need for AI. Traditional coding will work just fine. In some cases, solutions that don’t involve machine learning are better for automating processes.
Rethinking AI: not always the ideal solution
This might be surprising to you – even among people working in IT, automation in 2024 is basically synonymous with AI. Implementing AI wherever possible, however, is a big mistake because, in many cases, its use simply does not pay off. Furthermore, AI often constitutes only a part of the whole solution. It can be a key part, but we cannot forget about the rest of the system, which often does not require the use of AI.
Let’s illustrate this with an example from the autonomous vehicles sector. Contrary to common misconception, the vehicle control system is not a single ML model. It is a collection of components performing specific tasks, such as obstacle detection, reading traffic signs, maintaining distance between vehicles, etc.
Most components of this system do not actually use ML. It is cheaper, safer, more reliable, and much more efficient this way. So, when considering AI app development, analyze exactly where artificial intelligence will really be useful.
#2 Relying on generic solutions
Remember the study mentioned at the beginning of this article? The same MS New Future of Work Report cited statistics from another study, in which a language model was given a deliberately tricky task that included conflicting source information. As a result, BCG consultants relying on the language model made 19% more mistakes.
The conclusion of the study was, however, optimistic: these problems can be avoided by appropriately designing the implementation. This finding is 100% consistent with our experience. By adjusting the way knowledge is stored and retrieved and breaking down tasks into smaller parts, you can increase the accuracy of the results. Add dedicated prompt engineering and thoughtful UX to the mix, and the chances of a successful GenAI launch increase significantly.
Generic models, like MS Copilot or ChatGPT, lack these elements, which is their Achilles’ heel in business implementations. Even if they could integrate with all the organization’s data, their functionality without adaptation to processes will always be limited.
Even if you can operate on generic models, a custom solution will surely guarantee your users better results and greater convenience, as well as provide you with the highest return on investment.
#3 Implementing tools without their verification
AI tool demos often make the implementation look like pure magic. Just integrate the tool, and everything starts working as it should, with benefits flowing in abundance. Unfortunately, reality checks such promises, as we know from our own experience.
Here’s an example: Management at a large organization decided to implement a tool to automate report creation. It was supposed to automate the work of hundreds of analysts. For some reason, management did not consult this decision with the technical team or anyone with substantive knowledge of how language models work. A contract worth millions was signed, but it turned out that the solution had no chance of working.
The takeaway: always consult the implementations of the AI tools with independent specialists to avoid crying over spilled milk…or counting lost money.
#4 Lack of resources to build and maintain the solution
Before the implementation, verify whether your company has the right internal competencies. The best choice may be to join forces with an external Generative AI development services company who can help implement and maintain GenAI systems.
Often, it’s worth starting with external specialists because:
- Building specialized competencies takes a lot of time and effort. Even at Miquido, we sometimes collaborate with external agencies, consultants, or experts for support when we’re building or expanding new competencies.
- External specialists have experience and do not need to learn from costly mistakes. They have created similar solutions and have the necessary “know-how,” advising you throughout the process instead of just dealing with the code itself.
- Building internal competencies is expensive. Contrary to appearances, the costs of specialized consultations are lower and protect you from huge expenses associated with rescuing AI projects implemented without adequate knowledge.
- External specialists will support you for a specified time. If your company decides to end an AI project, you can do so without employment issues.
#5 Using non-production libraries
Due to the massive GenAI boom, many libraries have emerged aimed at simplifying the development of applications using LLMs. However, many of them are not yet suitable for use in production applications.
Take LangChain, for example – it’s a popular library that companies love because it makes building proof of concepts incredibly easy. But here’s the catch – it’s not even in its Beta phase, and new releases come out every few days. If you use it, you’ll constantly need to update your code. This means losing time and money, and it also introduces a lot of vulnerabilities.
By leveraging the support of companies specialized in generative AI services, you can avoid such problems. Those who have conducted numerous production deployments will help you choose the best infrastructure in terms of technological maturity, dependencies, and data security.
#6 Inadequate knowledge provision
For Large Language Models to be able to answer specific questions, you need to provide them with knowledge. Without appropriate data and control mechanisms, they will hallucinate and make mistakes. The key is not only the selection of knowledge but also its organization, especially with large volumes of sensitive data.
For example, if you have a document in your database that had previous versions, you need to properly mark the current one or get rid of the archives. Otherwise, the model might provide outdated information, not knowing which training data to use.
#7 Not estimating ROI
Before you join the AI bandwagon, make sure that the benefits justify the costs. Consider the scale of improvement to calculate a realistic return on investment in time.
For example, let’s assume you invest in custom automation for generating customer offers, which saves a salesperson 2 hours per offer. With 5 offers per week, it might pay off in 3 years. With 30 offers per week, the payback period is reduced to just 6 months. This is an important factor that must be taken into account.
Also, consider the costs of consumed tokens or the infrastructure costs if you deploy the model on-premise. With scale and the use of the most powerful generative AI models, costs can quickly rise.
Therefore, it is crucial to estimate them precisely before starting the work. The AI calculator created by Miquido can help with preliminary cost calculations for simple scenarios. We always provide a more accurate estimate after analyzing the client’s requirements.
Remember – ultimately, the costs of using LLMs are significantly lower than the costs of automated human labor. You just need to implement it wisely – preferably with an experienced partner.
#8 Ignoring risks of Generative AI in business
When implementing GenAI, your company should be aware of the risks and actively mitigate them with its implementation strategy. Consider aspects such as:
- Scalability and Flexibility: How can you build an generative AI solution that can easily adapt to the changing needs of the company?
- Compatibility with Existing Systems: How easily can AI be integrated with your current processes and systems? Will it work with your existing infrastructure, or will it require significant changes to work?
- Regulatory Compliance: What regulations will come into effect in the coming years (like the AI Act in the EU)? How will you address copyright issues that add up to risks of generative Artificial Intelligence in business? How will you adapt the generative AI tool to regulations specific to industries such as finance, insurance, medicine, or even publicly traded companies?
#9 Not testing enough
AI is a bit like…paracetamol. Though it is hard to believe, the pharmaceutical industry still doesn’t know how paracetamol works, despite it being widely regarded as one of the safest painkillers. We reached this point not through understanding, but through thorough research on its effects.
The same applies to AI. By continuously testing your AI solutions—both manually and automatically—you can ensure their safety and prevent hallucinations. This helps avoid material and reputational losses, such as those mentioned earlier with Canadian Airlines or Chevrolet. For systems intended directly for customers, unsupervised by humans, testing is a crucial preventive measure worth allocating time and resources to.
There are solutions you can implement for this purpose. For example, at Miquido, we developed a tool that automatically tests predefined scenarios. Additionally, every solution is tested by trained manual testers, mitigating risks of generative AI in business.
#10 Lack of organizational readiness for AI implementation
Implementing an AI tool can cause more problems than benefits if your industry and the company itself is not ready for it. If the end users are not familiar with the generative AI technology and are reluctant to use digital solutions, it’s worth considering training them and raising awareness first. Automating a process inevitably changes it, and the organization must be prepared for such a change.
Avoid these mistakes – your ultimate checklist
- Make sure that AI is the right technology for your purpose. If non-AI systems solve the issues you want to tackle with your project, it will be cheaper and more efficient to implement them.
- Don’t rely on generic solutions. Adjust knowledge storage and retrieval, UX and prompt engineering to the specifics of your processes.
- Verify the AI tools your company plans to implement with experts – internal, external, or ideally both. Sometimes demos don’t pass the reality check.
- Think of engaging an external team to avoid costly mistakes, save time and provide yourself with maximum flexibility. Don’t save on consultations – they will save you expenses later.
- Ensure that the tech stack you chose for your AI project is suitable, mature and secure.
- Organize the knowledge adequately to prevent hallucination.
- Estimate ROI, considering AI app development costs, the scale of the project, cost of consumed tokens and the infrastructure cost. You can do the initial estimates with our tool!
- Address risks of generative AI in business in your strategy, including scalability and flexibility, compatibility with existing systems and the regulatory compliance.
- You can prevent hallucinations and mistakes that cost you reputation with testing. We have a tool for that as well!
- Your organization should be ready for AI implementation. In some cases, it is better to start from educational efforts and training to make sure that the solution will serve its purpose.
As an AI development company, we partner with our clients to develop a wide range of artificial intelligence solutions. We help maximize the potential of their implementations and mitigate the risks of generative AI in business.
If you are in the research phase, we recommend our articles: on our blog, you can learn about AI trends in mobile apps, the difference between AI and gen AI, machine learning vs generative AI, and much more. And for more information or business insights, reach out to us!
Top AI innovations delivered monthly!
The administrator of your personal data is Miquido sp. z o.o. sp.k., with its ... registered office in Kraków at Zabłocie 43A, 30 - 701. We process the provided information in order to send you a newsletter. The basis for processing of your data is your consent and Miquido’s legitimate interest. You may withdraw your consent at any time by contacting us at marketing@miquido.com. You have the right to object, the right to access your data, the right to request rectification, deletion or restriction of data processing. For detailed information on the processing of your personal data, please see Privacy Policy.
