Opinion: Why GenAI isn’t a silver bullet

The rapid evolution of generative AI continues unabated, as does the corresponding hype, making this tumultuous landscape difficult to navigate.

As a result, many organisations are struggling to understand when and when not to apply generative AI within their business.

For many organisations or business units, generative AI will be their first experience with AI and will start conflating it with AI.

This risks generative AI overshadowing the broader AI landscape, yet it is only a small piece of the puzzle.

This is an important problem because different use cases may require different AI techniques. generative AI isn’t a silver bullet. The hype surrounding generative AI can lead organisations to apply the technology where it isn’t a good fit.

This increases the risk of higher complexity and failure in their AI projects.

Gartner predicts at least 30 percent of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value.

In addition, many business problems will require a combination of different AI techniques, which are likely to be ignored if organisations maintain a short-sighted focus on generative AI.

Avoid using generative AI for the wrong purpose

The question of whether to use generative AI models is a case-by-case decision. The first step is to determine if the use case is valuable and feasible, regardless of the AI technique being considered. Some just aren’t a good fit for AI and don’t merit further consideration.

After prioritising your use cases, evaluate the current usefulness of generative AI models for each one.

This is important as there are a few areas where generative AI models are typically misused in organisations.

Planning: Planning and optimisation are some of the key missing elements in current generative AI models. This limits their use for valuable use cases like inventory optimisation, field workforce scheduling, route optimisation, financial portfolio optimisation, pricing optimisation in retail, and resource allocation.

Prediction and forecasting: Don’t use large language models (LLMs) and other generative AI models for forecasting use cases, such as demand prediction, sales forecasting, time-of-arrival estimation, weather forecasting and supply chain forecasting. LLMs aren’t currently designed to do the kind of numerical predictive and statistical modelling required.

Decision intelligence: Current generative AI models aren’t built for decision making. Their output is unreliable, they lack explainability and they aren’t able to model decisions in an explicit way to achieve outcomes. Using generative AI outputs to make critical decisions is risky, such as hiring, budget and financial planning, supply chain management, marketing allocation and strategic decision making.

Autonomous systems: generative AI has an autonomy gap — current models require close human supervision given their inaccuracies and hallucinations. This limits the usefulness of generative AI for things like industrial robotics, delivery drones, smart factories, algorithmic trading and autonomous vehicles.

A broader reason why generative AI might not be a good fit is that the risks that come with generative AI are unacceptable and can’t be effectively mitigated. Risks include output unreliability, data privacy, intellectual property, liability, cybersecurity and regulatory compliance. These need to be considered for each individual use case.

Consider alternative AI techniques

Generative AI has taken focus away from existing and proven AI techniques. Alternative techniques, or their combination, may represent the best fit to support specific use cases.

Some of the main alternative techniques include nongenerative ML (or predictive ML), optimisation, simulation, rules/heuristics and graphs.

Map your use case against the suitability of different AI techniques to understand what alternatives are currently the most useful.

These alternatives can be more efficient, effective and reliable, and better understood than generative AI models for many use cases.

It’s key to consider what is needed for the specific use case in terms of explainability, performance and reliability.

Trying a simpler alternative AI technique first can be a good idea because it can be less risky, less expensive and easier to understand.

Combine generative AI models with other AI techniques

AI techniques aren’t mutually exclusive. They can often be combined in a way that makes for a stronger overall system. The combination of generative AI models with other AI techniques can be particularly powerful.

On one hand, some of the limitations of generative AI models — such as their lack of robustness, inaccuracies and hallucinations — can be mitigated by coupling them with more robust techniques. For example, knowledge graphs can help reduce hallucinations in LLMs.

On the other hand, generative AI models can be helpful additions for established techniques.

A common way generative AI models can be useful is by serving as a natural language interface to other AI or software systems.

The space of potential combinations of AI techniques is very broad. Organisations that develop an ability to combine the right AI techniques are uniquely positioned to build systems that have better accuracy, transparency and performance, while also reducing costs and need for data.

About the author

Leinar Ramos is a Senior Director Analyst at Gartner focused on generative AI and key management priorities relating to AI. The latest AI trends will be discussed during Gartner IT Symposium/Xpo on the Gold Coast (9-11 September).