An automation proposal normally has a clear before and after:
- twelve minutes becomes three;
- a queue becomes a dashboard;
- a person drafts and the model classifies;
- five handoffs become one.
That is useful, but incomplete. The process may become faster while the job becomes more fragmented, more closely monitored, harder to learn, or full of exceptions that the project never counted.
The relevant question is not whether work should remain inefficient to preserve dignity. It is whether the organization improves the process while treating the people doing it as participants with knowledge, judgement, and legitimate interests—not as spare capacity around a system.
Measure the job that remains after automation, not only the task that disappears. Include the affected workers before the design hardens and give them a real route to challenge unsafe or degrading outcomes.
Why job redesign belongs in the business case
Generative AI usually affects tasks before it replaces whole occupations. The International Labour Organization’s 2025 update on generative AI and jobs says transformation is the more likely effect for most exposed occupations because human input remains necessary.
Transformation can improve a job. It can remove copying, searching, repetitive formatting, hazardous inspection, or routine queue work. It can also concentrate the remaining day into angry customers, ambiguous exceptions, verification, and responsibility for errors generated elsewhere.
The OECD’s current AI and work evidence overview reports that training and worker consultation are associated with better worker outcomes. These sources were rechecked on 29 July 2026.
This is not a prediction about a particular workplace. It is a reason to put job quality and participation into the project design rather than treat them as communication after the technical decision.
Start with the whole job
Map the job before selecting the tool:
| Work element | Questions |
|---|---|
| Routine production | What repeated output consumes time? |
| Judgement | Where does context change the correct action? |
| Relationships | Which work builds trust with customers or colleagues? |
| Learning | How does a newcomer develop competence? |
| Coordination | Who notices conflicts and joins information across teams? |
| Exceptions | Which unusual cases require investigation? |
| Accountability | Who explains and repairs a wrong outcome? |
| Recovery | What happens when systems or data are unavailable? |
Do the map with people who perform the work. A process owner sees the intended workflow. Staff see the corrections, workarounds, emotional labour, and edge cases that make it function.
Assess six dimensions
1. Worker voice
Participation should be able to change the proposal.
Ask affected workers to identify:
- hidden steps and dependencies;
- cases where the written rule is insufficient;
- data the system will misinterpret;
- customers or colleagues likely to be disadvantaged;
- new work the proposal creates;
- monitoring that would feel disproportionate;
- skills the current task develops;
- conditions under which the workflow should stop.
Record what changed because of that input. A listening session after procurement is not design participation.
“Worker voice” takes different forms depending on company size and local obligations: direct involvement, elected representatives, works councils, unions, health and safety structures, or formal information and consultation. This article is not employment-law advice; verify the requirements that apply to your organization and jurisdiction.
2. Retained judgement and autonomy
Identify decisions the person still owns after automation.
A system may be officially “advisory” while its output becomes the default because:
- it appears first;
- rejecting it takes extra work;
- disagreement is measured;
- the reason is hidden;
- the employee lacks source access;
- or throughput targets leave no time to check.
For every model recommendation, define:
- evidence the worker can inspect;
- authority to edit or reject;
- whether rejection needs justification;
- whether approval is logged or scored;
- what happens when source and model conflict;
- how repeated system errors are escalated.
If staff remain accountable, they need the information, time, and authority required to exercise judgement.
3. Skill and progression
Some repetitive work is merely repetition. Some is how people learn.
First-pass reviews can teach what normal looks like. Drafting can develop structure and judgement. Routine customer contact can expose a new employee to the product and its failure modes.
Ask:
- Which capability did the removed task build?
- Who needs that capability to verify automated work?
- How will a newcomer acquire it now?
- Does automation remove the entry-level route into senior work?
- Which practice must remain unaided?
- Can rotation, supervised cases, simulation, or sampled manual work preserve it?
Do not claim that every displaced task will be replaced by “more meaningful work.” Name the actual work and ask the people doing it how it affects development.
4. Workload and exception burden
Automation can move work rather than remove it.
Measure:
- review time;
- correction time;
- queue monitoring;
- exception handling;
- data cleanup;
- customer explanation;
- incident response;
- maintenance and prompt changes;
- time lost when the tool is unavailable.
Also examine work intensity. If a system removes natural pauses and feeds a continuous queue, the same number of paid hours may become more demanding.
Compare the distribution, not only the total. A saving in one team can create invisible work in another. The automation retirement audit provides a fuller operating-cost test.
5. Monitoring and privacy
Automation often creates new data: prompts, activity logs, quality scores, timestamps, recordings, model evaluations, and override rates.
Before collection, define:
- the business purpose;
- the minimum data needed;
- who can access it;
- how long it is retained;
- whether it is used for individual performance decisions;
- how a worker can inspect or challenge it;
- what inference is prohibited;
- when the collection stops.
Do not quietly turn workflow observability into employee surveillance. System-quality metrics and individual performance management are different purposes and need explicit governance.
6. Accountability and appeal
When the workflow is wrong, somebody needs a route to challenge it.
Define:
- business owner;
- technical owner;
- final decision owner;
- incident route;
- correction authority;
- worker and customer appeal route where relevant;
- pause conditions;
- evidence retained for review;
- deadline for response.
“A human is in the loop” is not enough. The human needs capacity and authority. Human-in-the-loop design patterns explains the control choices.
Use a job-redesign impact assessment
Complete this before the pilot and again before scaling:
Workflow and affected roles:
Business problem:
People involved in the assessment:
Task removed or changed:
New review, exception, and maintenance work:
Judgement retained by workers:
Evidence and authority available to them:
Skill or progression path affected:
Replacement learning design:
Monitoring data collected:
Purpose, access, retention, and prohibited uses:
Expected time, quality, safety, or capacity benefit:
Expected workload and work-intensity effect:
Groups receiving benefit:
Groups carrying new burden:
Accountable owner:
Challenge and appeal route:
Stop conditions:
Pilot evidence and decision date:
Require a worker representative or affected employee to confirm that the before-state and new-work description are recognizable. Confirmation does not mean agreement with the decision; it prevents management from grading its own map as complete.
Decide: accept, redesign, pilot, or stop
Accept
Use when the task change is low consequence, the benefit is clear, workers retain appropriate control, and new workload or monitoring is limited and understood.
Redesign
Use when the outcome is useful but the job design is weak. Examples:
- restore source access;
- make rejection easy;
- remove individual activity scoring;
- add exception capacity;
- preserve sampled practice;
- move a customer-visible action behind approval;
- reduce the scope of collected data.
Pilot
Use when important effects remain uncertain. Set measures for both process and job:
- output quality and time;
- correction and exception rate;
- workload by role;
- worker confidence in checking output;
- incidents and near misses;
- skill and training gaps;
- override use and reasons;
- customer effect where relevant.
Publish the decision criteria before the pilot. The team AI adoption playbook gives the broader rollout sequence.
Stop
Stop when the business gain is marginal, safe operation requires permanent hidden labour, affected people cannot meaningfully verify the system, monitoring is disproportionate, or the redesign removes capabilities the organization still needs.
Stopping is a valid project result. It prevents a weak workflow from becoming infrastructure.
Communicate without pretending the trade-off disappeared
Tell the team:
- which task is changing;
- why;
- what has and has not been decided;
- what worker input changed;
- how jobs and workload may change;
- what is being measured;
- which data is collected;
- who can pause the workflow;
- when the decision will be reviewed.
What to tell your team when you automate includes a complete briefing sequence and the promises managers should avoid.
If the purpose includes headcount reduction or avoided hiring, state that in the business case. “Freeing people for higher-value work” is not an honest substitute for the actual staffing intention.
The honest limit
A framework cannot equalize power inside a workplace. It cannot make a redundancy harmless, make consultation meaningful by itself, or guarantee that leaders will act on inconvenient evidence.
It can make omissions visible before launch:
- people affected but not heard;
- judgement retained without authority;
- verification expected without time;
- skills removed without a new learning path;
- savings calculated without exception work;
- monitoring introduced without boundaries;
- accountability assigned without appeal.
Automation should remove avoidable work and improve outcomes. The standard is not whether every existing task survives. It is whether the redesigned work remains safe, intelligible, learnable, and open to challenge by the people expected to carry it.
