“Who participated in the last election?” looks like an easy question, yet it is one of the best examples of why generic questions break AI chatbots. To answer it well, a chatbot needs natural language processing (NLP) to understand the words, named entity recognition (NER) to work out which election, which country and which time period are meant, and reliable data analytics to pull the right facts from trusted data, and if any of these steps fails, the answer is wrong, vague or made up.
The same problem appears every day inside companies, where people ask their AI assistants questions like “How did the last batch perform?” or “Which machines failed this week?” without saying which line, which site or which product they mean.
In this article, we explain in clear and simple terms why generic questions confuse AI, how modern AI systems can handle them, and how much time and money a well-built assistant can save, based on our hands-on experience at Tenplus building data, analytics and AI platforms for industrial companies.
- Why "Who Participated in the Last Election?" Is So Hard for AI
- How AI Chatbots Read a Question
- Named Entity Recognition: The Key to Understanding "Which One?"
- How Chatbots Fail on Generic Questions
- How to Build Chatbots That Handle Generic Questions Well
- The Same Problem on the Factory Floor
- What Is the Value? Hours and Dollars Saved
- Conclusion: Build AI Assistants That Understand Real Questions With Tenplus
- FAQs
Why “Who Participated in the Last Election?” Is So Hard for AI
A person who asks this question usually has a clear picture in mind, but almost none of that picture is in the words themselves. A human expert would quickly fill in the gaps from context, while an AI chatbot has to work out every missing piece on its own, and each piece can go wrong.
Which Election?
There are national, regional, local, party and even company board elections happening all over the world every year. Without knowing the user’s country, city or topic of interest, the chatbot cannot tell whether the question is about a presidential race, a parliament vote or a local council.
What Does “Last” Mean?
“Last” depends on when the question is asked, and an AI model trained on older data may not even know that a newer election has taken place. If the chatbot does not check today’s date and up-to-date sources, it can confidently describe an election that is no longer the most recent one.
What Does “Participated” Mean?
The word can refer to the candidates, the political parties, the voters, the voter turnout or the countries that sent observers. Each meaning needs a completely different answer and different data, so the chatbot must first decide what the user really wants to know.
Missing Context Everywhere
Put together, a single short question hides at least three open choices, which means there are dozens of possible “correct” answers. A chatbot that simply guesses will often be wrong, and a chatbot that refuses to answer will frustrate the user, so the real skill lies in handling the gaps in a smart and transparent way.
How AI Chatbots Read a Question
Before a chatbot can answer, it has to understand what was asked, and this is the job of natural language processing. NLP is the branch of AI that helps computers read, understand and produce human language, and it breaks a question down into a few key parts.
Intent: What Does the User Want?
The intent is the goal behind the question, such as asking for a list, a number, a comparison or an explanation. In the election example, the intent could be “list the candidates” or “give the voter turnout,” and each intent needs a different kind of answer.
Entities: What Is the Question About?
Entities are the specific things in a question, such as people, places, organisations, dates, products or machines. In our example, “the last election” is an entity that still needs to be pinned down to one real event with a place and a date.
Context: What Else Do We Know?
Context includes the user’s location, language, role, earlier messages in the conversation and today’s date. Good chatbots use this context to fill in missing details, in the same way a colleague would assume you mean your own country’s election unless you say otherwise.
Data: Where Will the Answer Come From?
Finally, the chatbot needs trusted data to answer from, such as an up-to-date database, official documents or a company’s own systems. Large language models alone only know what they saw during training, which is why they must be connected to live, well-organised data through data analytics and retrieval tools.
Quick link: Business Process Automation for Small Businesses
Named Entity Recognition: The Key to Understanding “Which One?”
Named entity recognition, often shortened to NER, is the NLP technique that finds and labels the important names in a piece of text. It is the step that turns a vague sentence into something a computer can look up, and it is often the difference between a useful answer and a confident mistake.
How NER Works
An NER model reads a sentence and marks words or phrases as types of entities, such as PERSON, ORGANISATION, LOCATION, DATE or EVENT. In a sentence like “Who ran for president of France in 2022?”, it would tag “president” as a role, “France” as a location and “2022” as a date, which gives the chatbot clear search terms.
Entity Linking: From Words to Real Things
Finding a name is only half the job, because the same name can point to many different things. Entity linking connects each recognised name to one exact record in a database or knowledge graph, for example linking “Georgia” to the US state or to the country, based on the rest of the conversation.
Handling Time Words Like “Last” and “This Week”
Words such as “last,” “latest,” “this week” or “previous shift” are relative, which means they only make sense when compared with a reference date. A good NLP pipeline turns these words into real dates using today’s date and the relevant calendar, so “the last election” becomes “the most recent national election before today in the user’s country.”
When an Entity Is Missing
Sometimes NER finds that a key entity is missing, such as the country in the election question. That is a valuable signal, because it tells the chatbot that it must either use a sensible default from context or ask a short follow-up question instead of guessing.
How Chatbots Fail on Generic Questions
When a chatbot is not built to handle vague questions, it usually fails in one of a few predictable ways. Knowing these failure patterns makes it much easier to test a chatbot and to design one that avoids them.
Silent Guessing
The chatbot picks one meaning, such as the most famous election in its training data, and answers as if there were no doubt. The answer may be correct for someone else, but it is wrong for this user, and because the guess is not shown, the user may never notice.
Outdated Answers
A language model that is not connected to live data can only describe events it learned about during training. It may present an older election as “the last one,” which is a serious problem for any question about recent events, prices, stock levels or machine status.
Made-Up Details
When the model cannot find a clear answer, it may fill the gap with names, numbers or dates that sound right but are not true, a problem often called hallucination. This is the most damaging failure, because it destroys trust in the chatbot and can lead to bad decisions.
Unhelpful Refusals
On the other side, some chatbots are set up to refuse anything uncertain, replying with “I don’t know” or a long list of questions. This feels safe, but it pushes users back to slow manual work and makes them stop using the tool.
How to Build Chatbots That Handle Generic Questions Well
The best AI assistants do not need perfect questions, because they are designed to understand intent, fill gaps sensibly and stay honest about what they assumed. The following building blocks work together to make that possible.
Use Context Before Asking
The chatbot should first try to fill missing details from what it already knows, such as the user’s country, site, team, role and earlier messages. If a production manager at Plant A asks about “the last shift,” the assistant should assume Plant A and the most recently finished shift, just as a colleague would.
Ask One Short Clarifying Question When Needed
When context is not enough and the possible answers are very different, the assistant should ask one quick, specific question, such as “Do you mean the last national election in Sweden?” A single, well-chosen question takes seconds and avoids a long, wrong answer.
Show Its Assumptions
When the chatbot does use a default, it should say so in one line at the top of the answer, for example “Showing results for the most recent national election in Sweden, based on today’s date.” This lets the user correct it instantly and builds trust over time.
Ground Answers in Trusted Data
Instead of relying on what the language model remembers, the assistant should fetch facts from approved, up-to-date sources, using retrieval augmented generation (RAG) for documents and text-to-SQL or semantic layers for databases. A well-organised data analytics platform, with clear names, definitions and access rules, is what makes this reliable.
Keep Testing With Real Questions
Collect the vague and tricky questions that real users ask, and test every new version of the assistant against them. Tracking how often answers are correct, how often clarifying questions are needed and how often users rephrase gives a clear picture of quality.
The Same Problem on the Factory Floor
The election question is a public example, but industrial companies face the very same challenge when they add AI assistants to their data. Engineers, operators, service teams and managers ask short, natural questions, and the assistant must work out which machine, site, product or time period they mean.
| Generic question | What is missing | How a well-built assistant handles it |
| “How did the last batch perform?” | Which line, which product, and what “perform” means (yield, quality or energy) | Uses the user’s home line, picks the most recent finished batch, and shows yield and scrap with a note on the assumption |
| “Which machines failed this week?” | Which site, what counts as a failure, and whether “this week” starts on Monday | Links “machines” to the user’s site in the asset register, uses the company’s failure codes and calendar, and lists stops longer than an agreed limit |
| “What is the downtime on Line 3?” | Which Line 3, since several plants may have one, and over what period | Uses entity linking to the right plant from context, and asks one quick question if the user works across sites |
| “Show me the energy use of the compressor.” | Which compressor and which time range | Recognises the asset type, offers the compressors the user has access to, and defaults to the last 24 hours |
| “Are we on track for the customer order?” | Which customer and which order | Finds open orders linked to the user’s accounts and asks the user to pick one if there is more than one |
For companies that build and sell industrial hardware, the same skills matter in customer-facing assistants and service portals. A field technician who asks “Why is the pump alarming?” expects the assistant to know which pump, at which customer, with which software version, and that only works when NER, entity linking and clean asset data are in place.
What Is the Value? Hours and Dollars Saved
An assistant that understands generic questions saves time in two ways, because users stop wasting time on wrong or vague answers, and data analysts receive far fewer simple requests. To make this concrete, here is a worked example for a mid-sized industrial company, based on careful assumptions that you can replace with your own numbers.
Example Assumptions
About 200 employees use an internal AI data assistant and ask around 1,000 questions per month. With a basic chatbot that does not handle vague questions well, about 30% of answers are wrong, outdated or unusable, and each failure costs the user about 20 minutes of follow-up work to find the answer another way. The data team also handles about 300 simple ad hoc data requests per month at about 45 minutes each, and we assume a full staff cost of $55 per hour.
Example Yearly Savings
| Value source | How it is calculated | Hours saved per year | Money saved per year |
| Fewer failed answers | Failure rate drops from 30% to 5% = 250 fewer failures per month × 20 minutes = about 83 hours per month × 12 | about 1,000 hours | about $55,000 |
| Fewer simple requests to the data team | 60% of 300 monthly requests are now answered by the assistant = 180 × 45 minutes = 135 hours per month × 12 | about 1,620 hours | about $89,100 |
| Total | about 2,620 hours | about $144,100 per year |
In this example, the company gains back about 2,620 hours per year, which is more than one full-time employee, and analysts can spend that time on deeper data analytics instead of answering the same simple questions. The time to get a simple answer also falls from hours or days, while waiting for an analyst, to under a minute. We have not counted the value of fewer wrong decisions based on bad answers, which can be much larger in areas such as quality, maintenance and customer orders. These figures are an illustration based on the assumptions above, not the results of a specific customer, and the real value will depend on how many people use the assistant and how they use it.
Conclusion: Build AI Assistants That Understand Real Questions With Tenplus
“Who participated in the last election?” shows that the hardest part of building an AI chatbot is not producing fluent text but understanding what people actually mean. By combining natural language processing, named entity recognition, entity linking, smart use of context and a strong data analytics foundation, companies can build AI assistants that handle everyday, generic questions with accurate, trusted and up-to-date answers, and save thousands of hours in the process.
At Tenplus, we design, build and run data, analytics and AI platforms for industrial companies and the hardware makers who supply them. Our team connects your machines, systems and documents into a clean data foundation, builds the NLP and entity recognition layers that link everyday words to the right assets, sites and products, and delivers AI assistants that state their assumptions and answer from your own trusted data.
We start with a focused pilot built around the real questions your teams ask, prove the value quickly, and then scale it across sites and users while your own team gains the skills to run it with confidence.
If you want an AI assistant that understands your people as well as your best colleague does, contact Tenplus today to book a free first conversation with our data and AI experts.
FAQs
Why do AI chatbots struggle with generic questions?
Generic questions leave out key details, such as which place, which time period or which meaning of a word the user has in mind. Without those details, a chatbot has to guess, and guesses often lead to wrong, outdated or made-up answers.
What is named entity recognition (NER)?
Named entity recognition is a natural language processing technique that finds and labels important names in text, such as people, organisations, places, dates, products and machines. It helps AI systems understand exactly what a question is about so they can search the right data.
What is the difference between NER and entity linking?
NER finds and labels a name in the text, while entity linking connects that name to one exact record in a database or knowledge graph. For example, NER spots “Line 3” as an asset, and entity linking decides which plant’s Line 3 the user means.
Should a chatbot ask clarifying questions or make assumptions?
The best approach is to use context first, make a sensible assumption when the risk is low and clearly state it, and ask one short clarifying question only when the possible answers are very different. This keeps conversations fast while avoiding costly mistakes.
How does data analytics help AI chatbots give better answers?
A chatbot is only as good as the data behind it, so a well-organised data analytics platform with clear names, definitions and up-to-date records is essential. It lets the assistant look up real facts instead of relying on what the language model remembers from training.
