Not Everything Needs a Chatbot: How to Choose the Right Type of AI for a Product
Artificial intelligence does not add value simply by being visible in a conversational interface, but rather by better resolving a specific task, decision or bottleneck. Before choosing a technology, it is essential to understand the process, the user, the available data and the type of improvement you wish to achieve.
Applied AI · Product Strategy · Machine Learning · Generative AI · RAG · NLP · Automation
AI begins before the model
Choosing the right AI for a product starts with understanding what the user needs: to converse, automate, make decisions, anticipate, search for or understand information.
In many digital projects, the discussion about artificial intelligence starts both too late and too early. Too late, because it comes up once it has already been decided that ‘we need to incorporate AI’ without having properly understood the problem. And too early, because people start talking about chatbots, models or assistants before analysing what the product actually needs.
It is becoming increasingly clear to me that choosing the right type of AI does not start with comparing technologies, but with understanding the process we want to improve. Which tasks are time-consuming? Which decisions are repetitive? Where is information scattered? Which users need support? What data exists, what is its quality, and when is it generated? What is the level of risk if the system makes a mistake? And what degree of human control must be maintained?
A chatbot can be useful, but it shouldn’t be the automatic go-to solution for every AI opportunity. Sometimes the value lies in a recommendation, an alert, a classification system, a predictive model, intelligent automation, a semantic search engine, information extraction, or contextual assistance that doesn’t even feel like ‘AI’ to the user.
AI shouldn’t be implemented simply because it’s fashionable, but rather on the basis of a thoroughly understood product.
When the chatbot becomes the default response
One of the most common mistakes in today’s projects is assuming that incorporating artificial intelligence means adding a chatbot to the interface. Just as people used to say ‘let’s do more design’ – which sometimes translated into ‘let’s make the screens look nicer’ – many conversations now boil down to ‘let’s throw in some AI’, and almost automatically, the idea of plonking a bot in a corner pops up.
The problem isn’t the chatbot. The problem is using it as the default solution without understanding what task it is supposed to solve. A chatbot can help users look up information, guide them through processes, answer questions or assist them with open-ended tasks. But not every problem requires a conversation. There are situations where asking the user to type, ask a question, wait for a reply and manually verify the result creates more friction than it eliminates.
On a platform with complex social processes, for example, it may be tempting to think of a conversational assistant as a tool to help technicians, coordinators or management teams. But perhaps the real value lies in automating the recording of information, extracting data from documents, suggesting next steps, detecting inconsistencies, generating draft reports, recommending resources or prioritising cases according to defined criteria. In this context, AI can take many forms, and the chatbot is just one of them.
Something similar is happening in digital health. If the aim is to anticipate high-risk situations, perhaps we do not need an assistant that converses with the healthcare professional, but rather models capable of analysing signals, detecting anomalies, identifying deviations from a patient’s usual behaviour, or triggering early warnings. The interface could be a dashboard, a notification, a case prioritisation system, or a brief explanation of the reason for the alert. The intelligence lies in the analysis and the decision it facilitates, not necessarily in a conversation.
It is also easy to confuse automation with intelligence. Automating a repetitive task can add a great deal of value, but it does not always require generative AI. Sometimes rules, integrations, well-designed forms or business logic are sufficient. At other times, it does make sense to use NLP, embeddings, classification models, RAG or LLMs. The key lies in choosing the right capability for the right problem.
When you start with the technology, the product runs the risk of becoming a showcase for AI rather than a useful tool. You end up with chatbots that respond but don’t solve problems. Automations that look impressive but don’t fit into the actual workflow. Models that generate content but force the user to review everything. Or solutions that seem advanced but rely on incomplete, poorly structured or unreliable data.
Implementing AI means choosing the right kind of value
Before deciding whether a solution requires generative AI, traditional machine learning, NLP, RAG, embeddings or intelligent automation, it is important to understand what kind of value you want to create. Not all AI capabilities serve the same purpose, nor should they all be integrated into the product in the same way.
Talking doesn’t always mean finding a solution
A conversational assistant makes sense when the user needs to explore, ask questions, seek guidance or navigate complex information in a flexible way. But if the task is quick, repetitive or highly structured, a conversation may be unnecessary. In such cases, the value may lie in filling in fields, suggesting actions, displaying alerts or reducing the number of steps within the existing workflow.
Classical and generative AI can coexist
Generative AI does not replace predictive models, classifiers or recommendation systems. It solves different problems. A traditional model can anticipate risk, prioritise cases, segment users or detect patterns. A large language model (LLM) can summarise, draft, explain, interpret natural language or assist with knowledge retrieval. In many products, the best approach is not to choose one or the other, but to combine them judiciously.
The data determines the solution
You cannot design AI without examining the available data. Its quality, structure, traceability, timeliness and biases determine what can be automated, predicted or generated. A RAG system can be very useful if there is a reliable knowledge base; a predictive model can add value if there is consistent historical data; automation can fail if the input data is incomplete or in mixed formats.
Risk, explanation and control matter
The more sensitive the context, the more important it is to define what AI can do and what must continue to be validated by a human. Suggesting a label is not the same as recommending an intervention, prioritising a case or anticipating a clinical risk. In complex products, AI must be designed with criteria such as explainability, human review, trust and accountability in mind. It is not enough for it to simply work; it must be understandable, monitorable and sustainable.
From the problem to the type of AI
I work to identify which intelligent features add value before deciding how to incorporate them into the product.
When I’m exploring opportunities in applied AI, I try not to start with the format. I don’t ask myself first whether a chatbot, a co-pilot or a predictive model is needed. I start by understanding which part of the process needs improving and what kind of assistance would make sense for the user.
The first question is usually which task or decision we want to support. It could be a time-consuming administrative task, a complex search for information, a decision that depends on a large amount of data, a repetitive sorting task, a need to anticipate risks, or a process where the user needs to draft, summarise or interpret information. Each of these needs requires different capabilities.
Next, I look at the context of use. Who interacts with the system, at what point, under what pressure, with what level of autonomy, and what the consequences would be if the AI fails. Assisting an expert user who already knows how to interpret results is not the same as helping someone with little technological familiarity. Nor is proposing an operational improvement the same as intervening in a clinical, social or regulated process.
The next point is data. This is where much of the feasibility is determined. If there are reliable, up-to-date and well-structured documents, an RAG-based approach can help retrieve knowledge without relying solely on the model’s memory. If there is natural language in forms, interviews, tickets or CVs, NLP and embeddings can help to extract, compare or classify information. If there is sufficient historical data and consistent variables, predictive machine learning can be used to anticipate patterns, prioritise cases or detect anomalies.
You also need to decide how the output is integrated into the user experience. Sometimes the AI should appear as a conversational response. At other times, it might take the form of a structured recommendation, a score, an alert, an autocomplete suggestion, an editable summary, a contextual explanation, or an automated action that the user can accept, correct or dismiss. The interface should be tailored to the type of decision, not the hype surrounding the technology.
Finally, AI needs validation criteria. What constitutes ‘working well’? What errors are acceptable? What needs to be reviewed by a human? How do we measure whether it reduces time, improves data quality, increases accuracy, facilitates decision-making or enhances the user experience? Without that layer of validation, AI may appear useful in a demo but fail in real-world use.
Applying AI means deciding where it adds value
It’s not about making the product seem smarter, but about making it more useful
Choosing the right artificial intelligence solution is not simply a matter of deciding whether to use a chatbot, an LLM, RAG, embeddings, predictive machine learning or intelligent automation as if they were interchangeable components. It involves understanding the problem we want to solve, the type of data we have, the users involved, the decision-making process we wish to improve, and the level of risk we are prepared to accept.
When this is done properly, AI ceases to be a mere cosmetic feature and becomes a genuine product capability. It can help reduce manual workload, organise information, anticipate situations, generate drafts, recommend actions, detect anomalies or facilitate complex decision-making. But for this to happen, it must be connected to the process, integrated into the user experience and designed with a focus on control, trust and human validation.
Let’s talk about AI applied with a product-centric approach
We can discuss the product, UX and the application of artificial intelligence to identify which type of AI makes sense, how to integrate it into real-world workflows, and how to turn a technological opportunity into tangible value for users, the business and technical teams.