Your Exposure Score Fails at the Worksite
An occupational score becomes useful only after it meets the actual task, error cost, and deployment evidence at your workplace. Run the two-stage exposure audit.
Occupational-exposure supplies the published classification; automation-exposure supplies the task and deployment checks that test whether it survives contact with one worksite.
A careful audit cannot stop management from using a crude occupational score, and workers often cannot obtain the adoption evidence needed to settle an unresolved result.
The website has put a red score beside the name of your occupation.
Your title is correct.
The difficult part of your day is nowhere in the description.
Maybe the code is wrong. Maybe the code is right and the work is wrong for the code. Maybe the score describes a U.S. occupational category while your job is organized somewhere else, with different customers, controls, software, and responsibility.
Do not argue with the colour.
Audit what it is attached to.
An occupational score is a shortcut.
It saves a researcher from observing every workplace in a country.
That is useful for a national comparison.
It fails when somebody uses it as a verdict on the work in front of you.
The Wrong Desk
Start with the famous number, because it is probably what the website is borrowing.
Frey and Osborne classified 702 detailed U.S. occupations.
At a 0.70 high-risk cutoff, their employment-weighted result placed 47% of total U.S. employment in the modelled high-risk category.
Keep every part of that sentence.
It is U.S. employment.
It is an occupational classification.
It is a category created by a cutoff.
It is potential computerisation over an unspecified number of years, not a count of jobs actually automated.
The study models a broad occupation system because broad systems are what employment data contains.
Your employer does not employ an occupation system.
It employs a process, with people doing different versions of the job inside it.
That difference produces the other number people like to quote.
Arntz, Gregory, and Zierahn used worker-task information and estimated 9% of jobs automatable on average across 21 OECD countries.
That is not the correction to apply to a 47% headline.
It is a different measurement.
The occupation method gives every worker attached to a category the category’s score.
The worker-task method allows two people with the same code to have different work.
One is not reassuring. One is not alarming.
They tell you where the measurement changes.
The unit decides what the percentage can say.
The Code Behind the Colour
Find the occupation code behind the red score.
Do not accept a title alone.
O*NET holds 923 data-level occupations inside 1,016 titles.
A workforce file may use another national classification.
A crosswalk can merge several local roles into one category and mechanically flatten the difference that matters.
Write down the code.
Write down the code system and version.
Write down the study year and the country it covers.
Then run stage one.
Nothing here is a scoring system.
It is a record of what has to be true before a published number earns any influence over your decisions.
Record six things.
The technology. Is the study about computerisation, currently demonstrated technologies, a named automation system, or something else?
The task source. Does it use an occupational database, a worker-task survey, activity time, or a modelled capability profile?
The threshold. Where does high risk begin? Does the source also name a middle band as significant task change?
The code. Which code and classification put your title into the score? Does your worksite use that same category, or only a similar name?
The country. Is the employment base U.S. employment, an OECD group, or a local workforce? A title does not move a task profile across borders.
The outcome claim. Does the source report technical potential, task change, adoption, employment, wages, hours, or a scenario? Circle one. Leave the others blank.
That last line stops the score from borrowing an outcome it did not measure.
Exposure names a possible technological contact.
Adoption names whether a system enters the work.
Impact names what happens to employment, hours, earnings, quality, or the work left behind.
Keep the three columns separate.
If the code is absent, the task source is opaque, the threshold is missing, or the country has been replaced by a familiar title, stage one does not point toward the score.
It points toward unresolved.
The Cutoff
Treat the cutoff as part of the finding.
In the OECD Skills Outlook comparison, counting only a proportion of workers in high-risk occupations as high risk produced country estimates of 4% to 10% of employment.
Assigning the occupational high-risk status to every worker in those occupations produced 19% to 48%.
Much of the classification can stay in place while the result moves from 4% to 48%.
The within-occupation assumption did the work.
That is why a score can look correct and still be wrong for your shift.
It may be treating every worker in a category as if the category had no internal differences.
Those differences are often the work.
The Ten-Task Audit
Now run stage two.
Take out a sheet of paper.
Sample ten actual tasks from your worksite.
Use a recent shift, a recent week, or a recurring cycle.
Do not use the job description.
Write verbs.
Receive a damaged record.
Compare a claim against an exception.
Explain a delay to a customer.
Correct an incomplete entry.
Decide whether a case can leave the queue.
The task has to be observable.
“Support operations” is not a task. “Reconcile the missing information before release” is.
Give every task four marks from 0 to 2.
Give technical feasibility 0 when no system can credibly do the actual task, 1 when it can do a bounded or clean version, and 2 when it can handle the normal work as it arrives.
Give adoption evidence 0 when there is no observable deployment, 1 when there is a demonstration, trial, license, or announced plan, and 2 when the relevant workflow is in routine use.
Give exception cost 0 when errors, rework, escalation, and recovery are low and contained, 1 when they require ordinary checking, and 2 when the exception can damage a customer relationship, control, audit trail, safety condition, or adjacent team.
Give human accountability 0 when nobody must meaningfully own the outcome, 1 when a person reviews some cases, and 2 when a person stays responsible for the release, decision, recovery, or explanation.
These marks are not probabilities.
Do not convert them into one fake precision number.
They force the four things a public score leaves outside the frame onto the same page.
Technical feasibility is not deployment.
Deployment is not a safe process.
A safe process is not a staffing decision.
The Exception Cost
Look down the ten rows.
The exposed work looks clean at first.
Collect the data. Make the comparison. Draft the response. Move the record.
Then the actual process appears.
The input arrives incomplete.
The customer has already been told something different.
The system cannot show why it made the recommendation.
The result must survive an audit.
The ordinary case may be cheap.
The exception carries the cost.
McKinsey’s activity analysis makes the point at a different unit.
It found 45% of paid activities technically susceptible using then-demonstrated technologies, and fewer than 5% of occupations that were candidates for full automation.
There is no contradiction.
An occupation is a bundle.
A worksite process is a bundle with handoffs, controls, escalation, and somebody who owns the bad outcome.
One ONS example is worth keeping in view.
Retail assistants can receive a high occupation-level score in the Oxford method while only 4% perform their work without social interaction in the task-based evidence.
The transaction is not the whole job.
Neither is the title.
The Two Stages
Put stage one beside stage two.
Use the published score as a local signal only when both stages point in the same direction.
Stage one must show a documented code, technology, task source, threshold, geography, and outcome claim that fit your question.
Stage two must show actual tasks with credible technical feasibility and real adoption evidence, without exception cost and human accountability carrying the workflow back to people.
That does not prove displacement.
It identifies a task family that deserves close monitoring.
If stage one says your occupation is exposed but the ten-task audit finds high exception cost, human accountability, or no deployment evidence, label the result unresolved.
If the task audit shows a live process but the public score has the wrong code or hides its threshold, label the result unresolved.
Unresolved is not a soft answer.
It is an evidence condition.
Write the missing evidence beside it: workflow coverage, task hours, error rate, rework, customer escalation, staffing, or the human work after deployment.
Then watch the thing that would resolve it.
One OECD study linked a task-based exposure measure to occupational growth across 21 countries through 2019.
The riskiest half of occupations grew 6%; the least risky half grew 18%.
That is an association with later relative occupational growth.
It does not establish that automation caused the difference.
Trade, demand, structural change, and workforce composition can travel with the score.
The same study found no support for net job destruction at the country level.
A relative occupational signal and aggregate employment growth can coexist.
Your worksite can also change without following the headline.
It can reduce task time, move work to another team, add checking, expand output, improve quality, leave headcount unchanged, freeze hiring, or cut jobs.
The score cannot choose among those outcomes.
What the Label Hides
Do not let the audit become an argument that your job is safe.
That is not what it can establish.
Management can still use a crude occupational score, because it is cheap, legible, and available before anyone has measured the work.
You often cannot see the procurement file, the data integration work, the error record, or the staffing plan that would settle stage two.
Say that plainly.
Keep the unresolved label until there is evidence.
Then study the mismatch.
It shows where the occupational label has been standing in for work the model did not see.
That may be exception handling.
It may be accountability.
It may be the customer who arrives with the wrong problem.
It may be the old system that still has to be reconciled before anything can leave the building.
Make the missing work visible.
A red badge is a screening signal that arrived before the evidence did.
Do not let a website finish the calculation for you.
Minimum Viable Action
Before you sleep tonight, find the occupation code behind one exposure claim about your work.
Write it down.
Under it, write one task the code misses.
Leave a space beside that task for the evidence that would settle it.
What you will be able to say afterwards
- An occupational title must be matched to a documented occupation code, code system, and version before its exposure score can be interpreted.
- A high-risk threshold and a task source define the category a published exposure score reports.
- Technical potential, workplace adoption, and employment loss are separate claims requiring separate evidence.
- Ten observed tasks can be checked for technical feasibility, adoption evidence, exception cost, and human accountability.
- Within-occupation task variation can make a correct title and an incorrect occupational score coexist at one worksite.
- Conflicting published and worksite evidence should be labelled unresolved rather than converted into a personal job-loss forecast.
Questions this raises
What does a red automation score for an occupation actually mean?
A red occupational score usually means that a study classified an occupation above its stated exposure threshold under a particular technology premise. It is a screening signal about a coded occupational category, not a personnel decision. The score does not establish that a particular employer has adopted a system, that a particular worker performs the category's average tasks, or that employment will be lost.
Why can the job title be right but the exposure score still be wrong?
A title can match an occupation label while the underlying code, code version, or task organization does not match the worksite. O*NET contains 923 data-level occupations within 1,016 titles, and workforce records can use another classification. A crosswalk can merge different local roles and flatten their differences. A title match is therefore not evidence that the published task profile applies.
What information has to be recorded before using an exposure study?
A usable record identifies the occupation code and code system, technology definition, task source, scoring rule or threshold, country or geography, employment base, and outcome claim. It also records the data year and method version when they are available. Those fields show whether the figure measures technical potential, adoption, task change, employment, wages, or another outcome.
What does the 0.70 cutoff in an automation study mean?
The 0.70 cutoff is a rule for placing a modelled occupation score in a high-risk category. Frey and Osborne applied it to 702 detailed U.S. occupations and reported that 47% of total U.S. employment lay in that category. The cutoff does not mean that a worker has a 70% chance of dismissal. It defines a category within one measurement method.
Why does one study say 47% while another says 9%?
The figures measure different objects. Frey and Osborne classify occupations and place 47% of U.S. employment in a high-risk category at a 0.70 cutoff. Arntz, Gregory, and Zierahn use worker-task information and estimate 9% of jobs automatable on average across 21 OECD countries. Task methods allow workers within one occupation to have different task profiles.
How can a worksite test an occupational exposure score?
A worksite can sample ten actual tasks and score each from 0 to 2 for technical feasibility, adoption evidence, exception cost, and human accountability. The exercise records whether a system can perform the real task, whether deployment evidence exists, what error handling costs, and who remains responsible. It is a structured prompt for inspection, not a validated prediction instrument.
What counts as adoption evidence instead of a capability claim?
Adoption evidence concerns a system entering the actual workflow: a specified process, usable inputs, integration, accountable review, active use, and observed effects on hours, quality, staffing, or task allocation. Technical capability establishes only that a task may be possible. Cost, benefits beyond labour savings, competition with labour, technology development, and social and regulatory acceptance affect whether deployment occurs.
Can a job contain a great deal of automatable work without disappearing?
Yes. McKinsey reports that then-demonstrated technologies could automate 45% of paid activities and that about 60% of occupations contained at least 30% susceptible activity, while fewer than 5% of occupations were candidates for full automation. Jobs bundle activities. A routine activity can change while exceptions, oversight, customer contact, and accountable decisions remain.
Why do two estimates with the same high-risk occupations produce different totals?
The within-occupation assumption changes the total. OECD Skills Outlook reports lower-bound estimates of 4% to 10% when only a proportion of workers in high-risk occupations are counted as high risk, and upper-bound estimates of 19% to 48% when all workers in those occupations are counted. The endpoint is a scenario result, not a count of actual job loss.
When should a published exposure score be called unresolved?
A published score is unresolved when its code, task source, threshold, technology premise, geography, or outcome claim does not fit the worksite, or when observed tasks point in another direction. The label preserves the evidence gap. It is more accurate than calling a worker safe or replaceable when adoption data, exception costs, and accountable review arrangements remain unknown.
Does higher exposure prove that automation caused lower employment growth?
No. An OECD study linked a 2012 task-based measure to 38 occupations in 21 countries through 2019 and found that the riskiest half grew 6% while the least risky half grew 18%. The association supports limited predictive validity for relative occupational growth. It does not isolate automation from demand, trade, structural change, or workforce composition.
What does a worker do when management uses a crude score anyway?
A worker can document the occupational code, the actual task mix, exceptions, controls, customer contact, and the human work left after any proposed deployment. That record does not compel management to use better evidence. It does create a concrete account of where the score stands in for work the classification did not observe, and it separates exposure from an asserted staffing outcome.