AI Applications: Which Technology for Which Problem?

Is your support team answering the same questions every day? Do you have thousands of products but customers can't find what they're looking for? Are you trying to predict next month from your sales data? All of these can be solved with AI, but "AI" is such a broad term that it's hard to know where to start. In this post we sort AI applications not by technology, but by the problem they solve.

1. If you work with text: Natural language processing

Customer questions, emails, forms, documents, reviews... If text sits at the centre of your work, what you need are natural language processing (NLP) applications:

• Chatbots and assistants: Answering common questions 24/7, booking appointments, reporting order status
• Automatic classification: Sorting incoming requests by topic and routing them to the right person
• Summarising and extraction: Pulling summaries or specific fields out of long documents
• Review analysis: Seeing which topics come up most in customer feedback

On the technology side, most projects today combine large language models (APIs from providers such as OpenAI, Anthropic and Google, or open-source models like Llama and Mistral) with your own data. If you need a language other than English, test models with real examples in that language before deciding; quality varies noticeably between models.

2. If you work with images: Computer vision

If you work with photos, video or camera feeds, computer vision applications come into play:

• Document reading (OCR): Extracting data from invoices, forms and receipts automatically
• Visual search: Showing products similar to a photo the customer uploads
• Quality control: Catching defective products on a production line with a camera
• Object detection: Shelf monitoring, counting, security and similar scenarios

Ready-made cloud services (such as Google Cloud Vision or AWS Rekognition) get you started quickly; for specific needs, models like YOLO are trained on your own images. Image processing needs more computing power than text, so the number of images you process per second directly drives cost. For uses involving personal data, such as face recognition, plan for GDPR and local privacy rules from the start.

3. If you want personalisation: Recommendation systems

The "people who bought this also bought" logic is used in e-commerce, content platforms and apps to show each user the right product or content. Methods either learn from user behaviour (collaborative filtering), start from product attributes (content-based), or combine both.

A caveat: recommendation systems only work with enough user interaction. If your traffic is still modest, starting with simple rules such as "best sellers" and "more from this category" is both cheaper and more accurate.

4. If you want to see ahead: Forecasting and anomaly detection

Used to predict the next step from historical data, or to catch situations that deviate from normal:

• Demand forecasting: Predicting how much of each product will sell to plan stock
• Anomaly detection: Spotting suspicious transactions or unexpected drops early
• Predictive maintenance: Getting warnings before a failure from machine sensor data

The methods here are often "classic" machine learning (time-series models, algorithms like XGBoost). Success depends more on data quality than on the model: even the best model built on missing, wrong or inconsistent data will make bad predictions.

5 questions to ask before you start

• Does this problem really need AI? If it can be written as a clear "if this, then that" rule, simple automation is cheaper and more reliable
• Do I have enough data? Rather than training a model from scratch on little data, adapting an existing model or API to your needs is usually the smarter move
• Do I need to explain the decision? In sensitive areas such as credit, hiring or health, you must be able to show why the model decided what it did
• Real-time or batch? A chatbot must answer in seconds; a monthly forecast report can run overnight. Real-time systems cost more
• Cloud or my own server? Cloud APIs get you started fast; if the data is sensitive or volumes are very large, open-source models running on your own servers are worth considering

Off-the-shelf tool or custom build?

If your need is standard (a general support chatbot, basic document reading), SaaS tools are usually enough and can be running within days. A custom build makes sense when you have data and needs specific to your business, when you want AI embedded inside your own app or system, when privacy means data can't leave your infrastructure, or when the feature is what sets you apart from competitors.

We cover where small businesses can realistically start in our post "Does AI actually help small businesses?".

Common mistakes

• Ignoring data quality: Bad data never produces good results
• Starting with the most complex solution: First show the problem can be solved with a simple method
• Forgetting the model after launch: As user behaviour changes, performance drops; it needs regular measurement and updates
• Calculating cost too late: API costs grow with usage; map out volume scenarios from the start
• Using output unchecked: Language models can produce wrong information with confidence; put limits and review on anything that reaches customers

Start small, grow by measuring

Even with a big AI vision, keep the first step small. If you want personalised product recommendations, for example, start with simple rules, then add suggestions based on what the user viewed, and only then move to a model that learns from other users' behaviour. Measure concrete metrics such as clicks and conversions at every step. We explain the thinking behind this approach in our posts "What is an MVP" and "Digital product development".

If you have an AI idea in mind, in a free assessment we can work out together which problem truly needs AI, which technology fits, and what a small first step could look like.