US dataswitch to UK
Correspondence Clerks
ensuring that money collected is properly, processing orders for goods requested in correspondence and compiling data from records to prepare periodic reports. If that's your week, this page is about your job.
The honest answer
Most tasks in this job are the kind AI has learned to do: reading incoming correspondence to ascertain nature of writers' concerns and to determine disposition of correspondence. The tasks, though, are not you.
It would be a lie to soften that, and another lie to promise a rebuilt version of this job inside this job.
So the hope here is what you already carry: the judgment you bring to money collected is real, and the moves below are built from it. The first step is down this page.
Your week, as this page understands it
Compose letters or electronic correspondence in reply to requests for merchandise, damage claims, credit and other information, delinquent accounts, incorrect billings, or unsatisfactory services. Duties may include gathering data to formulate reply and preparing correspondence. The job title says “correspondence clerks”. The real job is the part underneath: ensuring that money collected is properly. That is the thing someone has to be right about.
The exposed part of this job is specific, and we won’t pretend it is coming back. But correspondence clerks is not one task. It is 17 scored ones, and they are nowhere near equally exposed. The most durable of them, on this evidence, is ensuring that money collected is properly, and the ledger below shows exactly why.
Where the work sits, by task weight
- shifting to AI
- 79%
- changing shape
- 6%
- staying human
- 15%
These bars are tasks changing hands, not people being counted out. The ledger below shows which.
Whole-job exposure score 76 out of 100 (71–81 allowing for uncertainty): high exposure, across 17 scored tasks. The number is the support for the sentence above it, not a headline about anyone’s future.
How we know this
What is measured: Every published task statement for correspondence clerks is rated on five dimensions: can a model produce the output, does the work need a body in a room, does it need a legally accountable person, does it depend on a person being trusted in the moment, and how much data exists. A published formula turns those five ratings into the score; the model never writes the number.
How the bar is built: Each task’s share of the bar is its published importance weight, so a task you do all day counts for more than one you do twice a year.
Release: 2026-q4.1, scores computed 2026-08-05. Read the full method.
Your job, task by task
These are the official task statements for this occupation, in plain English, sorted by what the evidence says is happening to each one. The official wording sits under every line so you can check the rewrite against it.
- 2 of this occupation's scored task statements carry a score that was measured under a different occupation's context, because the statement is shared between jobs and has only been scored once so far. Each one names the occupation it was measured under in the free routes below; none is presented as a measurement for this job.
- Task clusters are not derived in this release, so the task-cluster field is empty and no Collab365 Space routing is attached to this occupation yet.
Shifting to AI
13 tasksTasks today’s tools can already do most of. This is the part we will not soften: where these rows are the bulk of your week, the week changes.
Maintaining files and controlling records to show correspondence activities
This is reading one thing and writing another: files in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Maintain files and control records to show correspondence activities.” (O*NET task statement)
How this row was scored
Exposure score: 69 out of 100 (62–76 allowing for uncertainty): high exposure, medium confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Keeping the correspondence log up to date is simple record work, though some files are still physical folders.
The five ratings: output a model can produce 4/4 · needs a body in a room 1/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Reading incoming correspondence to ascertain nature of writers' concerns and to determine disposition of correspondence
This is reading one thing and writing another: correspondence in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Read incoming correspondence to ascertain nature of writers' concerns and to determine disposition of correspondence.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Reading a letter to work out what it is about and where it should go is classic sorting work for language tools.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Gathering records pertinent to specific problems
This is reading one thing and writing another: records pertinent in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Gather records pertinent to specific problems, review them for completeness and accuracy, and attach records to correspondence as necessary.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Finding the right records, checking they are complete and attaching them is routine work software does well.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Preparing documents and correspondence, such as damage claims, credit and billing inquiries, invoices and service complaints
This is reading one thing and writing another: documents in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Prepare documents and correspondence, such as damage claims, credit and billing inquiries, invoices, and service complaints.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Producing claims, invoices and inquiry letters from standard formats is exactly the kind of drafting tools do well.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Routing correspondence to other departments for reply
This is reading one thing and writing another: correspondence in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Route correspondence to other departments for reply.” (O*NET task statement)
How this row was scored
Exposure score: 69 out of 100 (62–76 allowing for uncertainty): high exposure, medium confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Deciding which department should answer a letter and passing it on is straightforward sorting, though paper still gets carried.
The five ratings: output a model can produce 4/4 · needs a body in a room 1/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Compiling data from records to prepare periodic reports
This is reading one thing and writing another: data in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Compile data from records to prepare periodic reports.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Pulling figures out of records into a regular report is a standard, repeatable job for software.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Changing shape
1 taskTasks where the machine takes the producing and a person keeps the checking, the deciding, or the answering-for-it. For most jobs this is the biggest group, and it is where "transformation, not termination" is literally visible.
Presenting clear and concise explanations of governing rules and regulations
The software now makes the first pass at clear, but the part that matters is a person saying it and standing behind it. So the job becomes checking and deciding rather than producing.
importance 4 · SupplementalSource: “Present clear and concise explanations of governing rules and regulations.” (O*NET task statement)
How this row was scored
Exposure score: 53 out of 100 (46–60 allowing for uncertainty): partial exposure, medium confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; the value is that a specific person does it.
The rating behind it: Clear explanations are easy to write, but people asking about rules that affect them want a person.
The five ratings: output a model can produce 3/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 2/4 · how much data exists 3/4.
Staying human
3 tasksTasks that stay with a person, because they happen in the physical world, because the rules need someone accountable, or because the value is that a specific person does them.
Ensuring that money collected is properly
This work happens in the physical world: money collected, in a real place. Software cannot follow it there.
importance 4 · SupplementalSource: “Ensure that money collected is properly recorded and secured.” (O*NET task statement)
How this row was scored
Exposure score: 25 out of 100 (18–32 allowing for uncertainty): low exposure, medium confidence.
Why it sits in this group: the same decision, made over and over; mistakes that are cheap to catch; work that happens in the physical world.
The rating behind it: Recording takings is simple bookkeeping, but physically securing the money has to be done by someone there.
The five ratings: output a model can produce 2/4 · needs a body in a room 2/4 · needs an accountable person 1/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Conferring with company personnel regarding feasibility of complying with writers' requests
The value here is that a specific person handles company personnel regarding feasibility and stands behind it. That is earned, not computed.
importance 3 · SupplementalSource: “Confer with company personnel regarding feasibility of complying with writers' requests.” (O*NET task statement)
How this row was scored
Exposure score: 35 out of 100 (28–42 allowing for uncertainty): low exposure, medium confidence.
Why it sits in this group: the value is that a specific person does it.
The rating behind it: Asking colleagues whether an unusual request can be met takes negotiation and judgment about internal appetite.
The five ratings: output a model can produce 2/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 2/4 · how much data exists 2/4.
Preparing records for shipment by certified mail
This work happens in the physical world: records, in a real place. Software cannot follow it there.
importance 3 · SupplementalSource: “Prepare records for shipment by certified mail.” (O*NET task statement)
How this row was scored
Exposure score: 19 out of 100 (15–23 allowing for uncertainty): minimal exposure, high confidence.
Why it sits in this group: the same decision, made over and over; mistakes that are cheap to catch; work that happens in the physical world.
The rating behind it: Preparing a package for certified mail means physically assembling, labeling and lodging it at the counter.
The five ratings: output a model can produce 3/4 · needs a body in a room 3/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Show the other 7 tasks
Typing acknowledgment letters to persons sending correspondence
shifting to AIThis is reading one thing and writing another: acknowledgment letters in, a record out. That is the shape today's tools are built for.
importance 4 · SupplementalSource: “Type acknowledgment letters to persons sending correspondence.” (O*NET task statement)
How this row was scored
Exposure score: 100 out of 100 (96–100 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Acknowledgment letters follow a fixed template, which is the clearest case of writing that software already produces fully.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 4/4.
Completing form letters in response to requests or problems identified by correspondence
shifting to AIThis is reading one thing and writing another: form letters in, a record out. That is the shape today's tools are built for.
importance 3 · SupplementalSource: “Complete form letters in response to requests or problems identified by correspondence.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Filling in form letters from the employer’s own templates is repetitive drafting that tools complete reliably.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Composing letters in reply to correspondence concerning such items as requests
shifting to AIThis is reading one thing and writing another: letters in, a record out. That is the shape today's tools are built for.
importance 4 · CoreSource: “Compose letters in reply to correspondence concerning such items as requests for merchandise, damage claims, credit information requests, delinquent accounts, incorrect billing, or unsatisfactory service.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Writing replies about billing, claims and complaints follows familiar patterns that language tools reproduce at a usable standard.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Computing costs of records
shifting to AIThis is reading one thing and writing another: costs of records in, a record out. That is the shape today's tools are built for.
importance 4 · SupplementalSource: “Compute costs of records furnished to requesters, and write letters to obtain payment.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Working out charges and writing a payment request letter is calculation plus standard letter writing.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Processing orders for goods requested in correspondence
shifting to AIThis is reading one thing and writing another: orders in, a record out. That is the shape today's tools are built for.
importance 4 · SupplementalSource: “Process orders for goods requested in correspondence.” (O*NET task statement)
How this row was scored
Exposure score: 93 out of 100 (89–97 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Turning a written request into an order in the system is routine data work already largely automated.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Reviewing correspondence for format and typographical accuracy
shifting to AIThis is reading one thing and writing another: correspondence in, a record out. That is the shape today's tools are built for.
importance 4 · SupplementalSource: “Review correspondence for format and typographical accuracy, assemble the information into a prescribed form with the correct number of copies, and submit it to an authorized official for signature.” (O*NET task statement)
How this row was scored
Exposure score: 81 out of 100 (77–85 allowing for uncertainty): very high exposure, high confidence.
Why it sits in this group: reading one thing and writing another; the same decision, made over and over; mistakes that are cheap to catch.
The rating behind it: Checking format and typing accuracy and assembling documents for signature is routine document work, with the signature itself elsewhere.
The five ratings: output a model can produce 4/4 · needs a body in a room 0/4 · needs an accountable person 1/4 · needs to be trusted in the moment 0/4 · how much data exists 3/4.
Compiling data pertinent to manufacture of special products for customers
shifting to AIThis is reading one thing and writing another: data pertinent in, a record out. That is the shape today's tools are built for.
importance 4 · SupplementalSource: “Compile data pertinent to manufacture of special products for customers.” (O*NET task statement)
How this row was scored
Exposure score: 68 out of 100 (61–75 allowing for uncertainty): high exposure, medium confidence.
Why it sits in this group: reading one thing and writing another; mistakes that are cheap to catch.
The rating behind it: Pulling together data about a custom product order relies on internal product knowledge, but the compiling itself is routine.
The five ratings: output a model can produce 3/4 · needs a body in a room 0/4 · needs an accountable person 0/4 · needs to be trusted in the moment 0/4 · how much data exists 2/4.
What this job pays, and how many people do it
- Median pay
- $46,800a year, the middle of the range, so half earn more and half earn less.bls-oews, 2025 · May 2025 estimates (national_M2025_dl.xlsx)
How we know this
Source: bls-oews
Reference period: May 2025 estimates (national_M2025_dl.xlsx)
Rounding: Shown as published.
- People doing this job
- 4,290in the US, 2025.bls-oews · May 2025 estimates (national_M2025_dl.xlsx)
What is deliberately not here: a forecast of how many of these jobs exist in ten years. Where an official projection exists for a market we publish it with its vintage; where it does not, we leave the space empty rather than borrow the other country’s number.
Why this is shifting
The reason is boringly specific. Most of what is shifting here is reading one thing and writing another: correspondence in, a record out. The rows above are exactly that shape: reading incoming correspondence to ascertain nature of writers' concerns and to determine disposition of correspondence and maintaining files and controlling records to show correspondence activities. What it cannot do is be there in the room, and that is still where money collected gets done. Which is why this page talks about your tasks changing, not your job ending.
Your move
Over a pint: what I’d tell you if you were my friend
The exposed part of your job is the biggest part, and I am not going to dress that up: reading incoming correspondence to ascertain nature of writers' concerns and to determine disposition of correspondence is work today's tools do quickly and cheaply, and that is not coming back.
So, given all that: 79% of this job's task weight sits in rows the software is already learning, 6% in rows that change shape rather than disappear, and 15% in rows it is nowhere near. That is the position, measured across 17 scored tasks. It is not a forecast about you.
What you have that the software does not is ensuring that money collected is properly, plus the years of knowing when something looks wrong before you can say why. That is the raw material for everything below.
This week: one thing
Sit on the machine's side of the desk. Pick one real piece of correspondence you would normally do yourself, let whatever software you already have take the first pass at it, and then go through what it produced and write down every single thing it got wrong. One evening this week. Do not fix anything yet. Just catch it.
- What you end up holding
- a written list of the machine’s mistakes, in your handwriting
- How long it takes
- an evening, or an hour if you pick one job rather than one client
If you can’t run software on your employer’s or your clients’ data: Do the same hunt on paper. Take one printed piece of correspondence, work through it the way you always do, and mark every point where you made a call rather than followed a rule. Same evening, same list, nothing to log into and nobody to ask permission from. That list is the same artifact: it is the judgment written down.
Over the next 90 days
Change one sentence about what you do. Not on a CV. Out loud, to whoever pays you. From “my job is reading incoming correspondence to ascertain nature of writers' concerns and to determine disposition of correspondence” to “I check what the software does and tell you what it means.” Your error list from this week is what makes that sentence true instead of a claim, so use it: show it once, to one person, inside the next ninety days. Same skills, priced as judgment rather than as typing.
Over the next 12 months
Walk toward the end of this job that answers for things, and get it recognised. Pick the one part of ensuring that money collected is properly you are already best at, and spend the year making it formal: a qualification, a named responsibility, a specialism people ask for by name. Price it honestly: that is evenings, it is months rather than weeks, and the seats are competitive because everyone in your position is looking at the same door. Before you pay for anything, use CareerOneStop - Find local training. It is free, it is the Labor Department's own service, and it is listed below with the rest of the free routes.
The roads out of here, and why I am not sending you down them
I looked at the obvious moves out of this job, and here is what I found.
I checked the 12 nearest US occupations to correspondence clerks (nearest by the work that AI is not taking, not by job title), and none of them survived. The closest was postal service clerks: only about 2% of its durable work is work you already do. I am not going to pretend that is comfortable news: 79% of your own task list is already in the top exposure band. But the answer on this evidence is not a sideways jump into a job with the same problem. It is to walk toward the end of this one that answers for things. “ensure that money collected is properly recorded and secured” is the part that stays with a person, and spending a year getting formally recognised for it beats a standing start somewhere else.
How that was checked: this job was compared against all 830 US occupations in this release on their official task statements, and the 12 nearest were examined one by one. A move that turns on an industry, an employer or a qualification rather than on the work itself will not show up in a check like that. And this release carries no licence register, so anything you are weighing needs that looked up separately.
3 moves I checked and rejected
These are the obvious-looking jumps. They are here with their reasons rather than quietly dropped, because the ones that fail are worth knowing about. It is one less thing to turn over at night.
Postal Service Clerks
Why it looked obvious: It came up as a near neighbour on the work AI is not taking: you already ensure that money collected is properly recorded and secured, and their equivalent is to cash money orders. Across both published task lists that is about 2% of the durable work in that job.
Why I am not recommending it: Almost none of it is work you already do: about 2% of the durable side of that job. That is a different job, not a next step.
Secretaries and Administrative Assistants, Except Legal, Medical, and Executive
Why it looked obvious: It came up as a near neighbour on the work AI is not taking: you already prepare records for shipment by certified mail, and their equivalent is to prepare and mail checks. Across both published task lists that is about 1% of the durable work in that job.
Why I am not recommending it: Almost none of it is work you already do: about 1% of the durable side of that job. That is a different job, not a next step. I will not move you off one melting floe onto another: 58% of its own task list already scores in the top exposure band (61/100 in this release), so the same software is eating it.
Legal Secretaries and Administrative Assistants
Why it looked obvious: It came up as a near neighbour on the overall shape of the two task lists, but nothing in your day matched a specific piece of theirs closely enough to name.
Why I am not recommending it: The two task lists look alike from a distance and share almost nothing close up: no single piece of their work matched a piece of yours. That is a resemblance, not a route. I will not move you off one melting floe onto another: 69% of its own task list already scores in the top exposure band (65/100 in this release), so the same software is eating it.
What I’d stop worrying about
A friend tells you what not to spend fear on. This is that list.
The headline number you read somewhere
The big “X% of jobs” figures are about the whole economy, not about you. The number that describes your job is on this page: 79% of its task weight, across 17 scored tasks. Every row behind it is printed above with the source. Check ours; ignore theirs.
“It’s too late for me to become something else”
You are not starting from zero, and the page shows why: ensuring that money collected is properly is work the software cannot do and you already do it. The move above is a repricing of what you know, not a new career. Nobody who has just left college has that.
“I should learn to code”
Almost certainly not. The value in your job is moving toward checking, deciding and answering for the output, not toward writing the software. A weekend of Python will not change your position; the error list from this week will.
The “obvious” next job everyone suggests
I checked the obvious moves and most of them did not survive. The reasons are printed with the routes above, including the pay and the gate. A move that fails on the numbers is worth knowing about so you can stop turning it over at night.
You are reading the United States figures
The United Kingdom splits this work across more than one official group, of which Sales administrators is the closest. The pay and employment figures are not directly comparable, and we do not average them together.
Switch to the United Kingdom page →partial match
The other groups this work is counted across:
In UK official statistics this job is counted as Sales administrators and Customer service occupations n.e.c.. Pay is shown separately for each of those groups (medians cannot be averaged together), while the task list and the scores on this page are for this group only.
Your route through this
Where to go next, and what it costs
Free, and complete
The moves above cost nothing. These are the real services that go with them: public, government-funded, and free at the point of use. Nothing on this page is behind an email address or a payment.
Anywhere in the US:
CareerOneStop - Find local training
Search what's running near you, from the Labor Department's own database, before anyone sells you a course.
Free to search; individual programs vary, and some are funded
Anywhere in the US:
An American Job Center will sit down with you for free. Find yours by ZIP code.
Free
Anywhere in the US:
CareerOneStop - Licensed occupations finder
Check what your state actually requires before you pay for anything.
Free
Anywhere in the US:
Free
No Space for this job, but one for what is happening to it
Nothing Collab365 runs is built for correspondence clerks, and we are not going to point you at the nearest one and call it a fit.
There is one that is not about a job title at all. The AI Authority is about being the person who directs these tools at work rather than the person they get compared to. That is worth saying here, because 79% of the work on this page is already inside what they can do.
7 days free, no card needed. Explore up to 2 Spaces before you choose a plan: you pick a plan later, not now.
The AI Authority is a general community about working with AI, not a course for correspondence clerks. You do not need it to act on anything here: the moves above cost nothing and stand on their own. The data on this page is the same either way.
Noted, and thank you. We’ll email you if a Space for correspondence clerks launches. Nothing else.
That did not look like an email address, so nothing was saved. Have another go below.
We could not save that. The fault is ours, not yours, and your address was not stored. Please try again later.
No deadline on any of this. The page will still be here, and the data is refreshed on a published schedule rather than when someone wants a headline.
Questions people ask about this job
- Will AI replace Correspondence Clerks?
- Not as a job, but it is already doing parts of the work. Across the 17 official task statements scored for Correspondence Clerks (United States, SOC 43-4021), 79% of the importance-weighted core work is made of tasks today's AI could already do most of. The overall exposure score is 76 out of 100 (range 71–81, band: high). That is a statement about tasks, not about headcount: this measures what AI could do, not whether any employer adopts it, whether the law allows it, or whether doing the routine parts faster creates more demand for the human parts. Figures are from release 2026-q4.1.
- Which tasks in “Correspondence Clerks” can AI already do?
- The highest-scoring tasks in release 2026-q4.1 are: “Type acknowledgment letters to persons sending correspondence” (100/100, very high); “Prepare documents and correspondence, such as damage claims, credit and billing inquiries, invoices, and service complaints” (93/100, very high); “Compile data from records to prepare periodic reports” (93/100, very high). Each score comes from five published 0–4 ratings turned into a number by a published formula, and each carries the model's one-sentence reason on the page.
- Which tasks in “Correspondence Clerks” stay human?
- About 15% of this job's task weight sits in work that scores low for AI exposure. The lowest-scoring tasks in release 2026-q4.1 are: “Prepare records for shipment by certified mail” (19/100, minimal); “Ensure that money collected is properly recorded and secured” (25/100, low); “Confer with company personnel regarding feasibility of complying with writers' requests” (35/100, low). Low scores usually mean the task needs a body in a room, a legally accountable human, or trust built in real time. Those are the three things the scoring rubric treats as gates rather than obstacles.
- What should someone working in “Correspondence Clerks” do about AI?
- Start from the ledger rather than the headline: 79% of this job's weighted core work is exposed, and roughly 15% is not. The practical move is to spend more of your week on the tasks that score low, the ones above, and to get fluent at directing AI through the tasks that score high, because those are the parts that change whether or not you are ready for them. This page does not predict your job, and nothing here is career advice tailored to you: the score describes the occupation, not the person.
- How is the AI exposure score for Correspondence Clerks calculated?
- Each official task statement for the occupation is rated on five published 0–4 dimensions (output replicability, physical embodiment, licensed accountability, real-time human trust, and data availability) by claude-opus-5 using scoring prompt task_scoring_v1.0. The model never writes the score; a published formula turns the five ratings into a 0–100 number, so every score can be recomputed by hand. The occupation figure is the importance-weighted mean across 17 scored tasks. The prompt, the rubric, the formula and the full dataset are published at https://futureproof.collab365.com/method and https://futureproof.collab365.com/data/2026-q4.1 under CC BY 4.0.
Where these numbers come from
Worth knowing about these figures
- The match between this job and its counterpart in the other country is partial, so the two pages count slightly different groups of people.
- 2 of this occupation's scored task statements carry a score that was measured under a different occupation's context, because the statement is shared between jobs and has only been scored once so far. Each one names the occupation it was measured under in the free routes below; none is presented as a measurement for this job.
- Task clusters are not derived in this release, so the task-cluster field is empty and no Collab365 Space routing is attached to this occupation yet.
- Task statements
- onet-dbProcessing: catalogue-bridge → onet-im-rt-weighting → task-scoring → score-aggregation
- Task weights
- onet-db (im-rt)
- Scores
- Rubric task_scoring_v1.0, model claude-opus-5, computed 2026-08-05.
- Pay and employment
- bls-oews (May 2025 estimates (national_M2025_dl.xlsx))bls-oews (May 2025 estimates (national_M2025_dl.xlsx))
Figures on this page come from release 2026-q4.1, published 2026-08-05. Every release keeps its own permanent address, so a figure you cite in March is still there, unchanged, in November.
The plain-English wording on this page is assembled directly from the task statements and the published ratings, not written by hand for this occupation. That is why it is specific, and it is also why we say so.
The routes and free resources further up are today’s, not the release’s (last reviewed 2026-08-05). A route is an offer, not a historical fact, so it moves on its own clock.
Using these figures?
Cite this
Everything on this site is published under CC BY 4.0. Quote it, chart it, sell something built on it. Just say where it came from, and cite the dated release rather than the site, so the figure you quote stays checkable.
Plain text
Collab365 (2026). Collab365 Futureproof: task-level AI exposure for US and UK occupations, release 2026-q4.1 (methodVersion 2.0.0, promptVersion task_scoring_v1.0). https://futureproof.collab365.com/data/2026-q4.1. Licensed CC BY 4.0. Built with O*NET data (USDOL/ETA, CC BY 4.0); ONS data (Open Government Licence v3.0); GAISI task framework (arXiv:2507.22748, MIT); BLS data (public domain).
BibTeX
@misc{collab365futureproof2026q41,
title = {Collab365 Futureproof: task-level AI exposure for US and UK occupations, release 2026-q4.1},
author = {{Collab365}},
year = {2026},
url = {https://futureproof.collab365.com/data/2026-q4.1},
note = {Release 2026-q4.1, methodVersion 2.0.0, promptVersion task_scoring_v1.0, CC BY 4.0}
}Data as of release 2026-q4.1, published . Releases never change after publication; when the figures move, a new dated release is published beside this one and this one stays exactly where it is.