Artificial intelligence in business: promises, limits and concrete uses for SMEs

Since ChatGPT launched at the end of 2022, artificial intelligence has left the laboratories and invited itself into every office conversation. You are promised that it will “change everything”. But behind the media excitement, a very concrete question arises for an SME manager: what can this technology really bring me, tomorrow morning, in running my company — and what will it not do in my place?

AI in plain terms: what are we really talking about?

Artificial intelligence refers to a set of technologies able to perform tasks that usually require human intelligence: understanding a text, recognising an image, analysing data or making a decision based on learned rules. It is not a new concept — spam filters, product recommendations and voice recognition have been part of it for years.

What changed at the end of 2022 was the mainstream arrival of generative AI: a family of AI able to produce original content — a text, a summary, an analysis — from a simple request phrased in natural language. It is this capability, embodied by conversational assistants, that triggered the current wave of enthusiasm.

For a manager, what matters is not understanding the technical machinery, but grasping one nuance: generative AI does not “know” in the human sense. It predicts the most likely continuation of a text from everything it has learned. That makes it remarkably useful for assisting, drafting and summarising — and structurally fallible as soon as precise factual truth is at stake.

A second distinction, often missing from the discussion, deserves to be made straight away, because it determines the type of project you will launch:

• Generative AI produces content: a text, a summary, a rewrite, an answer. It works on language, is used without configuration and is judged on the quality of its writing. It is the one everybody knows.
• Predictive AI writes nothing: it calculates a probability from series of figures — an expected sales volume, a risk of late payment, an abnormal variance. It is older, less spectacular, but often more profitable in management.

The two families have neither the same prerequisites nor the same risks. The first mainly requires clear rules of use; the second requires clean management history with enough depth. An SME is well advised to know which one it is aiming at before comparing offers, otherwise it risks buying a conversational assistant when what it needed was a forecast, or the other way round.

What AI does well for an SME

Rather than looking for “where to put AI”, start from the tasks that consume time without creating value. Here are the areas where, as of today, the contribution is tangible and accessible, even for a small organisation.

1. Assistance with writing and communication

This is the most immediate use. Writing a reply to a customer, rephrasing a payment reminder, producing a product sheet, translating a message or summarising a long set of meeting minutes: generative AI turns a blank page into a first draft in a few seconds. The gain is not to “replace” the writer, but to get them past the slowest step — starting — so they can concentrate on review and nuance.

2. Reading and analysing data

An SME accumulates data across its sales, purchasing, stock and treasury. AI helps to question it in plain language and to draw a first reading from it: spotting customers who no longer order, identifying products whose margin is deteriorating, comparing two periods. It is a natural extension of dashboards and business intelligence: where the dashboard shows the figure, AI helps explain it and formulate hypotheses.

3. Forecasting and anticipation

From history, some AI approaches estimate trends: a seasonal sales peak, an approaching stockout, treasury that is about to tighten. This is not a crystal ball, but useful alerts issued before the problem occurs — which has direct value for anyone who has to keep control of cash flow in uncertain times. This aspect deserves separate treatment: we cover it in detail in our article on the AI-augmented ERP and predictive management.

4. Intelligent automation

AI takes the automation of management processes further. Where a classic rule simply “executes if”, AI can handle less clear-cut cases: automatically sorting incoming e-mails, pre-filling a form from a document, routing a complaint to the right department. It absorbs part of the low-value repetitive work, freeing teams for what requires judgement.

5. An internal assistant on the company's knowledge

Finally, the use that increasingly interests managers: an assistant able to answer teams' questions from the company's own documents — procedures, catalogues, history, business data. “What is this customer's outstanding balance?”, “What is our returns procedure?”: instead of digging through folders, you question an assistant that knows the internal context. It is also the most demanding use in terms of confidentiality, since it assumes the AI has access to your most sensitive information; the subject is developed in our article devoted to the private AI chatbot in business.

The limits to know before getting started

A successful AI project starts with clear-headedness about what the technology does not do. Four limits deserve your full attention.

Quality depends on your data

AI does not invent reliable information out of disorder. If your business data is scattered across contradictory Excel files, AI will produce equally contradictory analyses. Clean, centralised and up-to-date data is the essential fuel. That is why an SME's first “AI” step often consists of collecting and making its data reliable in a single system, well before adding a layer of intelligence.

“Hallucinations”: misleading confidence

A generative AI can produce a false answer with complete assurance. It can invent a figure, a date, an article of law or a reference that does not exist. In an inspirational text, that is harmless. In a tax return, a contract or a payroll calculation, it is unacceptable. Any output intended for a decision or an official document must be checked by a competent human.

Data confidentiality

This is the most dangerous blind spot: pasting a customer file, a price list or payroll data into a public AI tool amounts to entrusting that information to a third-party service, with no guarantee as to how it will be reused. The answer is not to give up on AI, but to draw a clear line between what may leave the company and what must stay inside, relying on a private AI for the second case. The subject is too structural to be handled in a few lines: it is covered in a dedicated article on the private AI chatbot and data protection, alongside our advice on the confidentiality of business data.

The need for human control

AI bears no responsibility. It knows neither your customer context, nor your commercial stakes, nor your regulatory obligations. It proposes; the human decides. An organisation that blindly delegates its decisions to a tool exposes itself to costly mistakes and a loss of control. The right model is augmentation: the human stays at the controls, AI multiplies their capacity.

The hidden cost: verification time

Here is the limit that is discussed least, even though it often decides the real profitability of a use case. An AI that writes an e-mail in ten seconds has only saved time if reviewing it takes less than writing it by hand. On a simple, error-tolerant task, the arithmetic is largely favourable. On a technical task where every figure has to be re-checked, the gain melts away, or even reverses: you spend more time verifying a plausible output than you would producing it yourself.

The practical consequence is simple: before rolling out a use case, measure the net time, verification included, and not the generation time. That figure is what will tell you whether the tool deserves to be deployed to the whole team — and it is also what will stop you paying a subscription for productivity that only exists in the sales demonstration.

Promise versus reality: a table to see clearly

To tell the hype from the real contribution, here is a comparison of received ideas and what an SME can reasonably expect today.

The promise heard everywhere	The reality for an SME in 2023
“AI will replace my employees”	It automates tasks, not jobs. It frees up time for value-added work and customer relationships.
“You have to be a large company to benefit from it”	The tools are accessible and inexpensive. The real prerequisite is reliable data, not a big budget.
“AI always gives the right answer”	It can be wrong with total confidence. Any sensitive data must be checked by a human.
“I can entrust it with everything”	Not with a public tool: sensitive data requires a private, controlled solution.
“It is magic, it works all by itself”	AI is only useful when plugged into clean, centralised data and framed by clear rules.

What an SME really needs to get started

Many managers imagine that an AI project requires a substantial budget and in-house technical skills. That is less and less true — but four prerequisites do remain unavoidable, and none of them is technical.

A real problem, not a technology

The starting point is never “we want to do AI”, but “this task costs us too much time” or “we always discover this problem too late”. Put the pain point into a sentence anyone in the company can understand, then look at whether AI eases it. The projects that fail are almost always those that started with the tool and then looked for somewhere to plug it in.

Data you can look at without blushing

No model makes up for a customer database full of duplicates, or for sales entered from memory at the end of the month. This prerequisite is not a technical precondition, it is a matter of organisational discipline — and it is often the real work behind an AI ambition. The subject is covered step by step in our guide to collecting and making your company's data reliable.

Written rules of use, however short

Half a page is enough: what may be submitted to an external tool, what may not, who validates what before it goes to a customer or to the authorities, and to whom a doubtful result should be reported. Without that framework, every employee improvises their own — and the risk is no longer controlled, it is simply invisible.

An internal champion, not an expert

This is not about recruiting a technical profile, but about appointing someone who tests, notes what works, trains the others and reports the slip-ups. Without that role, use stays anecdotal: two or three curious people make use of it, the company gets nothing collective out of it, and the subscription ends up being cancelled without anyone being able to say whether it paid off.

Where to start in practice

There is no need to launch a major programme. Adopting AI, like any digital transformation of an SME, succeeds through small measurable steps. Here is a realistic approach.

1. Make your data reliable first. Centralise your business information in a single system. AI plugged into clean data is worth a thousand promises plugged into disorder.
2. Choose a single pilot task. Take a time-consuming, low-risk activity: writing e-mails, summarising minutes, a first reading of figures. Measure the time actually saved.
3. Set clear rules. Decide what may go through a public tool and what must stay internal. Make your teams aware of the risk of data leakage.
4. Keep the human in the loop. Any AI output intended for a customer, for the authorities or for a decision is reviewed and validated before use.
5. Broaden gradually. Once a use is mastered and profitable, extend it to another process. Value is built by accumulation, not by revolution.

One clarification on the second step, because that is where most experiments get bogged down: “measuring” does not mean asking the teams whether they found the tool pleasant. Set a baseline before you start — how long the task takes today, how many times a week it comes up — then measure again three or four weeks later, verification included. A modest but real figure is worth more than enthusiasm that fades in the first busy month.

Finally, be aware that a pilot may perfectly well end in abandonment, and that this is not a failure. Establishing that a task does not lend itself to AI saves you a pointless rollout and sharpens your judgement for the next use case. The companies that progress fastest are not those that succeed at everything, but those that decide quickly.

Key takeaways

Artificial intelligence is neither the miracle solution announced by some, nor the imagined threat feared by others. For an SME, it is a productivity lever whose value depends entirely on two conditions: the quality of its data and the maintenance of human control. The most solid uses today — assistance with writing, reading data, anticipation, automation, an internal assistant — are accessible and concrete, provided you start from a real problem rather than from the technology.

The real dividing line, for a company, is not “with or without AI”, but “with AI mastered or AI endured”. Centralising and making your data reliable, choosing an approach that respects confidentiality, and keeping the human as decision-maker: that is the foundation of an AI that serves the company instead of exposing it. To go further on data control, discover Swifto's ERP solution for SMEs, which lays exactly those foundations.

Frequently asked questions