An Automation Forecast Is Not a Redundancy Notice
A forecast can warn you where to inspect your work without telling you that you will be fired. Run the claim card before letting a percentage choose your next move.
Reading-forecasts supplies a ten-box claim card; occupational-exposure supplies the task test that keeps an exposure percentage from becoming a personal job-loss prediction.
A clean card cannot make a role with a real closure date safe, and income protection cannot wait for a better forecast.
The headline is open beside the laptop where you are refreshing job listings.
47% of jobs are at risk.
Maybe it arrived in an employer email. Maybe it was on a consultant slide in a meeting where nobody said the word layoff. Maybe a relative forwarded it with a sentence that needs an answer you do not have.
Do not answer the headline yet.
Copy it before it chooses your next move.
The number may describe a real danger.
It may also describe a different country, a different unit of work, a technical capability rather than an employer decision, or a scenario assembled to show assumptions.
Those are not minor qualifications.
They are the claim.
What the Headline Dropped
That 47% came from a study of occupational categories, weighted by U.S. employment, under a stated cutoff, over an unspecified span of years.
Every one of those conditions was removed on the way to the slide.
The headline does not count redundancies.
It does not report an employer adoption date.
It does not say that every task performed by every person in a high-scoring occupation is equally exposed.
It says a classification crosses a threshold in a model.
That is a large finding.
Keep it large.
It is not the sentence the headline makes it sound like.
The percentage was priced to travel farther than its method.
That tells you more about the market for alarming certainty than about the value of your particular work.
The Claim Card
Take out a sheet of paper.
Draw ten boxes.
Write the automation claim exactly as you received it at the top.
Do not improve its wording.
Do not translate “at risk” into “will be fired.”
Fill the boxes.
Producer. Name the institution or author who made the estimate, not the person who reposted it.
Date. Write the publication date. A claim about then-demonstrated technology has a date built into it.
Technology. Write what system or capability is actually being assessed: computerisation, a particular automation technology, AI, or a bundle of tools.
Unit. Write what receives the score: occupations, workers, activities, jobs, hours, wages, or another object.
Population. Write the numerator and denominator. “Employment” is not enough. Ask whether it means total employment, a sample of workers, paid activity time, or something else.
Geography. Write the country, countries, region, or global base.
Horizon. Write the target date or interval. A 10–20-year susceptibility estimate is not a payroll result.
Threshold. Record the category rule. A 0.70 cutoff does not decorate the result. It creates the high-risk class.
Method. Write occupation classification, worker-task survey, activity-time analysis, scenario model, or a named method from the source.
Measure. Circle one: exposure, adoption, or employment. If the source measures exposure, leave employment uncircled.
Score one point for every blank.
0–2 blanks: usable context. Keep the claim, with its boundaries, as a reason to inspect your work.
3–5 blanks: warning. The statistic may be useful, but it cannot carry the decision someone is trying to make with it.
6–10 blanks: panic material. Treat it as an object designed to travel, not as a decision input.
Nothing about this card has been validated.
It is a way of stopping an unknown percentage from occupying the space where facts should be.
The Unit
Do not let the word jobs do all the work.
McKinsey assessed more than 2,000 activities in over 800 occupations.
It found 45% of activities people were paid to perform technically automatable using then-demonstrated technologies.
That is not a 45% unemployment forecast.
Its unit is activity time.
The same research found fewer than 5% of occupations were candidates for full automation.
Your job can contain routine data collection, processing, and structured physical work, and still keep exceptions, customer contact, accountability, and judgment.
Partial potential is not a small version of full replacement.
It is a different operational question.
Use the number to list activities.
Then list what surrounds them: inputs, handoffs, controls, error cost, rework, escalation, and the human task left after deployment.
A system that removes 10 hours of copying can produce 10 hours of checking.
It can shift work to customers.
It can shorten a queue.
It can allow more volume.
It can support fewer hires.
It can remove jobs.
An exposure score does not decide which one happened.
Make your employer state it.
Record baseline hours per week, throughput, defects, rework, customer escalation, staffing, and the human work left after deployment.
Those are adoption and employment evidence.
The occupation percentage is not.
Do not move a score across borders by matching a title either.
O*NET supplies more than 900 U.S. occupation profiles and connects task statements to a common taxonomy.
That makes it useful research infrastructure.
It does not make a U.S. task score automatically true in another labour market.
Demand a documented occupational crosswalk and an argument that the work is organized comparably.
A title match is not evidence of comparable tasks.
The Threshold
Find the threshold before you repeat the headline.
The OECD’s worker-task method places roughly 14% of workers at high risk and another 32% in a band for significant task change.
High risk means a probability above 70%.
Significant change runs from 50% to 70%.
Those two bands distinguish likely whole-job automation in the method from major task redesign.
The distinction prevents a common fraud of language.
Significant change becomes gone.
It is not gone.
The category is changed work.
Do not average 47%, 14%, and 32% into a sensible-looking number.
They are not rival counts of the same thing.
One assigns a common occupational classification. The other allows people inside the same occupation code to report different task profiles.
Different population. Different unit. Different threshold treatment.
The words at risk are doing mathematical work.
Make them show it.
The Missing Middle
Technical capability is not deployment.
Integration cost matters.
Reliability matters.
Workflow data, liability, labour supply, competition, regulation, and customer acceptance matter.
Name them, because they are where a capability headline becomes an employer plan, or fails to become one.
Ask for the affected task group.
Ask for the vendor claim.
Ask for the planned adoption date, validation steps, monitoring capability, expected staffing effect, training pay, redeployment options, and decision rights.
If there is a consultation period, find the trigger.
A 90-day consultation only changes anything if it begins before procurement and requires an impact assessment with a named worker group and a baseline.
Do not let “AI will transform our jobs” end the conversation.
Make it name the process, system, date, evidence, and staffing claim.
Exposure tells you where to inspect.
Adoption tells you whether a system is entering the work.
Employment evidence tells you what happened to hours, jobs, pay, hiring, or transitions after that.
Keep the three columns separate.
Useful and Wrong
Do not demand prophecy from a projection.
BLS evaluated its 2014–24 occupational projections and found 86% directional accuracy for occupational-group growth or decline.
In the same evaluation, projected average growth was 6.5% and actual growth was 12.9%.
Keep both results.
The projection identified direction often enough to structure inquiry.
It missed the average rate by enough to make a personal plan built on the point estimate irresponsible.
Ask whether the category itself stayed stable long enough to test.
BLS notes that SOC codes changed in 2010 and 2018, and evaluation needs documented mappings when labels split, merge, or move.
An apparently precise occupation forecast can fail because the target changed, not only because the estimate missed.
For industry employment, BLS projected 149.1 million nonagricultural wage-and-salary jobs for 2024 against 158.6 million actual: an error of 6.0%.
A naive trend model was off by 13.0%.
The point is not that official forecasts are safe.
The point is that a serious forecast can be imperfect and still improve on an empty trend line.
Look for the evaluation, the comparison model, the code mapping, the data year, and the original wording.
A forecast is worth keeping and worth re-reading later.
It is not worth believing.
Keep the publication date.
Keep later revisions.
Keep the observable signals that would update the claim.
A six-month review shows whether the adoption pathway exists before the distant horizon arrives.
The Closure Date
A clean card will not make a threatened role safe.
If your employer has named a closure date, cut hours, ended a contract, announced a procurement, or removed a function, that is not merely exposure.
It is an operating condition.
Deal with the income first and finish the appraisal afterwards.
Use the claim card to stop a general percentage from impersonating your redundancy notice.
Do not use it to pretend a real notice is only a model.
If you have a closure date, you cannot wait for a better research design before you make a call, pursue bridge work, test a transition, or cut an expense.
But do not let urgency make the headline true.
Retraining is not a universal exit.
OECD Skills Outlook finds that, in a scenario allowing up to six months of training, about half of high-risk occupations had no acceptable move to a lower-risk occupation.
That is the part the headline and the course advertisement have in common.
Neither one names where you would go.
Do not buy a course because a forecast frightened you.
Study destination demand, entry requirements, pay, local employers, training cost, and the work you would actually perform.
Then return to the card.
Minimum Viable Action
Before you sleep tonight, find the one automation statistic that has been frightening you.
Make the ten boxes.
Fill every box you can.
Count the blanks.
Write one sentence underneath: This claim measures ____; it does not yet measure ____.
Do not leave the second blank empty.
What you will be able to say afterwards
- Exposure, adoption, and employment loss are separate measures that require separate evidence.
- A forecast claim can be tested by recording its producer, date, technology, unit, population, geography, horizon, threshold, method, and measure type.
- A high-risk threshold and the numerator and denominator change what an automation percentage means.
- An occupation score does not describe identical task exposure for every worker with that title.
- A scenario midpoint is a model assumption rather than a personal probability.
- A task score is a prompt for a worksite process inventory, not a staffing instruction.
Questions this raises
Does a headline saying that 47% of jobs are at risk mean that 47% of workers will be fired?
No. Frey and Osborne classified 702 detailed U.S. occupations, used a 0.70 high-risk cutoff, and reported that about 47% of total U.S. employment fell in that modelled high-risk category. The result describes potential computerisation, not observed layoffs, an adoption date, or a prediction that every worker in the category loses employment.
Why does one automation study say 47% and another say 14%?
The figures use different units and methods. Frey and Osborne apply an occupation-level classification to U.S. employment. OECD worker-task reporting finds roughly 14% at high risk and another 32% facing significant task change across participating countries. Worker-task methods allow people with the same occupation code to have different task profiles, so they do not assign one common risk to every worker in a title.
What does an automation exposure score actually measure?
An exposure score measures predicted technological contact or technical potential under a stated method. It can identify task bundles worth inspecting. It does not by itself measure whether an employer adopts a system, whether work is reassigned or expanded, whether hours fall, or whether employment is lost. Those outcomes require separate evidence about deployment and the worksite.
What belongs on a claim card for an automation statistic?
A claim card records the producer, publication date, technology, unit, population, geography, horizon, threshold, method, and whether the claim measures exposure, adoption, or employment. It also needs a readable numerator and denominator. Those fields show whether a percentage counts workers, occupations, activities, work hours, wages, or some other object before it is used in a decision.
What does a 0.70 automation-risk threshold mean?
A 0.70 threshold is a study rule for putting a modelled score into a high-risk category. Frey and Osborne used that cutoff in their occupation classification, and later OECD reporting defines high risk as a probability above 70%. The threshold does not establish that an individual worker is 70% likely to be fired; it defines a category within a particular method.
Can a job contain automatable work without being fully automatable?
Yes. McKinsey's activity analysis finds that 45% of paid activities could be technically automated using then-demonstrated technologies, while fewer than 5% of occupations are candidates for full automation. The difference arises because occupations are bundles of activities. A system can change copying, data processing, or drafting while leaving exceptions, oversight, customer contact, and other work in place.
Is the middle number in an automation scenario a probability?
Not necessarily. McKinsey's 2017 scenarios place potential automation by 2030 between almost zero and 30% of global hours, with a 15% midpoint. That midpoint is a central modelled case, not a 50% probability statement, unless the report supplies a probability model. Scenario ranges expose assumptions about adoption paths rather than sampling uncertainty.
Why does a capability forecast not tell anyone when jobs will disappear?
Technical capability and adoption are different events. McKinsey places realization of half of then-current work activities around 2055, with a range 20 years earlier or later. Integration cost, reliability, workflow data, liability, labour supply, competition, regulation, and customer acceptance affect timing. A technical-potential score therefore identifies possible task redesign more reliably than workforce timing.
What should an employer measure before claiming that automation saves jobs or removes them?
An employer needs a worksite process inventory: activities, handoffs, exceptions, controls, customer contact, baseline hours, throughput, defects, rework, escalations, staffing, and the human task after deployment. A reduction of 10 copying hours can become checking work, customer work, capacity, shorter queues, fewer hires, or job cuts. An occupational exposure score cannot choose among those outcomes.
Why can two estimates use the same high-risk occupations and still produce different shares?
Within-occupation assumptions change the result. OECD Skills Outlook reports lower-bound estimates of 4% to 10% when only a proportion of workers in high-risk occupations are treated as high risk, and upper-bound estimates of 19% to 48% when all workers in those occupations are treated that way. The dramatic endpoint is an assumption-driven scenario result, not a count of actual automation.
What can be learned from a forecast that later misses its target?
A forecast can still show whether its method adds information and which assumptions failed. BLS's 2014-24 occupational evaluation identifies growth or decline correctly for 86% of occupational groups, while projecting 6.5% average growth against 12.9% observed. The score and the miss need to remain together; neither proves that forecasting is perfect nor that structured projections have no use.
Should a worker delay income protection until an automation forecast is clearer?
No. A forecast does not settle an individual employment outcome, but a real closure date, lost hours, or a stated employer plan is operational evidence. Income protection, job search, and transition decisions use the observed condition. The claim card is for preventing a general percentage from impersonating a personal notice; it is not a reason to ignore a documented employment change.