Introducing artificial intelligence into a workplace is often described as a technology project. Leaders select a tool, approve a pilot and ask employees to adopt it. Yet the harder questions are not technical. They concern purpose, authority, work quality and trust.

Who decides which tasks should change? Who carries the risk when an AI-assisted decision is wrong? Can an employee challenge the system without being treated as resistant to innovation? Will efficiency gains create time for better work, or simply raise targets?

These are leadership questions. A human-centred approach therefore begins with the design of work, not the purchase of software.

The evidence supports this shift in emphasis. The International Labour Organization's 2025 global assessment estimates that one in four workers is in an occupation with some exposure to generative AI. It also concludes that transformation is more likely than wholesale job replacement because most exposed work still requires human input. The ILO consequently emphasises managing the transition through social dialogue rather than treating workers as passive recipients of change.

For leaders in Pakistan, this matters. Organisations may face uneven digital confidence, language barriers, limited training time, infrastructure constraints and strong workplace hierarchies. These conditions should not be framed as employee deficits. They are design variables that determine whether AI adoption broadens capability or concentrates advantage.

Human-centred leadership can be organised around five practical shifts.

1. Move from tool acquisition to a clear human purpose

An organisation should not begin with, “Where can we use AI?” It should begin with, “What problem in work are we trying to solve, and for whom?”

A useful purpose statement names the people affected, the current difficulty and the desired improvement. For example: reduce the time faculty spend formatting routine student communications so they can devote more attention to feedback and mentoring. This is better than a generic ambition to “automate academic administration.”

Purpose creates a boundary. It helps leaders distinguish valuable assistance from automation that merely shifts hidden work to employees or customers. It also makes evaluation possible. If the intended benefit was better feedback, success cannot be measured only by faster output.

2. Move from announcing change to designing it with employees

Employees understand the exceptions, workarounds and judgement calls inside a process. When they are consulted only after a system has been selected, leaders lose operational knowledge and increase the chance of avoidable resistance.

Participation does not mean giving every employee veto power over every tool. It means creating credible opportunities to shape the problem definition, pilot boundaries, failure tests and review criteria. Leaders can ask:

  • Which part of this workflow creates avoidable burden?
  • Where does context matter more than speed?
  • What error would cause the most harm?
  • What evidence would make this tool useful and trustworthy?

In a hierarchical organisation, leaders should provide channels that do not depend entirely on speaking openly in front of senior managers. Small-group sessions, anonymous feedback and role-based interviews can reveal concerns that a town hall may suppress.

3. Move from a “human in the loop” slogan to meaningful authority

Human oversight is often promised but poorly designed. A person cannot exercise meaningful oversight if they lack the time, information, competence or authority to question an output.

Leaders should identify the decisions that require human review, the evidence the reviewer will see, and the action available when confidence is low. For consequential uses, there should also be a route for affected people to seek explanation or correction.

This aligns with the OECD AI Principles, which call for human rights, fairness, transparency, robustness and accountability across the AI lifecycle. It also reflects the NIST AI Risk Management Framework's emphasis on incorporating trustworthiness into the design, development, use and evaluation of AI systems. Neither reference removes the need for contextual judgement. They help leaders ask better questions about responsibility.

4. Move from adoption metrics to evidence about work quality

Logins, prompts and hours saved can indicate use, but they do not demonstrate value. A human-centred review examines several dimensions together:

  • quality and accuracy of the work;
  • distribution of benefits and burdens across roles;
  • frequency and severity of errors;
  • employee confidence and ability to challenge outputs;
  • effects on learning, discretion and customer experience;
  • complaints, corrections and near misses.

These measures prevent a common mistake: counting visible efficiency while ignoring invisible checking, rework or anxiety. They also keep leaders from generalising too quickly from a successful demonstration to a complex operational environment.

5. Move from one-time training to adaptive capability

AI literacy is not a single software tutorial. Employees need practice in framing tasks, protecting information, testing claims, recognising uncertainty and deciding when not to use a tool. Managers need additional capability to redesign workflows, examine evidence and respond to concerns without punishing candour.

Because tools and work practices will continue to change, organisations need a learning loop. A monthly review can ask what assumptions failed, which safeguards were used, what new risks emerged and whether the pilot should expand, change or stop. This turns adoption into responsible experimentation rather than an irreversible rollout.

A clearly hypothetical Pakistan-based example

Consider a hypothetical Karachi-based services organisation testing AI-generated first drafts of customer responses. A technology-led rollout might measure the percentage of replies produced with the tool and average response time.

A human-centred pilot would go further. Frontline employees would identify sensitive categories that require manual handling. Reviewers would have explicit authority to reject generated text without a productivity penalty. The organisation would compare accuracy, rework, escalation and customer complaints across languages and service types. Employees would receive protected learning time, and the team would review the evidence before expanding the system.

The difference is not slower innovation. It is better-governed learning.

A 30-day leadership agenda

Leaders can begin without waiting for a complete enterprise policy.

Week 1: Define purpose and boundaries. Map one workflow, the people affected, the decision at stake and the harm that must be avoided.

Week 2: Listen before designing. Gather employee and user insight through more than one channel. Identify exceptions, language needs and unequal access.

Week 3: Run a bounded pilot. Specify meaningful human oversight, evidence requirements, an appeal path and conditions for pausing the pilot. The proposed HUMAN Test for AI-Enabled Work can support this review.

Week 4: Decide from evidence. Examine quality, burden, capability, incidents and distributional effects. Choose to expand, revise, watch or stop. Record what was learned.

Human-centred leadership does not reject AI. It refuses to treat adoption as complete when software is installed. Its real test is whether the organisation improves human judgement, distributes benefits fairly, keeps responsibility visible and learns before scaling.

Next step: Fiker Futures Academy can facilitate a leadership briefing or workflow-design session to help an organisation examine one proposed AI use through purpose, participation, oversight, accountability and future capability.

Authoritative sources

  • International Labour Organization. Generative AI and jobs: A 2025 update, published 20 May 2025. Checked 25 September 2026.
  • OECD.AI. OECD AI Principles, originally adopted in 2019 and updated in May 2024. Checked 25 September 2026.
  • National Institute of Standards and Technology. AI Risk Management Framework. The page states that AI RMF 1.0 is under revision; it is used here as a voluntary risk-management reference, not a compliance certificate. Checked 25 September 2026.

Qualification

The five leadership shifts, Pakistan-focused interpretation, hypothetical example and 30-day agenda are Dr. Salman Khatani's proposed analysis. They are not findings, standards or endorsements issued by the ILO, OECD or NIST. No client result, institutional adoption, ranking or validated impact is claimed.