Teaching exposed the real literacy gap
Designing and delivering an AI literacy course to 20 participants forced me to answer a deceptively difficult question: what does somebody now need to understand before they can use AI responsibly and productively at work?
The answer was much broader than a catalogue of tools. Interfaces will change, model rankings will change and many of today's products will disappear. Durable literacy has to sit beneath those changes. People need to understand how models behave, how work is framed, how evidence is evaluated and where accountability remains human.
Using AI is not the same as understanding it
A person can produce a polished response without understanding why it is strong, where it may be wrong or what information shaped it. Fluency with a chat interface can therefore create confidence faster than judgement.
Mandatory knowledge now begins with a practical mental model. Generative AI predicts useful outputs from patterns and context. It does not possess organisational truth, understand intent in the human sense or verify every statement automatically. Its responses are probabilistic, sensitive to framing and capable of sounding certain when the underlying answer is weak.
The same model can produce completely different work
One of the clearest lessons in the course was the variation produced by the same model. Two people can use identical technology and receive results that differ enormously in relevance, accuracy and usefulness. The difference is often not intelligence in the model. It is the quality of the task definition and the context supplied around it.
A vague request asks the model to infer the objective, audience, constraints, source material and definition of quality. A well engineered request makes those variables explicit. The model has less room to solve the wrong problem convincingly.
What prompt engineering actually includes
Prompt engineering is not the search for clever phrases. It is the disciplined specification of a task. A useful prompt defines the objective, the role the model should perform, the audience, the input material, the required output, the constraints and the standard against which the result will be judged.
It may include examples, counterexamples, a required structure, definitions, prohibited assumptions and instructions for handling uncertainty. Complex work benefits from decomposition: analyse first, identify missing evidence, perform the task and then review the result against explicit criteria.
Good prompts also create the right failure behaviour. The model should know when to ask a question, mark something as unknown, cite a source or refuse to infer beyond the available evidence.
Context engineering is the larger discipline
A strong prompt cannot compensate for missing or badly selected context. Context engineering determines what the model can see and how that information is assembled for the task. It includes source documents, data, conversation state, user preferences, organisational policies, examples, tool results and the instructions that govern their priority.
In a production system, context may be retrieved from a knowledge base, generated by software tools or inherited from earlier interactions. It must be current, relevant, permissioned and small enough for the model to distinguish signal from noise. More context is not automatically better context.
This makes information architecture an AI capability. Document quality, metadata, ownership, access control and retrieval design directly influence the quality of the model's work. Organisations with disordered knowledge do not escape that disorder by adding AI. They expose it through a new interface.
Evaluation is part of literacy
People need a method for deciding whether an output is fit for purpose. That means separating fluency from correctness and judging the result against the risk of the task. A brainstorming suggestion, customer communication, financial calculation and regulated decision should not share the same approval threshold.
Evaluation can include factual verification, source checking, calculations, policy checks, comparison with expected examples and review by a responsible specialist. Repeated use cases need reusable test sets rather than occasional subjective approval. When a model or prompt changes, those cases reveal whether quality improved or simply changed shape.
AI cannot sit beside transformation
AI crosses every part of digital transformation: customer journeys, operations, data, workforce capability, technology architecture, risk, governance and product design. Treating it as an independent innovation initiative separates the technology from the processes and decisions it is supposed to improve.
That separation creates a predictable pattern. Leaders request AI use cases to demonstrate adoption. Teams search for visible applications. Pilots are built and tested against artificial conditions. Many then fail because the use case was selected for its ability to showcase AI rather than its ability to remove friction, improve a decision or increase productive capacity.
Start with work, not use cases
The better starting point is the work itself. Where do people wait? Which decisions repeatedly lack information? Where is expertise consumed by repetitive interpretation? Which handovers introduce errors? What work is valuable but does not happen because capacity is limited?
Only then should the organisation decide whether the answer is generative AI, predictive models, conventional automation, better data, process redesign or no technology at all. AI earns its place when its characteristics fit the problem. It should not be forced into a process merely because an adoption target exists.
AI changes the control model
Conventional software is expected to produce deterministic behaviour from defined rules. Generative systems introduce variable outputs, model updates and failure modes that are harder to enumerate in advance. Governance therefore has to move beyond project approval and include continuous evaluation, monitoring and ownership.
Teams need to know which data may enter a model, where that data is processed, how outputs are stored, which actions require human approval and how incidents are reported. They also need protection against prompt injection, inappropriate tool use, excessive permissions and automation that turns a model error into an operational action.
Augmentation is a design decision
Productivity does not come from placing a chatbot next to every employee. It comes from redesigning the workflow so that the model performs the parts it can do well while people retain the judgement, accountability and relationships that matter.
Sometimes the right pattern is assistance: summarise, compare, draft or surface evidence. Sometimes it is controlled automation with validation and approval gates. In higher risk work, the model may only prepare a recommendation for a qualified person. The design should reflect consequence, reversibility and confidence rather than enthusiasm for autonomy.
The minimum standard has moved
AI literacy is becoming part of ordinary professional competence. People do not need to become machine learning engineers, but they do need to frame tasks, protect information, construct useful context, recognise weak evidence, evaluate outputs and understand when not to use AI.
For leaders, the standard is higher. They must connect AI investment to transformation outcomes, create governance that enables responsible experimentation and resist adoption theatre. The objective is not to produce the largest number of AI pilots. It is to create measurable improvements in how the organisation decides, serves, builds and operates.
Teaching the course reinforced my view that the most important AI capability is not access to a frontier model. Access is becoming common. The advantage comes from combining domain knowledge, quality context, disciplined evaluation and the authority to change the work around the technology.