AI decision rights for Slovak SMEs

A new European Commission survey published on 14 October 2026 shows that 42% of the interviewed European SMEs and mid-caps already use artificial intelligence or algorithmic management, mostly in a limited or experimental way; another 21% plan to do so. The important management question is therefore no longer simply whether a company uses AI. It is whether everyone knows when a system may suggest, who must decide and what evidence makes that decision defensible.

These systems are not confined to spectacular applications. Among surveyed users, 47% apply them in customer service and client relations, 36% in monitoring performance and productivity, and 30% in logistics, planning or staff scheduling. A recommendation can therefore affect a customer response, a shift, a target, a workload or a person's prospects before the company has called it an “AI decision”.

The survey contains optimism and a warning

Three quarters of the surveyed companies expect AI and algorithmic management to improve job quality. Of the businesses using these systems for decisions, 52% say the technology suggests while people decide; overall, 90% report that people make the final decision and only 2% describe decisions as fully automated. At the same time, 49% do not trust AI or automated systems for staff-related decisions. Legal uncertainty, investment costs and lack of expertise are similarly prominent barriers.

Those results are not contradictory. Companies can see operational value and still recognise that a technically plausible output is not automatically a fair, accurate or context-aware decision. The Joint Research Centre reported in the same week that AI adoption across EU strategic sectors rose from 7% in 2021 to 18% in 2025, while deeper value is associated with organisational change and integration into real processes—not merely adding an accessible tool.

Start with a decision inventory, not an AI inventory

A software list tells management what has been purchased. A decision inventory reveals where authority is changing. Ask each function to identify recurring decisions influenced by a score, forecast, recommendation, ranking or generated response. For every item, record:

the decision and the person or group affected;
the data used and where it originates;
whether the system informs, recommends, prioritises or acts;
the named decision owner and an available substitute;
the possible operational, financial, legal or human impact;
how an error can be noticed, challenged and corrected.

This exercise often finds “quiet automation”: an employee schedule accepted without review, a lead score that decides who receives attention, a productivity dashboard treated as a performance verdict, or a drafted customer answer sent because it sounds confident. None requires panic. Each requires an explicit boundary.

Use three decision classes

A proportionate model keeps controls practical. Classify decisions by consequence rather than by technical complexity:

Low-impact examples may include categorising an internal request or proposing a meeting slot when the action is easy to undo. Medium-impact examples include prioritising sales enquiries, suggesting stock replenishment or drafting a customer remedy. High-impact examples include hiring, dismissal, pay, disciplinary action, access to essential services, significant pricing decisions, safety-critical scheduling or any decision with a comparable effect on a person.

The class determines the minimum control: approval, explanation, data check, second reviewer, log or appeal route. It also prevents the common mistake of applying an elaborate committee to every minor use while leaving a consequential dashboard effectively unsupervised.

Make human review meaningful

A human clicking “approve” is not necessarily exercising judgement. Meaningful review requires enough time, information and authority to reach a different conclusion. The reviewer should understand what the output represents, what it omits, how current the inputs are and which indicators should trigger escalation.

For a staff schedule, that might mean seeing contractual limits, qualifications, approved absences and safety constraints—not only an efficiency score. For lead prioritisation, it may mean checking whether past data systematically favoured one market or customer type. For performance assessment, managers should examine the work context and avoid turning a proxy such as response time into a complete judgement of contribution.

Require the reviewer to state a short reason for high-impact decisions rather than merely confirming the system. Track how often people override recommendations. A zero-override rate is not proof of perfect accuracy; it may indicate automation bias, weak authority or a review screen that hides the necessary context.

Give affected people a workable route to question the result

EU-OSHA stresses a human-centred approach, worker participation and a sound occupational-safety-and-health system when AI-based worker management is introduced. Participation is practical risk control: employees often know which data are incomplete, which exceptions occur in the real workflow and where a metric can create counterproductive behaviour.

Before deployment, explain the purpose, data sources, expected benefit and limits in plain language. Invite representatives of the affected roles to test edge cases. After launch, provide a route to report a wrong input or harmful outcome without having to debate the mathematics of the model. Name who investigates, how quickly the person receives an answer and how the company corrects the underlying record.

Separate operational monitoring from employee judgement

Dashboards are useful for finding queues, delays and capacity constraints. They become risky when process measures silently become individual verdicts. A backlog may reflect demand, system downtime or upstream quality; a low number of completed cases may reflect complex assignments; a location signal may say nothing reliable about productive work.

Define in advance which metrics support process improvement and which may be used in an employment decision. Limit access, retention and secondary use. If personal data are involved, assess the GDPR implications and obtain specialist advice where necessary. Decisions based solely on automated processing can raise specific legal issues, and the AI Act also treats some employment-related uses as high risk. A generic “human in the loop” label is not a substitute for a case-specific assessment.

Buy evidence, not promises

For a vendor system, ask for documentation that supports your decision class: intended purpose, data requirements, known limitations, configuration options, change history, access controls, logging, export and deletion, human-override functions and incident support. Clarify who may retrain or alter the system and how the company will be told when behaviour changes.

Test the tool on representative cases before it influences live decisions. Include normal cases, rare exceptions and deliberately incomplete inputs. Compare results across relevant groups where fairness may be affected. Record the baseline against which benefit will be judged; otherwise a polished dashboard can become its own proof of success.

A practical 30-day governance sprint

Week 1—Find: identify the ten most consequential decisions currently influenced by automated recommendations, scores or generated content.
Week 2—Classify: assign low, medium or high impact; name the owner, required evidence, override authority and challenge route.
Week 3—Test: run representative and exceptional cases with affected employees; check data quality, explanations, workload and failure handling.
Week 4—Embed: add the rules to the real workflow, brief reviewers, define monitoring and schedule the first review after live use.

Measure the quality of decisions

Tool usage is not the outcome. Monitor a small set of signals: error and correction rates, overrides, appeals, time saved after review, distribution of outcomes, incidents, employee feedback and whether the operational objective improved. Review changes in vendor versions, data sources and business conditions. A model that performed acceptably last quarter may drift when products, staffing or customer behaviour change.

The objective is neither full automation nor ceremonial human approval. It is a company in which technology can accelerate analysis without making authority ambiguous. When decision classes, owners, evidence and challenge routes are explicit, an SME can use AI with more confidence—and stop asking individual employees to absorb risks that management has not defined.

For support with decision structures, responsible process design and practical implementation, see our Consulting services.

Sources and further reading

Facts, figures and current context were checked against the following official European sources on 18 October 2026. The decision classes, governance sprint, practical recommendations, conclusions and wording are original editorial work by Merkle s. r. o.

The Eurobarometer surveyed privately owned SMEs and mid-caps with 1 to 749 employees across all 27 EU countries; its figures describe the interviewed sample and should not be read as the adoption rate of an individual sector or company. This article offers general operational guidance and does not replace an assessment under employment, data-protection, occupational-safety or AI law, nor individual legal advice.