Generative AI — language models such as ChatGPT, Claude, or Gemini that can write, summarize, analyze, and answer questions — has generated more excitement (and more confusion) than any technology in recent years. Every month brings promising new use cases and new examples of disappointing results.
Separating what works from what still needs to mature, specifically in business use, is more useful than any list of theoretical possibilities.
Use cases that create real, measurable value
Summarizing and extracting information from documents. This is probably the use case with the best ratio of value to implementation effort. Contracts, reports, meeting minutes, long emails: language models can summarize them, extract specific information (dates, parties, terms), and answer questions about the content with high reliability. For document-heavy businesses (law firms, insurers, accounting and administrative firms), the time savings are significant and measurable.
Drafting documents and communications. Generative models speed up writing sales communications, proposals, FAQs, product descriptions, and technical documentation. Note: they speed it up, they don’t replace it. The draft needs human review, especially for content with legal or commercial implications. The real gain is going from a blank page to a reviewable draft in minutes instead of hours.
Customer service assistants backed by documentation. A business chatbot built with RAG (retrieval-augmented generation) — which checks the company’s internal documentation before answering — can handle a meaningful share of frequent questions on its own; results depend on the quality of the knowledge base and the type of questions. Unlike the chatbots of five years ago, these systems understand natural-language questions and can answer in context. The value lies in lightening the support team’s load, not in eliminating the team.
Classifying and routing content. Sorting emails by department, prioritizing support tickets, categorizing complaints, tagging documents by type: generative AI is particularly effective here because it understands the content, not just keywords. A company can process hundreds of items a day without anyone reading them one by one.
Help with writing code. For in-house development teams, coding assistants (GitHub Copilot, Claude, Cursor) speed up writing repetitive code, documenting functions, and tracking down bugs. Studies on developer productivity consistently show significant gains. The impact varies a lot with the type of task and the developer’s experience.
Use cases that need more care than they seem to
Generating factual content without verification. Language models can make factual errors in a completely confident tone. For content that requires factual accuracy — market analysis, product data, regulations — review and verification are essential. The model speeds up the writing; verification is still human work.
Automating decisions with significant consequences. Using generative AI to make or recommend decisions on hiring, credit assessment, or claims handling without human oversight has legal and ethical implications under the EU AI Act. The model can inform the decision, but in these contexts the final decision needs human oversight.
Complex quantitative analysis. Language models are better with text than with numbers. For statistical analysis, financial modeling, or quantitative forecasting, specialized tools (Python with analytics libraries, purpose-built ML models) are more reliable. Generative models can help explain the results of a quantitative analysis, but running the analysis directly carries more risk of error.
Privacy: what data you can use
Sending company data to third-party APIs such as OpenAI or Anthropic is a decision that needs analysis. For non-sensitive data (public content, generic text, code without proprietary data), the risk is low. For personal data about customers or employees, or confidential company information, you need to review the provider’s data processing agreements and check whether they comply with the GDPR.
The alternative for sensitive cases is deploying open-source models locally (Llama via Ollama, Mistral via LM Studio), which keeps data inside the company’s perimeter at the cost of some performance on certain tasks.
How to start without making the most common mistakes
The most common mistake is giving the team access to a language model with no usage framework. The result is chaotic adoption, with everyone using it differently, uneven results, and no way to measure impact.
A more effective approach is to identify two or three concrete use cases with clear potential for improvement, design the workflow that includes the model, and define how output quality is checked before it’s used. Start small, measure the impact, and scale what works.
Frequently asked questions
Do we need to train the team to use generative AI well? Yes, although the training isn’t technical. What has the most impact is teaching people to write clear instructions (basic prompt engineering), to validate critical outputs, and to recognize where the model can go wrong. A well-designed two-hour session can multiply a team’s effectiveness with these tools.
Does generative AI replace jobs? In most business contexts, it changes the nature of work more than it directly replaces it. Tasks that used to take up skilled people’s time (writing standard communications, searching documents for information, classifying content) get significantly faster. That frees up time for work that requires human judgment, customer relationships, and strategic thinking.
How much does it cost to implement a generative AI assistant for the company? It depends heavily on the scope. A basic chatbot integration with internal documentation can be built in weeks at a relatively low development cost. A system integrated with the ERP, the CRM, and multiple data sources is a more complex project. Recurring model API costs are usually manageable for most companies at early usage volumes.