Innovation & AI
AI can enter everywhere. The real decision is where to give it power.
As access to technology becomes commonplace, the advantage shifts to the quality of the decisions made around it.
AI has become easy to try. Choosing is becoming the real issue.
Companies have never had such easy access to artificial intelligence capabilities. Writing, analysing, searching, forecasting, classifying, recommending, conversing, triggering an action: the technical frontier is moving quickly. For a business leader, the problem is therefore changing in nature. It is no longer simply a matter of adopting AI. The challenge is to decide where it deserves to enter the business, what we are willing to entrust to it, and how to prove that it truly creates value.
of EU enterprises with 10 employees or more used AI in 2025.
for European SMEs covered by the statistic.
in Belgium, among the highest rates in the Union.
of SMEs using GenAI use it in their core activities.
The scene is composite. It does not describe one specific company, but brings together situations that have become commonplace.
A management meeting. A few leaders around the table. In the middle, a list of ideas.
An assistant that would prepare sales meetings. An AI capable of reading contracts. An agent for customer requests. Automation for invoices. A sales forecasting tool. Copilot for teams. A system capable of finding knowledge scattered across the company.
There is no shortage of ideas.
Yet one question is enough to change the meeting.
Which of these ideas genuinely deserves our money, our time and our attention?
This is no longer a discussion about what the technology can do. It is a discussion about the company. And that is probably where the difficult part begins.
In 2025, 20% of European enterprises with ten employees or more reported using at least one artificial intelligence technology, compared with 13.5% a year earlier. Among SMEs, the proportion reached 19%, compared with 55% in large enterprises. At 35%, Belgium was among the European countries where adoption was most widespread. [1]
Another survey, conducted by the OECD at the end of 2024 among more than 5,000 SMEs in Austria, Canada, Germany, Ireland, Japan, Korea and the United Kingdom, found an even higher figure for generative AI alone: nearly 31% of the companies surveyed reported using it. The two statistics measure neither exactly the same population nor exactly the same technologies and should therefore not be compared as though they came from the same thermometer. But they describe the same movement. Adoption is moving beyond the circle of pioneers. [2]
The most interesting detail in the OECD survey lies elsewhere. Among SMEs using generative AI, fewer than three in ten reported using it in their core activities, those that directly produce their main goods, services or revenue. Usage remained more concentrated on peripheral tasks, often simple or occasional. [2]
That may be where many companies are today. They are no longer discovering AI. They are still looking for where to put it.
And soon even that question will no longer be enough.
Because when a technology can theoretically intervene almost everywhere, the advantage no longer comes from being able to find a use case. It comes from being able to eliminate the bad ones.
The problem with AI in business may already no longer be a lack of ideas. It is the absence of sufficiently robust criteria for choosing between them.
The first choice is sometimes not to use AI.
Choosing an AI project first means choosing between problems, not between models.
The temptation is understandable. A technology appears. It is spectacular. Its entry cost is low. Competitors are talking about it. Employees are already experimenting. Vendors are adding AI features to almost every software product.
The question comes naturally: What could we do with it?
That may be the wrong first question.
Management rarely has infinite time, money and capacity for change. Choosing an AI project almost always means not doing something else.
The starting point should therefore be in the business, not in the technology.
A sales cycle that is too long. Information that is difficult to find. Administrative errors. An overloaded team. Expertise concentrated in two people. An inability to handle more cases without hiring. A degraded customer experience. A decision that consistently arrives too late.
The problem comes before the tool.
This nuance seems trivial until we discover that the smartest solution sometimes contains no artificial intelligence at all.
Removing an approval can be more effective than automating it. Connecting two software systems can be more robust than asking an agent to shuttle between them. Changing a form can prevent hundreds of corrections. A conventional business rule can be cheaper, more predictable and easier to audit than a generative model.
Innovation is not about using the maximum amount of technology available. It is about finding a better way to produce an outcome that matters.
AI is no exception to this rule.
Before automating a step, check whether it still needs to exist.
A conventional integration can be more robust and more predictable.
AI is not a substitute for a business rule or a well-designed form.
A good problem is not enough.
The official process and the real process are not always the same.
Now let us assume the problem is real.
An SME wants to reduce the time required to produce its sales proposals.
On paper, the idea of AI seems obvious. It could read the request, find the relevant information, prepare a response and propose a quotation.
Then we look at the work as it actually exists.
Some requests arrive through a form, others by email. Long-standing customers call the salesperson directly. Prices are in the ERP, except for certain negotiated terms. Discounts follow a grid, except for a few accounts considered strategic. A capacity check is required, but bypassed when the case is urgent. Some essential information still lives in the heads of experienced people.
The official process exists. The real process does too. They are not always the same.
Connecting AI to a vague process may simply allow the vagueness to circulate faster.
It would nevertheless be a mistake to conclude that a process must be perfectly standardised before AI can be used.
That would ignore part of what makes these technologies interesting in the first place.
Current models can work with natural language, imperfect documents and situations that are less structured than older automations. They can therefore sometimes handle variability that would previously have made automation too complex.
The real issue lies elsewhere.
The company must know how much ambiguity it wants to leave to the machine.
Reading a request is one thing. Detecting that information is missing is another. Proposing a price is another again. Granting a discount, committing capacity or sending an unusual offer without human validation completely changes the nature of the system.
At that point, we are no longer talking only about automation. We are talking about power.
At a certain level, we are no longer talking only about automation. We are talking about power.
Not every AI has the same right to act.
The question is not only what the system can do, but what the organisation is willing to delegate to it.
The conversation becomes much clearer when projects stop being classified by technology and are instead classified by autonomy.
An AI can observe. It reads, classifies, searches, extracts or detects.
It can assist. It summarises, prepares, translates, reformulates or produces a first draft.
It can recommend. The human remains the decision-maker but receives a proposal.
It can also act. Update an application, send a message, create an order, launch a sequence.
Finally, so-called agentic systems seek to pursue an objective by chaining several actions together with greater autonomy.
The boundary between these levels is anything but academic.
An error in the summary of an internal meeting does not carry the same weight as an incorrect price sent to a customer. A poor wording suggestion does not have the same consequence as a decision affecting recruitment, payment or a contractual obligation.
Technical progress is rapid, but autonomy still does not mean absolute reliability. The Stanford AI Index 2026 shows major gains across several benchmarks while also finding that agentic systems remain imperfect on structured tasks. In the organisations surveyed, general AI adoption reached 88% in 2025, while agent deployment was still in the single digits in almost every function studied. [3]
This gap tells us something important.
Companies are much further along in using AI than in delegating action to AI.
And they are probably right to distinguish the two.
The question is therefore not only whether the technology can perform a task. It is to decide what level of error, autonomy and consequence the organisation is prepared to accept.
Read, search, extract, detect.
Summarise, prepare, reformulate.
Propose; the human decides.
Write, send, create, trigger.
Chain actions towards an objective.
The best model is not necessarily the best decision.
A benchmark chooses a model. Strategy chooses a capability.
Problem, volume, data, acceptable level of error.
Native feature, model, API, RAG, business solution, agent.
Integration, costs, control, reversibility, accountability.
Choose a capability
Once the problem has been chosen and the level of autonomy defined, the market comes into play.
ChatGPT, Copilot, Gemini, Claude, AI features in CRM or ERP systems, specialised business solutions, APIs, internal document search engines, automation platforms, open models, agents.
The choice is dizzying.
It naturally pushes companies towards model comparisons. Which is the most powerful? Which benchmark puts one ahead of another? Who released the latest feature?
These questions are useful. Yet they describe only a fraction of the decision.
For a company, a slightly less capable model that is already integrated into its environment, properly governed and sufficiently reliable can be far more useful than a benchmark champion isolated from the rest of the system.
What the company is really choosing, therefore, is not just an artificial intelligence. It is choosing an architecture of use.
How does the data get there? What can the system access? What does the provider retain? Which software systems must it communicate with? How is an answer verified? What happens when it is uncertain? How much will the system cost when it leaves the pilot stage and processes thousands of operations? And above all: will it be possible to change tools without rebuilding the entire way of working?
The issue is not the logo displayed at the top of the screen. The issue is the dependency being built around it.
Govern use
While management prepares its policy, teams are already experimenting.
AI governance has a particular feature: in many companies, it begins after use has already started.
One employee summarises a contract in a chatbot. Another prepares a customer response. Someone in HR rewrites a job posting. A salesperson analyses a proposal. A manager summarises confidential notes.
No one necessarily waited for an executive committee to adopt an official policy.
That is precisely why governance cannot be limited to prohibition.
A rule that is too rigid pushes usage into the shadows. An absence of rules leaves each employee to decide alone on questions of data, confidentiality, quality or accountability that they may have no reason to master.
The European regulatory framework now makes the issue more concrete. Provisions of the AI Act concerning, among other things, definitions, certain prohibitions and AI literacy have applied since 2 February 2025. Since 2 August 2026, the transparency rules in Article 50 and several other provisions have entered into application, and enforcement has begun for the relevant rules. Some obligations relating to high-risk systems will follow later according to the planned timetable. [5]
For an SME, however, immediately turning this into a legal treatise would miss the essential point.
The first layer of governance is practical.
Teams need to know which tools they may use, which information must not be entered into them, in which situations an output must be checked, and who remains responsible for the decision.
They must also be able to report a problem. And the company must know how to stop what it has put into production.
The NIST voluntary framework formalises this logic around four functions: Govern, Map, Measure and Manage. It notably emphasises testing systems before deployment, monitoring them during operation, recovery capability and mechanisms for stopping or withdrawing them.
Behind the terminology lies a simple idea.
A deployed AI is never a finished project. [6]
It is a system that must continue to be observed.
Responsibilities, rules, skills and policy.
Use, context, people, data and consequences.
Tests, quality, errors, risks, performance.
Monitoring, correction, recovery, stopping or withdrawal.
Innovation begins where certainty ends.
The right to experiment has no value without the right to stop.
There is still one problem that procedures cannot solve.
No one can know in advance with certainty which AI uses will create the most value.
That is precisely why experimentation is necessary.
But an experiment is not a small deployment. It is an organised way of reducing uncertainty.
Suppose a sales team spends a great deal of time preparing for meetings.
The temptation is to install an AI assistant, let a few salespeople test it, then ask whether they like it.
The experiment is appealing. It teaches little.
A real hypothesis would be more demanding: We believe that an assistant capable of automatically preparing customer context can reduce preparation time without reducing the quality of meetings.
From there, the test changes in nature.
You need to know the situation before the tool. Choose a sufficiently small scope. Observe what happens. Compare. Understand the errors. Look at what users actually do, not only what they say they do.
Then decide.
Continue. Modify. Expand. Or stop.
That last possibility is essential.
Otherwise, the company does not have a portfolio of experiments. It gradually acquires a portfolio of costs.
A mature management team facing AI does not pretend to know the future. It can say: we do not yet know whether this will work well enough, but we know what we want to learn, how much we are prepared to invest to learn it, and what evidence will make us continue or stop.
That may be the most useful form of innovation governance today.
Neither absolute control nor fascination. A discipline of uncertainty.
Starting situation and hypothesis.
Controlled scope, duration and exposure.
Actual use, errors, observed effects.
Expand, modify or stop.
Saving time is not yet the same as saving money.
more problems solved per hour, on average, in the NBER study.
among novice or lower-performing agents.
of SMEs using GenAI report improved performance.
report having saved money.
AI-related productivity gains are not theoretical.
In a study of 5,179 customer support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond observed an average 14% increase in the number of problems resolved per hour after the introduction of a generative AI assistant. The effect reached 34% among novice or lower-performing agents, while it was small among the most experienced. [4]
That is an impressive result.
It is also an excellent example of how studies should not be read.
It does not mean that ‘AI increases productivity by 14%’.
It means that a specific tool, applied to a specific activity, in a specific environment, using a specific measure, produced that result.
Context is part of the result.
SMEs surveyed by the OECD also report benefits. Among those using generative AI, 65% say it improved employee performance, 45% say it enabled them to save money, around 35% say it enabled them to carry out new tasks or offer new products or services, and 26% say it contributed to increasing revenue. The OECD nevertheless specifies that its survey does not measure the magnitude of these effects.
That nuance matters enormously. [2]
‘AI saves us time’ is not yet a business case.
Imagine a salesperson saves five hours a week. What happens to those hours?
If that capacity is used to meet more prospects, prepare better proposals or reduce response times, the company can convert the gain into value.
If the person simply works under slightly less pressure, the effect also exists. It may be humanly desirable and operationally useful. But it is not the same economic outcome.
That is why the number of activated licences, prompts sent or even hours theoretically saved tells us little about the success of an AI strategy.
The interesting measure lies further downstream.
Lead time. Margin. Conversion. Quality. Capacity. Error. Cost to serve. Customer satisfaction. Revenue. Cash. Or simply an activity that was previously impossible with the available resources.
The indicator depends on the original problem. It should therefore exist before the company falls in love with the solution.
Released capacity has economic value only if the company knows what it wants to do with it.
Human work does not disappear. Its value shifts.
When part of execution is taken over, where do we want to move judgement, relationships and expertise?
The conversation about AI often returns to a spectacular question: Which jobs will disappear?
For an SME leader, the useful question is often much closer to home.
What will I now expect from this person?
If AI produces a first analysis, more value shifts towards the ability to understand and challenge it.
If it prepares the sales meeting, the human can spend more time understanding the customer, negotiating or building the relationship.
If it produces a first draft of a document, the skill no longer lies only in being able to write, but in the judgement required to detect what is wrong, incomplete or dangerous. [4]
The study by Brynjolfsson, Li and Raymond also suggests that assistance can spread some of the practices of the highest-performing workers to less experienced ones, potentially changing learning mechanisms themselves. [2]
The OECD observes the same paradox from another angle. Among SME users that experienced a skills shortage, 39% believe generative AI helped them compensate for it. At the same time, the survey concludes that more SMEs report an increase rather than a decrease in their need for highly skilled workers.
AI therefore does not necessarily reduce the need for expertise.
It can shift where expertise becomes valuable.
That shift deserves to be managed.
Because automation does not only change a process. It sometimes changes a job. And when a job changes, the question eventually reaches management.
of SME users that experienced a skills shortage say GenAI helped them compensate for it.
Six dimensions to bring the discussion back to a management decision.
A company does not need a fifty-page committee report to challenge an initial idea. It can start with six dimensions.
Value
What business problem do we actually want to reduce, or what opportunity do we want to create?
Work
Which part of the work do we want to transform, and do we understand well enough what happens there today?
Technology
Why is AI needed here, and why this solution rather than a simpler approach?
Autonomy
Does the system observe, assist, recommend or act? Where does the human take back control?
Governance
What data does it use? Who is accountable for the outcome? How is the system tested, monitored, corrected and, if necessary, stopped?
Impact
What evidence will determine what happens next: continue, expand, modify or abandon?
This grid is not a standalone scientific standard. It translates into an operational SME framework principles of value creation, work transformation, experimentation and risk management found notably in OECD work and in the NIST AI Risk Management Framework.
The test that often reveals more than the answer
Take a single AI project currently being considered in the company. Ask the people concerned to answer these six dimensions separately. Then compare.
If three leaders identify three different problems, the technology is probably not yet the issue.
If everyone knows the product but no one can explain why AI is necessary, the solution was chosen too early.
If the benefits are clear but no one knows how far the system may act, the work to be done is about autonomy.
If the experiment has no stopping condition, it may already no longer be an experiment.
And if all six answers are identical?
Then the project is not guaranteed to succeed. But at least the organisation knows what it is trying to prove.
The next advantage will not be access to AI.
Technology will continue to advance. New functions will enter everyday software. Activities that are experimental today will become routine. Some current promises will probably disappear as quickly as they appeared.
In that environment, being able to say ‘we use AI’ will soon have as little strategic value as being able to say ‘we use the internet’.
Access is becoming commonplace. Judgement cannot be downloaded.
You still need to know which problem deserves to be addressed. Understand the work well enough to decide what can be delegated. Choose a technology without creating unnecessary dependency. Grant autonomy in proportion to the consequences of an error. Allow teams to experiment without abandoning accountability. Measure value rather than activity. Accept stopping what does not work. And redesign human work when the machine takes over part of execution.
The strategic question may therefore not be: ‘What is our AI strategy?’
It can begin much more concretely.
Which decision are we prepared to let this technology make in our place, and why that one?
[1] Eurostat, Digitalisation in Europe - 2026 edition, ‘AI in business’. 20% EU, 19% SMEs, 55% large enterprises, 35% Belgium in 2025. ec.europa.eu/eurostat
[4] NBER, Brynjolfsson, Li & Raymond, Generative AI at Work, Working Paper 31161. 5,179 agents, +14% on average, +34% among novice and lower-performing agents. nber.org
[2] OECD, Generative AI and the SME Workforce: New Survey Evidence, 2025. Representative survey of more than 5,000 SMEs in 7 countries. oecd.org
[5] European Commission, Timeline for the Implementation of the EU AI Act, consolidated 2026 version. AI Act Service Desk
[3] Stanford HAI, AI Index Report 2026, Economy and Technical Performance. Organisational adoption and agent performance. hai.stanford.edu
[6] NIST, Artificial Intelligence Risk Management Framework, Govern, Map, Measure and Manage functions. airc.nist.gov
What use of AI is your company considering today without yet being able to state clearly what problem it solves, what power you want to give it, and what evidence would justify deploying it at a larger scale?
lucid.dtsc.be · lucid@dtsc.be
Written and created by Junior Cantos for LUCID by DTSC
LUCID · SME Management by DT Services & Consulting
11
The first 30-minute conversation is free. That's where it gets framed, with no commitment.
