The Role Rewrite: How to Redesign Jobs When AI Changes the Work

Most companies are introducing AI into jobs that were designed before the technology could perform any meaningful part of them.

That mismatch is now measurable. Deloitte's State of AI in the Enterprise 2026 surveyed 3,235 senior business and technology leaders and found that 84% had not redesigned jobs to accommodate AI. Access to approved AI tools, meanwhile, had risen by around half.

This helps explain why a tool can be popular without producing a visible commercial return. Employees complete familiar tasks more quickly, then use the saved time on fuller inboxes, additional meetings, better-presented reports or work that had already been waiting. The individual feels more productive. The organisation carries the same roles, cost and performance measures.

Nothing is wrong with the tool. The business has improved a task without deciding what should happen to the job around it.

AI does not make a role 30% smaller

Leadership teams often discuss AI at enterprise level, through strategy and investment, or at product level, through licences and applications. Jobs change somewhere between the two.

A job is a collection of tasks, decisions, relationships and accountabilities. AI rarely removes that collection in one clean piece. It reduces the effort required for certain tasks, changes the level of expertise needed for others, and creates new work in checking, escalation, instruction and tool management.

The job title usually remains untouched. So do the grade, reporting line, objectives, hiring brief and development path.

This creates a misleading calculation. If AI removes 30% of the effort from each of five positions, the spreadsheet shows capacity equal to 1.5 full-time roles. Operationally, the business has five people with small pockets of time scattered across different days and responsibilities. Those pockets cannot be redeployed like a complete vacancy.

Within weeks, the time is claimed by whatever is nearest: another report, a longer meeting, more responsive internal service, or a more polished version of existing work. The capacity has not vanished, but it has become commercially invisible.

For mid-market companies, this fragmentation matters. There are fewer layers in which to hide duplication, and individual roles are often broad. A finance manager may combine analysis, supplier queries, cash reporting, systems administration and team supervision. Automating two tasks does not automatically create a coherent smaller job. It creates a different bundle that someone needs to redesign.

The correct question is therefore not, "How much time does AI save?" It is, "What role should exist when this work is done differently?"

Redesign begins with the work as it is

Job descriptions are a poor source for role redesign. They tend to record the role that was approved, advertised or inherited, not the role a person performs on a difficult Tuesday.

Real work includes informal checks, customer exceptions, duplicated entries, favours for other teams, manual reconciliations and tasks that survive because nobody has revisited them. It also includes judgement that is easy to miss when the output appears routine.

Before changing a role, build a task-level view with the person doing it. Fifteen to thirty tasks is usually enough to expose the shape of the work without creating an exhaustive time-and-motion exercise.

For each task, record:

• its trigger and intended output;
• approximate monthly effort;
• frequency and variability;
• information and systems required;
• consequence of poor execution;
• degree of business-specific context needed; and
• whether an error would be obvious and reversible.

This last distinction separates work that can safely move from work that merely looks easy to automate. Drafting a routine internal summary is very different from producing a customer recommendation whose error may not be noticed until renewal.

Once the real task list is visible, the role can be rewritten through five moves.

The five-move Role Rewrite

Start with one role where AI is already changing the workload, or one that is about to be recruited. The objective is not to produce a new organisation chart. It is to prove that one job can be redesigned properly.

Move 1: Expose the task ledger

Break broad responsibilities into observable units of work.

"Manage customer onboarding" is too broad. The work beneath it might include requesting documentation, validating details, checking commercial terms, creating accounts, configuring access, sending joining instructions, conducting the first review and resolving incomplete cases.

Those tasks have different economics and risks. Some are rules-based. Some require customer judgement. Some exist because the underlying system is awkward. Lumping them together conceals the design choices.

Estimate time honestly, but do not chase false precision. The aim is to identify material patterns, not account for every minute. Ask the jobholder which tasks take longer than colleagues assume, where they keep private workarounds, and what they would stop doing if permission were granted.

Run the exercise with the current jobholder, not solely with their manager. Managers understand expected outcomes. Jobholders understand the queues, exceptions and compensating work that make those outcomes possible.

Move 2: Make four explicit choices

Place every task into one of four categories.

Automate. The system performs the task, while people monitor performance and handle defined exceptions. This is suitable where rules are stable, volumes justify the effort, inputs are reliable and mistakes can be found and corrected.

Assist. A person retains responsibility, but AI reduces the production effort. Typical examples include preparing a draft, summarising evidence, comparing documents, conducting first-pass research or formatting information. For many SMEs, this is where the most credible near-term value sits.

Reserve. The task remains deliberately human. This may be because it depends on negotiation, commercial judgement, trust, tacit organisational knowledge or a consequence the business is unwilling to delegate. Record the reason. An undocumented boundary will be challenged later as the technology improves or cost pressure rises.

Retire. The task should stop. It may duplicate a control, serve an audience that no longer uses it, or exist only because an old process once required it.

Retire deserves its own decision, not a footnote beneath automation. Stopping low-value work releases capacity immediately, carries no licence fee and prevents the business from teaching a new system to reproduce yesterday's waste.

The four choices also prevent a common distortion: describing assisted work as automated work. If a person must read, correct and accept every output, the task has not disappeared. Its production method has changed.

Move 3: Build a coherent job from what remains

Sorting the tasks is analysis. Recomposition is design.

Do not ask only what remains in the existing role. Look across adjacent roles and ask what combination of work now makes organisational and human sense.

The answer may be a more focused operational job. It may be a broader role with genuine ownership of an outcome. Administrative capacity released across several jobs may create enough room for a specialist capability the business previously could not afford, such as commercial analysis, process improvement or customer insight.

Coherence matters because a collection of leftovers is not a viable job. A well-designed role should meet four tests:

1. It owns an outcome the business values.
2. Its tasks require a reasonably related set of skills.
3. Its decision rights match its accountability.
4. A capable candidate could understand why the role is worth doing.

Try writing a two-sentence recruitment advert. If the role sounds like three unrelated part-time jobs pushed together, return to the design.

Then examine grade and reporting line. When routine production falls away, the remaining work may require more judgement, not less. Assuming every AI-enabled role should be cheaper can leave the business with fewer tasks but a more demanding job priced at the wrong level.

Move 4: Give every released hour a destination

Capacity only becomes value when the business decides what it will replace or enable.

There are five legitimate destinations:

1. More volume: The same team serves more customers, processes more cases or supports growth without equivalent hiring.
2. Higher quality: Time is reinvested in work where better analysis, service or control has commercial value.
3. New work: Capacity funds activity that has been consistently deferred, such as account development, process improvement or management information.
4. Lower cost: The business uses attrition, vacancy control or a formal restructuring process to reduce the cost base.
5. Recovery: The team receives some capacity back because workloads have been unsustainable and retention or resilience is at risk.

None is automatically superior. The correct destination depends on strategy, service levels, growth plans and workforce conditions.

The error is leaving the choice implicit. "Improve efficiency by 15%" is not an operating decision. State where the hours will go, who owns the result, which measure will prove it and when the leadership team will review it.

If the intended outcome is higher volume, track output without extra headcount. If it is quality, define the defect, renewal or customer measure expected to improve. If it is recovery, measure whether overtime, absence or unwanted turnover changes.

Move 5: Replace the old working agreement

The redesigned role needs a one-page role contract. This is an operating document, not a substitute for an employment contract.

It should state:

• the outcomes the role owns;
• decisions the role can make independently;
• decisions requiring another check;
• escalation triggers;
• approved and prohibited AI uses;
• evidence that must be retained;
• the measures used to judge performance; and
• the route into the next level of responsibility.

This gives the jobholder and manager a shared basis for working in the new model. It also prevents people becoming personally liable, in practice, for machine-assisted decisions they have neither the time nor information to review properly.

The development path is particularly important. Junior employees traditionally build judgement by completing lower-risk versions of the work, observing feedback and learning where standard rules fail. When AI takes most of that practice work, the route to expertise weakens.

Leaders should decide what replaces it. Options include supervised exception handling, structured case reviews, rotations, simulations, paired decision-making and deliberate exposure to customer or operational context. Otherwise, the business may enjoy a short-term productivity gain and discover later that it has no credible internal successor for its senior roles.

Four failure patterns to prevent

Role redesign can produce an attractive organisation chart and a poor working reality. Four patterns deserve an explicit check.

The hollowed role

The system takes the routine and satisfying work. The person receives only disputes, unusual cases and failures.

This concentrates cognitive and emotional load while removing the sense of progress that came from completing a wider range of work. The role may appear more senior because every case is difficult, but it can quickly become exhausting and unattractive.

Keep enough ownership, variety and visible achievement in the human role. If the role consists entirely of exceptions, redesign the exception process and reconsider whether the workload can support a sustainable job.

The checker's trap

The person is described as the accountable reviewer, but the volume makes meaningful review impossible.

If forty outputs arrive each day and proper review takes fifteen minutes, the role requires ten hours before any other work begins. Calling each item "human checked" does not create control. It creates approval theatre.

Choose deliberately between full review and sampling. If full review is required, resource it honestly. If sampling is sufficient, define selection rules, error tolerances, escalation thresholds and what happens when performance deteriorates.

The stacked role

Prompt maintenance, output checking, exception handling, user support and tool administration are added to the existing job, while nothing is removed.

This often affects the most enthusiastic early adopter. Their competence attracts more requests until they become an unofficial support function alongside their substantive role.

Count the new work. Assign ownership. Remove or transfer something of comparable weight.

The patchwork role

Fragments from different teams are combined because each is too small to remain where it was.

The resulting role lacks a clear outcome, spans unrelated skills and depends on a particular individual's tolerance for ambiguity. It becomes difficult to recruit, manage or benchmark. Its fragility is exposed when the person leaves.

The test is simple: could the role be explained without listing its tasks? If it has no coherent purpose, it needs another design pass.

A twelve-question role diagnostic

Score each question from 1 to 5. Any score of 1 or 2 requires an owner, an action and a date.

1. Do we understand the work as it is actually performed?
2. Has every material task been classified as Automate, Assist, Reserve or Retire?
3. Is the reason for every reserved task recorded?
4. Have we stopped any work, rather than simply changing how it is produced?
5. Does the redesigned role have a clear purpose and a worthwhile outcome?
6. Are decision rights consistent with accountability?
7. Can the person genuinely complete the review workload assigned to them?
8. Does released capacity have a named destination, owner and review date?
9. Has work been removed from the role as well as added?
10. Do objectives and measures reflect the redesigned job?
11. Would a recruitment brief written today describe the real role accurately?
12. Is there still a credible path by which someone can develop into the senior version of the role?

Do not average away disagreement. If the manager scores role clarity at 5 and the jobholder scores it at 2, that gap is more useful than the average of 3.5.

A practical 60-day sequence

Days 1 to 10: Select and observe

Choose no more than three roles. Prioritise jobs where AI use is already material, a vacancy is open, someone is leaving, or the function is under pressure to release capacity.

Build the task ledger with the people performing the work. Use current cases and artefacts, not only interviews. Look at the reports, approvals, prompts, spreadsheets, system queues and private workarounds that reveal how the job operates.

Days 11 to 25: Classify and challenge

Sort every task into the four categories. Include the line manager, jobholder and one informed person from outside the function. An external perspective helps challenge the belief that every familiar task requires specialist judgement.

Report retired work separately. It is often the fastest capacity release and the clearest test of whether leaders are willing to remove work, not just buy tools.

Days 26 to 40: Recompose and commit

Design the future role or roles. Test each for purpose, skill coherence, decision rights, workload and development path. Run the four failure-pattern checks.

Choose the destination for released hours and place the commitment beside the original AI business case. Draft the one-page role contract and update any open recruitment brief.

Days 41 to 60: Implement and learn

Put one redesigned role into operation properly. Brief the jobholder and manager, change objectives and measures, establish exception and review routines, and book a 90-day review.

Retain the original task ledger. It provides the baseline for testing whether work genuinely moved, whether new tasks accumulated and whether the intended value appeared.

Material changes to duties, reporting lines, location, hours or grading may create consultation and contractual obligations in the UK. Involve HR or an employment solicitor before communicating changes that affect substantive terms. Allington Advisors is not a law firm, and this article is not legal advice.

Be equally clear about workforce intent. If lower cost is the chosen destination, say so and follow a proper process. Describing a redundancy exercise as role innovation will damage trust and produce unreliable information about how work is really done.

Measure work redesign, not tool activity

Licence activation and prompt counts show that people opened the software. They do not show that roles improved.

Use measures that reveal whether the operating design changed:

• tasks retired;
• roles with a current task ledger;
• released hours assigned to a named destination;
• capacity outcomes confirmed at the review date;
• review volume per person;
• exception and override rates;
• time from vacancy approval to a rewritten role specification;
• regretted turnover in redesigned roles; and
• internal promotion into roles whose traditional feeder tasks have changed.

Pair quantity with consequence. A high number of retired tasks means little if they consumed almost no time. A strong adoption rate can conceal a rising correction workload. A shorter process may be a poor result if customer outcomes weaken.

The most useful measure is the one linked to the chosen capacity destination. If the business said the team would absorb 20% more volume, that is the promise to test.

Five questions for the leadership team

1. Which job has changed most because of AI while its formal design has changed least?
2. Which activities have we stopped, rather than automated or accelerated?
3. Where did the capacity from our last AI investment go?
4. Who is now expected to approve more machine output than they can genuinely inspect?
5. How will our strongest junior employees acquire judgement if routine learning work disappears?

If the team cannot answer the third question, the next AI investment case is missing an operating decision.

Use the moment before the role hardens again

There are three particularly useful moments to redesign a role: when a new tool is introduced, when a vacancy appears and when the organisation is being restructured.

Each creates permission to question what the job is for. Once a person, manager and team adapt around the new tool, informal routines settle and the opportunity becomes harder to recover.

AI will continue to change task economics whether job design keeps pace or not. The businesses that capture the benefit will not simply deploy more capability. They will decide which work should stop, which work should move, what people should own and where the released capacity should create value.

Allington Advisors helps leadership teams redesign roles around the work the business now needs, make explicit choices about released capacity and ensure that future appointments reflect the job ahead rather than the job inherited. A focused Role Design Review can turn two or three materially changed roles into clear role contracts, credible capacity decisions and better hiring briefs.