eduKateSG · Career & Adulthood · Job Redesign and AI · Hidden Rework · 2 October 2026
Generative AI can make a task look faster at the moment output appears. That is not the same as making the workflow faster. A draft arrives sooner, but someone may still need to verify sources, correct logic, reconcile numbers, reformat the output, handle exceptions and repair mistakes that reach the next team. If those costs are invisible, productivity claims become stories about the easiest part of the task.
This guide develops SG-AUD-20260929-09 beneath Job Redesign and AI. It does not create a new AI-work root. Its narrower job is to compare faster output with hidden rework and retained human judgment, using task boundaries, verification effort and downstream corrections rather than invented productivity benchmarks.
The empirical literature already warns against one universal productivity story. In customer support, Brynjolfsson, Li and Raymond found substantial average productivity gains from a generative-AI assistant, with larger gains for less-experienced workers; the work was later published in the Quarterly Journal of Economics. In a different setting, METR’s early-2025 randomized study of experienced open-source developers found that allowing AI tools increased completion time by 19% in that sample even though developers believed the tools made them faster. METR’s February 2026 update says newer tools may be producing more speedup, but the follow-up study had serious selection effects and cannot provide a reliable current estimate. These results are not contradictory once task, worker and measurement are kept visible.
Another large field experiment across thousands of knowledge workers found that access to generative AI reduced time spent on email and changed some individual work patterns, while many coordination patterns changed less. The lesson is that AI often transforms pieces of work before it transforms the whole system. Job redesign therefore needs workflow measurement, not only tool adoption.
Output speed is not workflow speed
Output speed is not workflow speed is mainly about measuring the whole task rather than the generation moment. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is an AI drafting a report in two minutes while source checking and revision take another hour. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is calling the task 30 times faster because first-draft time collapsed. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that total cycle time includes verification, correction and handoff. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Visible automation can create invisible review work
Visible automation can create invisible review work is mainly about counting labour that appears after AI output. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is staff checking citations, calculations and policy references generated automatically. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is measuring only keystrokes saved. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that review effort is logged as part of the redesigned task. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Rework can move downstream
Rework can move downstream is mainly about seeing errors that survive early review and surface later. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is a generated customer reply creating a follow-up complaint that another team must resolve. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is declaring success because the original agent closed the ticket quickly. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that downstream corrections are included. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Human judgment should be attached to a decision point
Human judgment should be attached to a decision point is mainly about specifying exactly what the person must decide. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is a compliance officer approving whether AI-generated text matches current regulation. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is adding a generic human-in-the-loop label. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that the human has criteria, authority and time to intervene. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Verification should match the failure mode
Verification should match the failure mode is mainly about checking the properties that actually matter. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is source verification for research, tests for code, reconciliation for financial calculations. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is using one universal review checklist. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that controls are tailored to the task. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
High-confidence language can hide uncertainty
High-confidence language can hide uncertainty is mainly about treating fluency as presentation rather than evidence. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is an AI producing a polished explanation with an unsupported factual claim. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is reducing review because the prose sounds professional. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that evidence quality is checked independently of style. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Task decomposition reveals where AI helps
Task decomposition reveals where AI helps is mainly about separating research, drafting, calculation, checking and decision. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is using AI for a first structure while a human owns evidence and final recommendation. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is evaluating the whole job as either automated or not. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that task-level effects are visible. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Different workers can experience different effects
Different workers can experience different effects is mainly about recognising heterogeneous productivity. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is novices benefiting from suggestions while experts spend more time correcting mismatched advice. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is reporting one average as universal. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that worker experience and task familiarity are part of the interpretation. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Different tasks can produce opposite results
Different tasks can produce opposite results is mainly about preserving task context. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is customer support showing gains while one experienced-developer RCT found slowdown. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is generalising one study to all knowledge work. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that evidence remains task-specific. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Perceived speedup can differ from measured speedup
Perceived speedup can differ from measured speedup is mainly about checking subjective impressions against outcomes. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is developers believing AI saved time in a setting where measured task time increased. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is using self-report alone. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that perception and observed performance are compared. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Current AI effects change over time
Current AI effects change over time is mainly about treating model and tool capability as dated. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is METR noting in 2026 that late-2025 tools may differ materially from early-2025 tools but their new estimate is selection-biased. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is freezing one benchmark into an evergreen productivity claim. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that article separates mechanism from dated empirical results. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Coordination tasks may change more slowly than individual tasks
Coordination tasks may change more slowly than individual tasks is mainly about distinguishing work that one person can alter from work requiring team redesign. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is AI reducing email work while meetings remain largely unchanged in a field experiment. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is assuming individual tool access automatically transforms team process. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that coordination dependencies are mapped. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Local speed can create global congestion
Local speed can create global congestion is mainly about watching bottlenecks move. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is AI generating more sales proposals than reviewers can approve. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is celebrating higher draft volume despite a growing queue. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that system throughput is measured at the real constraint. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Quality thresholds can raise or lower rework
Quality thresholds can raise or lower rework is mainly about setting explicit standards. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is requiring source-backed claims before a report is accepted. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is letting every reviewer use a different intuition. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that acceptance criteria are shared. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Automation bias is a workflow risk
Automation bias is a workflow risk is mainly about protecting against over-reliance on plausible output. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is reviewers checking high-impact fields rather than assuming AI is usually right. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is rubber-stamping because errors are rare. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that verification remains active. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Excessive verification can erase the benefit
Excessive verification can erase the benefit is mainly about avoiding controls that cost more than the task warrants. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is sampling low-risk routine outputs while fully checking high-risk cases. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is requiring exhaustive review for trivial drafts. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that review effort scales with consequence. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Risk tiering changes the review design
Risk tiering changes the review design is mainly about matching human oversight to stakes. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is full review for legal advice and lighter review for internal brainstorming. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is using identical controls everywhere. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that risk level determines validation depth. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Reference tasks need source provenance
Reference tasks need source provenance is mainly about requiring links or citations that can be checked. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is AI summarising a policy with the current official source attached. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is accepting unattributed summaries. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that reviewers can trace consequential claims. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Calculation tasks need independent computation
Calculation tasks need independent computation is mainly about using deterministic tools where possible. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is recomputing a financial total with a spreadsheet or calculator. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is asking the language model to verify itself. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that verification uses an independent mechanism. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Code tasks need tests and repository context
Code tasks need tests and repository context is mainly about checking execution rather than prose plausibility. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is running tests, linting and code review after AI generation. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is accepting code because it looks idiomatic. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that software behavior is verified. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Customer-facing work needs outcome checks
Customer-facing work needs outcome checks is mainly about looking beyond handle time. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is tracking resolution quality, escalation and customer sentiment. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is optimising only speed. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that service measures remain balanced. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Knowledge work needs claim-level verification
Knowledge work needs claim-level verification is mainly about checking evidence behind conclusions. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is reviewing assumptions and sources in an AI-assisted memo. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is editing only grammar. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that reasoning and evidence are inspected. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
AI can shift work from creation to evaluation
AI can shift work from creation to evaluation is mainly about recognising a new skill mix. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is workers spending less time drafting and more time choosing among alternatives. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is assuming saved drafting time becomes idle capacity. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that evaluation becomes a core task. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Evaluation itself is skilled work
Evaluation itself is skilled work is mainly about training people to judge AI output. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is a junior analyst learning to test assumptions and evidence. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is assuming anyone can review expert-looking output. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that review capability is developed. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Junior learning can be weakened by bypassed practice
Junior learning can be weakened by bypassed practice is mainly about preserving developmental tasks. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is letting juniors attempt a first analysis before seeing AI output. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is using AI to eliminate every beginner task. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that the organisation retains routes for novices to build judgment. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Senior experts can be overloaded by review
Senior experts can be overloaded by review is mainly about avoiding concentration of all checking at the top. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is designing automated tests and distributed review rules. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is sending every AI draft to a small group of experts. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that expert attention is reserved for genuinely hard cases. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Exception queues are part of the redesign
Exception queues are part of the redesign is mainly about making escalation visible. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is routing low-confidence or high-impact cases to specialists. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is forcing the AI path to handle everything. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that exceptions have an owner and service level. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Rollback must remain possible
Rollback must remain possible is mainly about maintaining an alternative process when automation fails. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is reverting to the prior workflow during a tool outage or quality incident. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is making the AI system the only path immediately. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that continuity is preserved. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Audit trails make rework learnable
Audit trails make rework learnable is mainly about recording why output was changed. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is capturing common correction categories. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is fixing errors silently so the system learns nothing. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that rework data informs redesign. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Prompt changes are process changes
Prompt changes are process changes is mainly about controlling versioning when instructions affect outputs. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is testing a revised prompt on a sample before rollout. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is editing shared instructions casually. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that workflow changes are traceable. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Model changes are dependency changes
Model changes are dependency changes is mainly about rechecking performance when the underlying model changes. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is rerunning critical evaluations after a model upgrade. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is assuming newer always means better for the task. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that release changes trigger review. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Human approval should be meaningful
Human approval should be meaningful is mainly about preventing ceremonial sign-off. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is requiring the approver to see evidence and unresolved uncertainty. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is asking for a click without time to inspect. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that approval has substance. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Metrics should include correction cost
Metrics should include correction cost is mainly about measuring time spent repairing output. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is tracking draft time plus review, rework and exception handling. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is publishing generation-time savings only. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that net productivity is estimated at workflow level. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Metrics should include missed-error cost
Metrics should include missed-error cost is mainly about recognising that undetected errors can dominate. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is measuring downstream incident or customer impact where appropriate. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is treating zero visible rework as proof of quality. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that silent failures are considered. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
The best AI task may be the one with cheap verification
The best AI task may be the one with cheap verification is mainly about selecting opportunities where checking is easy relative to generation. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is using AI to draft code with strong tests or summarise structured data with known totals. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is automating tasks whose correctness is difficult to inspect. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that verifiability is part of use-case selection. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
Human judgment should get stronger, not merely remain
Human judgment should get stronger, not merely remain is mainly about using AI to expose alternatives while preserving final reasoning. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is a manager comparing several scenarios and documenting why one was chosen. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is letting AI supply the conclusion and human merely endorse it. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that the human can independently explain the decision. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
The final test is independent recovery
The final test is independent recovery is mainly about checking whether the team can still operate when AI is unavailable or wrong. The redesign team should draw the workflow from input to final accepted outcome. Mark where AI generates, where a person verifies, where information is handed off, where exceptions go and what happens if an error escapes. This makes hidden work visible before benefits are claimed.
A practical example is staff completing a critical process manually during a controlled exercise. The example matters because AI effects are conditional. A model can be excellent at producing plausible first drafts and still create expensive review in a high-stakes setting. It can be highly valuable in another task where errors are cheap to detect. The same model can therefore improve one workflow and harm another.
The common failure is allowing skill atrophy until fallback is impossible. This usually measures the visible acceleration and ignores the work displaced to reviewers, customers, downstream teams or future incidents. AI adoption then appears successful locally while total organisational effort rises. Good job redesign follows the work after the output leaves the prompt box.
The acceptance test is that the organisation retains recoverability. Measure net cycle time, quality, correction effort, downstream rework and exception load at the appropriate risk level. If human judgment remains necessary, specify what it is for. The goal is not to keep a human present ceremonially; it is to preserve accountable decisions and build a workflow that can recover when the tool fails.
A hidden-rework audit
- Define the final accepted outcome, not merely the AI-generated artefact.
- Measure baseline cycle time and quality before changing the workflow.
- Record generation time, review time, correction time and downstream rework separately.
- Classify common failure modes and which ones are cheap or expensive to verify.
- Define when human approval is mandatory and what the reviewer must inspect.
- Track exception volume and whether experts become a new bottleneck.
- Re-evaluate after model, prompt or policy changes.
- Keep rollback and independent-recovery capability for critical work.
Three evidence cases that should not be collapsed
Customer support
The customer-support study shows that AI assistance can raise measured productivity in a real operational setting, particularly for less-experienced workers. That is evidence for that workflow and tool configuration. It does not prove the same effect in software engineering, law, teaching or executive decision-making.
Experienced open-source software development
METR’s early-2025 randomized study found slowdown for experienced developers working in mature repositories they knew well. The authors explicitly caution against generalising this result to most software work. Their 2026 update also warns that newer-tool estimates are compromised by selection effects. The useful lesson is methodological: perceived speed, benchmark performance and real-task productivity can diverge.
Cross-industry knowledge work
The Microsoft/NBER field experiment found that AI access changed some individually controlled activities, such as email work, more readily than coordination-heavy activities. That distinction matters because organisations are networks of dependencies. Saving time in one person’s draft does not automatically redesign meetings, approvals or handoffs.
Frequently asked questions
Does AI increase productivity at work?
Sometimes, and sometimes not. Effects vary by task, worker, tool, workflow and measurement. Current evidence should be read at that level rather than as a universal percentage.
What is hidden rework?
Work created after the AI output appears: verification, correction, clarification, exception handling, downstream repair or incident response that is easy to omit from productivity measurements.
Why can people feel faster even when measured time does not improve?
AI can reduce aversive or effortful parts of a task, make progress feel faster, or shift work into review. Subjective speed and end-to-end completion time are different measures.
Should every AI output be fully reviewed?
No. Review depth should match risk, verifiability and consequence. Low-risk drafts may need lighter checks; high-stakes decisions need stronger controls.
What tasks are good candidates for AI?
Tasks with clear inputs, valuable generation assistance and relatively cheap verification are often easier to redesign safely than tasks where correctness is ambiguous and errors are hard to detect.
Can automated verification replace humans?
It can reduce review burden in domains with strong machine-checkable properties, but accountability and exception handling may still require human judgment depending on the task.
How should AI productivity be measured?
Measure the final workflow: cycle time, accepted quality, correction and review effort, exceptions, downstream rework and relevant outcomes—not just first-output speed.
What happens when models improve?
Re-evaluate. A workflow that was inefficient with one generation of models may become useful later, and a model change can also introduce new failure modes.
Could AI reduce learning for junior staff?
Yes, if it removes the tasks through which novices used to build judgment. Job redesign should preserve developmental routes rather than only optimise immediate output.
What is retained human judgment?
The decisions that remain accountable to a person: framing the problem, selecting evidence, interpreting exceptions, weighing trade-offs, approving high-stakes actions and explaining the final decision.
The quiet conclusion: speed belongs to the whole workflow
AI changes the economics of producing an answer. It does not automatically change the economics of trusting that answer. Verification, context, accountability and downstream integration still have costs, and those costs vary dramatically by task.
The mature question is therefore not whether AI is fast. It is whether the redesigned system reaches an accepted outcome faster, better and with risks we understand. When measurement includes hidden rework and human judgment remains real, organisations can distinguish genuine capability from impressive output velocity.