South Africa is investing in AI skills at an important moment. Telkom has announced plans for a R100m AI institute, while new workplace programmes are offering professionals practical training in responsible AI use. These initiatives recognise the right problem: the country cannot capture AI’s benefits without building human capability.Yet training alone will not change how organisations work. An employee can complete a course, return to the office and find that no manager has identified where AI belongs, what data may be used, who checks the output or how success will be measured. The result is familiar: certificates accumulate while workflows stay the same or employees experiment quietly with unapproved tools.South Africa should therefore complement AI courses with workflow apprenticeships. Instead of teaching tools in isolation, each participant should redesign one real, recurring task under a manager’s supervision. A bank team might improve first-pass customer query triage. A mine could test maintenance report summaries. A municipality might draft routine public notices more efficiently. A retailer could analyse recurring inventory exceptions.The apprentice would map the current process, identify the limited step where AI could help, define prohibited data, test outputs against human work and document who remains accountable. The manager would approve the experiment, protect time for practice and decide whether the revised workflow should be adopted, changed or abandoned.This approach matters because South Africa’s AI challenge is not only technical. It is organisational and emotional. Employees may fear that explaining their work will help automate their jobs. Professionals may worry that using AI makes their expertise appear less valuable. Managers may hesitate to permit experimentation because they fear errors, reputational damage or regulatory exposure.Clear guardrails reduce that uncertainty. Organisations should specify approved tools, confidential data rules, required human review and responsibility for final decisions. Employees should be able to report mistakes and disclose unofficial AI use without automatic punishment. Psychological safety does not mean tolerating recklessness. It means making problems visible early enough to correct them.The country also needs better measures of success. Training attendance, tool licences and prompt counts do not establish value. Employers and public programmes should track outcomes such as time saved on a defined process, error reduction, turnaround time, service quality, employee confidence and whether the improved workflow still functions after three or six months.This would make AI development more inclusive. Large companies can build internal academies and governance teams, but smaller firms, municipalities and community organisations often cannot. Public-private initiatives should offer reusable workflow templates, manager coaching and regional peer networks so smaller employers can learn without exposing confidential information.South Africa’s recent withdrawal of a draft national AI policy after fictitious sources were found in its references offers a broader lesson. AI governance cannot rest on polished outputs alone. It requires verification, named accountability and processes that make human review unavoidable.The country is right to invest in AI skills. The next step is to connect those skills to daily work. Workflow apprenticeships would turn training from an event into an operating capability — one that improves productivity while preserving judgment, trust and accountability.• Dr Tsipursky is a behavioural scientist and author of ‘The Psychology of AI Adoption at Work: From Resistance to Results’.