In recent years, many businesses have adopted AI by starting with the technology: they choose a tool and then look for somewhere to put it. That order produces demonstrations that never become part of the work. They function in the meeting where they are presented and then remain unused because they were not connected to anything that somebody needed to do every day.

The order that works is the reverse, and it is the same as for any other project. Start with work that currently consumes time or causes errors, examine how it is done and only then decide which technology should address it. Sometimes the answer is AI. Often it is ordinary automation. Sometimes it is traditional software, or simply a better-built module.

What a language model does well

It helps to be specific about where these tools are strong. The list is narrower than public debate suggests and more useful than scepticism implies.

Turning unstructured text into structured data. Extracting a company name, request type and urgency from an email; reading an invoice or document and returning its fields; assigning a free-form description to a category. It is the most reliable and least showy use case.

Classifying and routing. Assigning a ticket to the right category, distinguishing a sales enquiry from a technical problem, or recognising a tone that needs a quick response.

Summarising. Condensing a long conversation, a document or a series of recurring reports so that a person can decide without reading everything.

Answering from a defined knowledge base. Internal search that still works when a question uses different words from the documents. This application is often underestimated, yet regularly saves real time because the cost of failing to find company information is distributed and largely invisible.

Preparing a first draft. A reply, summary or product description: material for a person to correct, not material that should be sent as-is.

The common feature is that AI receives a defined input and produces an output that somebody checks, or that feeds a verifiable next step.

The value is in the context, not the model

An AI assistant without access to company data can only respond in general terms. The difference between a useful tool and a demo is almost always the material available to it and the actions it can take.

That requires preliminary work that is rarely exciting: finding where information lives, understanding its condition, deciding what is current, what may be exposed and what must remain private. A disorganised knowledge base produces disorganised answers, and no model compensates for a bad source.

The second element is the ability to act. An assistant that can check an order, search an archive or create a draft inside a management system is useful in a different way from one that can only write text. This is where AI and automation stop being separate subjects: the model interprets and chooses the step, while the automated flow performs it predictably.

The most reliable way to use AI in a business process is to give it the ambiguous part (understanding what a person is asking) and leave the part that must always work the same way to a deterministic system.

The problem with errors, plainly stated

These systems produce plausible answers, and plausible does not mean correct. They can confidently state something untrue, cite a procedure that does not exist or apply a rule from one case to another. Choosing a better provider does not remove this characteristic; the system has to be designed around it.

There are three practical consequences.

The first is where AI belongs. When an error is recoverable (a draft, a proposed category, a suggestion), the risk can be acceptable. When an error reaches a customer or changes an amount, human review is required.

The second concerns sources. An assistant that cites the document behind its answer lets the reader check it in seconds. One that only gives an answer demands blind trust.

The third is handover to a person. A system that communicates with customers needs an explicit, quick way to transfer the conversation, and must use it when uncertain. An assistant that keeps insisting is worse than a contact form.

Data, privacy and responsibility

Using AI with business material means deciding where that data goes. Several questions must be answered before starting: which information may be sent to an external service, which must stay inside the organisation, what is retained and for how long, and who can access conversation histories.

When personal data is involved, GDPR obligations apply as they do to any other processing, including provider selection and the information given to data subjects. This is not a legal issue to leave until the project is finished. It changes technical decisions, and changing them later is expensive.

When AI is the wrong choice

When the rule is already written. If the criterion is “orders over five hundred euros require a manager’s approval”, that is a condition, not an interpretation problem. A line of logic is faster, cheaper, reliable and easy to explain.

When the result must be identical every time. Calculations, reconciliations and documents generated from known fields call for a deterministic system. Variability is a defect here.

When the data does not exist. If the necessary information is missing, scattered or wrong, AI cannot reconstruct it. The right project at that point is to fix the data.

When the volume is low. Work that happens only a few times per month rarely justifies the maintenance, monitoring and controls required for a dependable system.

When it is there only to say that it is there. This is the most expensive case because the project has no criterion for success and therefore never ends.

A practical way to choose

For any process, three questions in order usually reveal the direction.

First: can the decision criterion be written as a rule? If yes, it needs software, not AI.

Second: does the work move or transform data that is already structured? If yes, it needs automation.

Third: does the work require interpreting natural language or producing text? This is where AI has a real advantage. Use it only for that part and leave the rest to systems that already perform it well.

One tool, with one job

AI is neither a revolution to adopt wholesale nor a trend to ignore. It is a concrete new capability, reading and producing language with an ability software did not previously have, that becomes useful when connected to a real process, real data, a verifiable objective and a person who remains accountable for the result.

The alternative, adding it to a product simply because it is AI, is visible. Especially to the people expected to use it every day.

Go deeperAI & Automation