AI changes tasks before it replaces jobs: a manager's guide to redesigning work

The question arrives in different forms, but the concern underneath it is usually the same.

An employee asks whether AI will take their job. A manager asks whether the team will need fewer people. A director asks which roles should be “automated.” Everyone is talking about jobs as if each one were a single switch that can be handed from a person to a machine.

That is rarely how work changes.

A job is a collection of tasks, relationships, decisions and responsibilities. Some parts may be easy to assist with AI. Others depend on context, trust, physical action or accountability. Even when an AI assistant produces an impressive draft, somebody still has to decide whether the draft belongs in the real world.

I would therefore begin with a better management question:

Which parts of this job could AI help with, what new work would that create, and what should remain firmly in human hands?

That question is less dramatic than predicting the end of a profession. It is also far more useful.

Exposure does not mean replacement

The International Labour Organization and Poland's National Research Institute assessed nearly 30,000 tasks in their 2025 global index of occupational exposure to generative AI. They estimated that one in four workers is in an occupation with some degree of exposure. Their central conclusion, however, was that transformation is more likely than complete replacement because most occupations still contain tasks requiring human involvement.[1]

This distinction matters. If software can help prepare meeting notes, that does not mean it can run the meeting, understand an unspoken disagreement, make the commercial decision and accept responsibility for the result.

Exposure is an estimate of how much the technology might affect tasks within an occupation. It is not an observed redundancy rate, a timetable or proof that every exposed task should be automated.

Leaders should be equally cautious with confident forecasts from either direction. “This role will disappear” and “AI will never affect this role” can both hide the work that management needs to examine now.

Start by unpacking the job

Consider a customer-service employee. The job may include:

• Reading new enquiries.
• Identifying the customer and account.
• Finding an order or service record.
• Recognising the customer's actual concern.
• Applying company policy.
• Drafting a response.
• Calming an upset person.
• Negotiating an exception.
• Recording what was agreed.
• Escalating a serious problem.

An AI assistant may classify an enquiry, retrieve approved information and prepare a draft. That could reduce routine searching and writing. It does not automatically know when a customer is vulnerable, when a policy exception is justified or when an apparently polite reply will make the situation worse.

The assistant also creates new tasks:

• Checking whether retrieved information is current.
• Reviewing the draft for invented claims or promises.
• Recording and investigating failures.
• Maintaining the source material.
• Deciding when to override the system.
• Explaining consequential decisions to customers.

Automation does not always remove work. Sometimes it moves the work from creation to checking, from routine handling to exception handling, or from frontline staff to the people maintaining the system.

REDESIGN THE TASKS, NOT THE JOB TITLE

One role becomes three different management decisions.

CUSTOMER-SERVICE ROLEEnquiry arrives → useful answer reaches the customer

AI ASSISTSRoutine preparationClassify enquiries · find approved information · prepare a draft

PEOPLE RETAINJudgement and responsibilityRead the situation · decide exceptions · make promises · approve

NEW WORK APPEARSChecking and stewardshipVerify results · maintain sources · investigate failures · improve

Use four lenses to redesign a role

For each important task, examine capability, consequence, context and responsibility.

1. Capability: can the assistant perform this task reliably enough?

Do not answer from a product demonstration. Test the task using representative examples, including incomplete information and difficult cases.

Research shows why. In a field experiment with 758 management consultants, AI improved speed and quality on tasks within its capabilities. For a task outside that boundary, participants using AI were more likely to produce an incorrect answer.[2] The researchers call this uneven boundary a “jagged technological frontier.”

The same system can therefore appear highly capable on Monday and confidently fail on Tuesday because the work changed in a way that was not obvious to the user.

2. Consequence: what happens when the result is wrong?

Errors do not carry equal weight.

A weak first draft of an internal announcement is easy to replace. An incorrect payment instruction, disciplinary recommendation or safety message may harm another person and expose the organisation to serious loss.

The greater the consequence, the stronger the case for expert review, restricted use or keeping the task entirely human.

3. Context: what knowledge does the task require?

Some tasks can be completed from information contained in one document. Others depend on customer history, informal agreements, changing market conditions and knowledge held by several employees.

AI may produce fluent output while missing the context that makes the answer useful. Managers should ask where the necessary information lives, whether the assistant is permitted to access it and who keeps it accurate.

4. Responsibility: who must answer for the outcome?

An AI tool cannot carry managerial accountability. Someone in the organisation must own the process, approve consequential decisions and respond when something goes wrong.

NIST's AI Risk Management Framework treats governance as a continuing responsibility across the AI lifecycle. It calls for defined roles, documented processes, training, measurement and management oversight that reflect the organisation's risk tolerance.[3]

If nobody can name the person accountable for an AI-supported process, the organisation has not redesigned the work. It has created a gap.

Decide whether AI suggests, prepares or acts

A useful role design separates three levels of responsibility.

A business does not have to aim for the third level to create value. Drafting a reliable response or preparing a weekly report can remove substantial assembly work while preserving human approval.

The temptation is to regard human review as a temporary inconvenience on the road to full autonomy. That assumption should be tested, not declared. In some processes, judgement and accountability are permanent parts of the design.

Different employees may benefit differently

AI's effects are not uniform across a team.

In a study involving 5,172 customer-support agents, access to an AI assistant increased issues resolved per hour by 15% on average. The gains were considerably larger for newer and less-skilled workers. The most experienced and highest-skilled workers saw small gains in speed and small declines in quality.[4] The researchers found evidence consistent with the assistant helping distribute practices associated with stronger performers.

A separate experiment involving 453 college-educated professionals found that ChatGPT reduced the average time taken on selected writing assignments by 40% and increased independently rated quality by 18%. Participants with weaker initial performance benefited more.[5]

These findings suggest opportunities for onboarding and support. They do not mean experienced employees have become unnecessary. Experienced people may be the ones who recognise exceptions, maintain standards, teach the system's users and notice when a plausible answer does not fit the situation.

Managers should measure results by experience level rather than relying only on a team average. A tool that helps new employees while creating extra checking work for senior staff has changed the distribution of work. That may still be worthwhile, but the full effect should be visible.

Protect the expertise the organisation still needs

When an assistant performs the first draft every time, employees may get fewer opportunities to learn how to create one. When it supplies the answer instantly, they may stop practising how to find and evaluate the source.

This is not an argument for preserving repetitive work simply because it is traditional. It is a reason to decide which capabilities the organisation cannot afford to lose.

Ask:

• Which skills help employees recognise a bad result?
• Which knowledge is needed when the system is unavailable?
• Who can handle a case that falls outside the normal process?
• How will new employees learn the underlying work?
• Are experienced employees becoming permanent error checkers?
• Does the redesigned role offer progression, or only exception handling?

Create opportunities for employees to practise core judgement without assistance. Rotate responsibility for reviewing difficult cases. Record why employees override recommendations. Use incidents as material for training rather than hiding them as embarrassing exceptions.

The goal is not to ensure that people can reproduce every automated step manually. It is to preserve the understanding required to supervise, challenge and improve the process.

Do not confuse adoption with performance

Managers sometimes measure an AI programme by licences activated, prompts submitted or weekly users. Those figures show activity. They do not show whether the work improved.

A target such as “every employee must use AI every day” invites performative use. The team will satisfy the metric even when a spreadsheet formula, a template or five minutes of concentrated thought would be better.

Measure the work instead:

• Time required, including checking and correction.
• Quality and error rates.
• Customer or colleague waiting time.
• Escalations and complaints.
• Employee confidence and workload.
• The number and type of overrides.
• Whether new employees reach competence sooner.
• Whether experienced employees spend more time on valuable exceptions or routine supervision.

A slower process with fewer serious errors may be the better design. A faster process that increases work intensity or weakens service may not be an improvement.

Involve employees before the role is redesigned

OECD workplace research found that training and worker consultation were associated with better reported outcomes for employees using AI. It also recorded concerns about job loss, personal data collection and automated employment decisions.[6]

Consultation should be specific. Do not ask, “How do you feel about AI?” and treat the meeting as complete.

Ask employees to help identify:

• The parts of the job that consume time without creating value.
• The exceptions that managers rarely see.
• The information needed for a correct decision.
• The situations where empathy or discretion matters.
• The mistakes that would be most damaging.
• The work employees would do if routine effort were reduced.
• The skills they want to retain and develop.

Then show how their input affected the design. Consultation without visible influence can feel like a presentation wearing a name badge.

Be careful when AI becomes the manager

Using AI to support employees is different from using systems to assign work, monitor behaviour, score performance or recommend employment decisions.

OECD research on algorithmic management found concerns about explainability, accountability, bias, worker awareness and the protection of employee wellbeing. Among firms using these systems in the countries surveyed, guidelines and worker consultation were among the governance measures managers reported most often.[7]

For Kenyan organisations, automated decisions involving personal data also require legal attention. Section 35 of the Data Protection Act gives people protections relating to decisions based solely on automated processing when those decisions produce legal or similarly significant effects, subject to specified exceptions and safeguards.[8] The Data Protection (General) Regulations identify certain automated decision-making, profiling and sensitive-data uses as high-risk processing that may require a data-protection impact assessment.[9]

This article is general business information, not legal advice. The important leadership principle is broader: the more a system can affect someone's livelihood, rights or opportunities, the less appropriate it is to hide responsibility behind a score.

A role-redesign conversation for managers

Choose one role and complete this exercise with the people who perform it.

Step 1: List the work

Write down the recurring tasks, exceptions, decisions and relationships. Do not begin with what the job description says. Begin with what people actually do.

Step 2: Mark the pain

Identify delays, repeated searching, duplication, avoidable errors and work that employees believe creates little value.

Step 3: Test assistance

Select one narrow task. Decide whether AI should suggest or prepare, and test representative cases without allowing automatic action.

Step 4: Name the human value

State what people contribute that the redesigned process must retain: judgement, empathy, accountability, negotiation, local knowledge, creativity or physical action.

Step 5: Count the new work

Record checking, correcting, maintaining information, handling exceptions and monitoring. If the work has merely moved, say so.

Step 6: Agree on development

Decide what employees need to learn and how they will continue practising essential skills.

Step 7: Measure and revisit

Compare the new process with the old one. Roles should be reviewed after the pilot because real use will reveal effects that the design workshop missed.

The manager's responsibility

AI may remove some routine effort, raise the standard of a first draft or help a newer employee find an answer. It may also create overconfidence, extra checking and new forms of monitoring. Which outcome appears depends partly on the technology and partly on management choices.

Leaders decide whether the goal is to improve service, increase capacity, reduce cost or change staffing. They decide whether employees are involved, whether training is real, whether risks are visible and whether a person remains accountable.

That is why the future of work is not only something that happens to an organisation. At the level of one role and one process, it is something management designs.

Do not begin by asking whether AI can replace a job. Take one job your organisation understands and ask which part should become easier, which part must remain human and how the employee doing it can become more capable—not less.

The next step is to turn that role-level thinking into a careful rollout. Read how to introduce AI into your business without disrupting your team.

Research and helpful links

1. International Labour Organization and NASK, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure,” 2025. Exposure estimates describe potential task-level effects, not observed job losses.
2. Fabrizio Dell'Acqua et al., “Navigating the Jagged Technological Frontier,” Organization Science, published online in 2026. The preregistered experiment involved 758 consultants.
3. National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework,” 2023, and “Generative Artificial Intelligence Profile,” 2024. NIST describes the framework as voluntary and is revising AI RMF 1.0.
4. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, “Generative AI at Work,” The Quarterly Journal of Economics, 2025. Results from customer support should not be assumed to apply to every role.
5. Shakked Noy and Whitney Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science, 2023. The study examined selected professional writing tasks.
6. Marguerita Lane, Morgan Williams and Stijn Broecke, “The Impact of AI on the Workplace,” OECD, 2023. The research covered finance and manufacturing in seven OECD countries; reported associations do not establish causation on their own.
7. OECD, “How Widespread Is Algorithmic Management in Workplaces?,” 2025, drawing on an employer survey of more than 6,000 mid-level managers in six countries.
8. Kenya Law, “Data Protection Act, 2019, Section 35,” revised edition.
9. Kenya Law, “Data Protection (General) Regulations, 2021, Regulation 49,” Legal Notice 263 of 2021.

Research checked on 12 September 2026. Employment, labour, privacy and AI rules differ by jurisdiction; obtain appropriate local advice before using AI for consequential workplace decisions.