AI and Automation in the Workplace

Start with augmentation, not automatic replacement. Most jobs are made up of different tasks, and each task has a different level of repetition, judgement, risk and human contact. AI might draft a first version, summarise a document, extract information or suggest patterns. A practitioner still checks accuracy, applies workplace knowledge and makes the final decision. This is augmentation: technology extends human capability. Full automation is more suitable when the task is stable, rules are clear, data is reliable, outcomes can be measured and errors have limited impact. Research on customer-support work found that a generative AI assistant improved productivity, with particularly strong gains for less experienced workers, showing how AI can share useful patterns while people remain responsible for the service (Brynjolfsson, Li and Raymond, 2025).

Practitioners should therefore break a process into tasks rather than asking whether an entire role can be automated. For each task, ask: What outcome is needed? What information is required? Where is judgement needed? What could go wrong? Who must approve the result? A sensible first use case is often a human-AI-human workflow. A person sets the purpose and provides approved information; AI produces a draft, classification or recommendation; then a person verifies, edits and decides. This creates a controlled way to learn where AI adds value before giving a system more autonomy. AI capability is uneven, so apparently similar tasks can produce very different results; pilots must test the actual work rather than assume the tool will transfer safely (Dell’Acqua et al., 2026).

Prompt engineering means giving a generative AI system instructions that help it produce a useful output. A practical workplace prompt should include six parts.

• First, state the task and intended outcome.
• Second, give relevant context, such as the process, audience or policy.
• Third, provide the input material the system is allowed to use.
• Fourth, specify the required output format, length and tone.
• Fifth, set constraints, including what the system must not assume and when it should state uncertainty.
• Sixth, define the quality check, for example asking for sources, assumptions or a checklist against agreed criteria.

OpenAI’s guidance similarly emphasises clear, specific instructions, relevant context and explicit output requirements (OpenAI, nd).

For example, ‘Write an email about the delay’ gives the system very little direction. A stronger prompt would explain who the email is for, why the delay has happened, which facts can be shared, the tone required, the word limit, the next step and anything the system must not invent. The output should still be checked. Prompt engineering improves the instruction; it does not guarantee that the response is correct. Practitioners should compare outputs with trusted evidence, test prompts on normal and unusual examples, record successful versions and update them when the task or source information changes.

Data matters because AI outputs depend on the information used to train, configure, prompt or ground the system. Workplace use may involve three forms of data: model training data, organisational reference data such as policies or product information, and operational input data such as forms, messages or transactions. Before using data, check whether it is accurate, complete, current, relevant, representative, lawful and secure. NIST warns that unsuitable or unrepresentative data and other quality problems can reduce AI trustworthiness (NIST, 2023). A polished prompt cannot repair missing facts, outdated guidance or biased records. Sensitive or personal information should only be used in approved tools and for an authorised purpose.

The reinvested time principle means that time saved should be planned as an organisational benefit rather than treated as an automatic gain. In practice, re-investable time is the gross time saved minus the time required for review, correction, exceptions, maintenance and governance. That remaining capacity can be directed towards work that creates greater value: speaking with customers, coaching colleagues, analysing causes, improving processes, assuring quality, learning new skills or resolving complex cases. This is a work-design decision, not only a technology measure (Parker and Grote, 2022).

To apply the principle, establish a baseline before the change: volume, time per task, quality, errors and user experience. Pilot the new method, measure the same indicators and calculate the usable time released. Then agree where that capacity will go, who owns the benefit and how its impact will be measured. Saving ten hours is not a successful outcome if five hours are lost to rework and the remaining five simply increase work pressure. The strongest workplace use of AI connects augmentation, good prompts, dependable data and deliberate reinvestment to better outcomes for people and the organisation.