What AI does well in a business
- Assistant on your documents: asking a question in natural language and getting an answer drawn from your procedures, contracts or product sheets, with the source cited.
- Automation of repetitive tasks: sorting incoming requests, preparing a standard reply, matching invoices with purchase orders.
- Data extraction: reading documents whose form varies (scanned forms, PDFs, emails) and producing structured fields, ready to be stored.
- Sorting and summarising: grouping customer feedback by theme, summarising a long exchange, spotting what needs action.
The first question: is AI really needed?
Not every repetitive task deserves to be automated, and not every task that does needs artificial intelligence. If the data is already structured and the rules are clear, conventional logic does the job better: it is simpler, cheaper to run and perfectly predictable. AI becomes the right tool when you need to understand natural language or documents that are never quite alike.
In a well-built project, the language model only handles one step anyway: understanding. The input (an email, a document), the business rule that decides the action and the action itself (writing to the database, notifying, filling in a field) remain conventional code.
A safety net, not a black box
The most neglected part of an AI project is not the choice of model, it is what happens when the model gets it wrong. Every integration I deliver includes:
- decision logging, to reconstruct what happened when something goes wrong;
- a low-confidence signal: the system must be able to say it does not know, rather than guess;
- human validation for ambiguous cases, instead of a decision applied blindly.
An automation without this net is no more reliable than a manual process: it just gets things wrong faster.
Your data stays under control
The choice of model is made after clarifying how sensitive the data is. For public or low-sensitivity content, the API of a major provider is often the simplest solution. For confidential data, an open model can run on infrastructure you control, with nothing leaving your environment. In both cases, the question is dealt with at scoping, not discovered afterwards.
A full-stack developer first
An AI integration is still a software project: it needs an interface, a database, secure access, hosting. That is why I approach it as a full-stack developer rather than as a specialist in a single model: AI is one more tool in the box, not a default answer. I explain this choice in Staying a generalist in the age of AI (in French).
Frequently asked questions
Does my business really need AI?
Not necessarily. Many repetitive tasks are better automated with conventional logic, which is simpler and more predictable. AI becomes useful when you need to understand natural language or documents whose form varies: emails, forms, invoices, customer feedback.
What happens to our data?
This is settled at the scoping stage, before the model is chosen. Depending on how sensitive the data is, we can use a provider’s API with suitable contractual guarantees, or run an open model on infrastructure you control.
What happens when the AI gets it wrong?
A well-built system plans for it: every decision is logged, the system flags when it is unsure rather than guessing, and ambiguous cases go through human validation.
Can AI be added to an existing tool?
Yes, it is in fact the most common case. The integration is usually built alongside the existing tool, through its API or its database, without having to rebuild everything.