Ireland’s sector AI plan needs pilots that can end in “stop”

Ireland’s Department of Enterprise, Tourism and Employment published A Sectoral Study on Artificial Intelligence in Enterprise in Ireland on 13 October. The 291-page Accenture study covers manufacturing, tourism, construction and ICT services. It maps more than 500 global use cases, reports more than 30 stakeholder interviews, 272 online consultation responses and sector workshops.

The evidence base is broad but deliberately heterogeneous. The report draws on the CSO enterprise technology survey, business demography, a credit-demand survey, European indicators, interviews, consultations and workshops. It says these sources are interpreted separately rather than merged into a synthetic maturity score because populations, questions and methods differ.

That boundary matters. The principal enterprise technology survey excludes micro firms, even though the report says 92.6% of active Irish enterprises had fewer than ten people engaged in 2023. In construction, micro enterprises represented 97% of firms. Approximate margins of error use simple-random-sampling assumptions because respondent weights and full sample-design parameters were unavailable to the research team. The sector recommendations therefore should not be read as precise economy-wide causal estimates.

The report’s strongest operational proposal is simpler than a national ranking. Workshop participants called for diagnostics organised around a business decision, pilots that establish a baseline before work begins and an explicit production decision at the end, plus a casebook that records failed assumptions rather than only successes. Tourism needs short support tied to live tasks; small engineering firms cannot retain every specialist; manufacturing and construction have different data and operational constraints.

Give every supported pilot an exit record

For each intervention, define the unit before funding: one process, site or role. Record the current cycle time, error or rework, service quality, labour input, data availability, safety constraints and who owns the decision. Choose one comparable outcome and one harm or distribution measure. Do not use “AI adoption” as the outcome.

At the end, require one of three signed decisions: continue to production, change the design and retest, or stop. A stop decision should state which assumption failed—data, integration, economics, skills, safety, demand or supplier capability. Publish aggregated failure patterns without exposing commercially sensitive details. That creates reusable evidence for firms that need to know cost, management time and prerequisites before investing.

Keep sector and firm-size denominators visible. Report micro firms separately when the survey supports them, and do not splice figures from surveys with different questions into a trend. For training, count supervised task performance and escalation behaviour, not only course attendance. For vendors, disclose whether the use case evidence comes from a comparable Irish firm or an international example.

This angle is distinct from the existing Irish worker survey: that evidence mapped unequal access and training among employees. The new study asks how enterprises in four sectors can move from use cases to normal operation. The decision test is whether public support produces a documented, reproducible production choice. Link the capability evidence to causal inference, while preserving “stop” as a successful learning outcome.