Everyone has AI.Not everyone has an outcome.

A leadership meeting can acquire a surprisingly expensive sentence: “Everyone is doing it.” Perhaps. But doing what? Experimenting with a chatbot, changing a department’s workflow and producing a measurable enterprise return are three different achievements.

McKinsey’s 2025 survey reports regular AI use in at least one business function among 88% of respondents. 39% attribute some enterprise EBIT impact to AI. Its much stricter high-performer group represents 6%: respondents reporting at least 5% EBIT impact and significant value. That does not mean only 6% benefit. [McKinsey]

These are definitions, not a funnel that allows us to calculate how many adopters “failed”. The distinction sounds pedantic until someone uses the wrong number to justify the next budget.

Eurostat’s 2025 data put AI use at 21.6% of Malta enterprises within its survey scope, compared with 20.0% across the EU. Malta’s rates are 17.9% for small enterprises, 30.7% for medium ones and 52.2% for large ones. [Eurostat]

The scope matters: at least ten workers, within covered activities. This is not a census of every sole trader, startup or regulated financial business. Nor is it directly comparable with McKinsey’s multinational respondent sample.

FreeMalta interpretation: start with the difference in operating capacity, not a national victory lap. A small business may need a deliberately narrower implementation: one reliable process, manageable oversight and an invoice that does not require its own financing round.

Deloitte’s 2026 edition reports productivity and efficiency gains for 66% of respondents, but revenue growth for 20%. Another 74% hope for revenue growth in future. Expectations are not realised returns. [Deloitte]

Consider a hypothetical accounts team. AI extracts supplier-invoice fields quickly. Someone then retypes them because the accounting system has not been connected. Another person checks the same details because approval responsibilities are unclear. The extraction became faster. The process did not necessarily become cheaper.

Draw the route from incoming information to the final decision. Mark every place a person waits, copies, reviews or asks for clarification. That is where the business case lives. A model should earn its place in that route, rather than become an extra stop on it.

KPMG’s September 2026 Global AI Pulse finds a formal AI harness layer in 55% of surveyed organisations, rising to 86% among those reporting established ROI. Only 12% consistently assess AI value against cost. These associations do not prove that any one control creates returns. [KPMG]

In practical terms, a harness is the machinery around the model: permissions, tool access, logs, tests and boundaries. It determines whether an impressive response can become a dependable business action.

A Malta service firm does not need to copy a multinational’s architecture. It does need to know which records the system may read, which actions require approval, where errors are recorded and how to recover when a provider changes. Without those answers, “automation” can be a charming name for an undocumented dependency.

Before deployment, record the current cost and quality of the complete process. Include ordinary cases and awkward ones. Measure time per completed case, rework, error severity and the queue left waiting at the end.

After deployment, add model usage, integration maintenance, human review and exception handling. A saved hour has a different value depending on whether it becomes additional billable work, less overtime or simply an empty hour that nobody can redeploy.

Protect the quality measure. A process that closes cases faster by passing errors downstream has not become more productive; it has outsourced the inconvenience. Agree in advance who may pause the system and what would trigger that decision.

Finally, choose a comparison that can teach you something. Similar cases handled through the old and new workflow are more useful than last month versus this month when demand, staffing and case complexity all changed.

Our suggested first cycle is intentionally unglamorous. Choose one recurring task with a clear owner, accessible data and a reversible failure mode. Establish the baseline. Run a supervised pilot. Review exceptions before widening access.

If the pilot works, document the operating procedure and the economics. If it does not, distinguish model limitations from bad inputs, integration gaps and an ill-chosen use case. Those require different remedies. Another subscription is not a universal solvent.

For Malta’s smaller teams, the useful ambition is not “AI everywhere”. It is a system that does a specific job well enough to be trusted, cheaply enough to be kept and transparently enough to be improved. Scale follows evidence. Preferably in that order.