Why some letters still land on a human desk
The work is simple to describe and harder to automate cleanly: read what someone wrote, find the records behind it, and write back. A lot of that is pattern work. A billing error, a delinquent account, a request for credit information, an unsatisfactory service complaint. Language software is good at patterns, which is why the answer to whether AI will replace correspondence clerks is less comfortable here than in hands-on jobs.
But two parts of the job resist. The first is checking records for completeness and accuracy before a reply goes out. Software can pull a file; deciding the file is thin, wrong, or contradicted by an earlier letter is judgment. The second is the correspondence nobody wants: a disputed damage claim, an angry second letter, a case where the company’s own paperwork is at fault. Those replies carry money and liability, and someone has to own the wording.
Scale matters too. The Bureau of Labor Statistics counts about 4,290 people in this occupation in the US, with median pay near $46,800, and projects employment falling 5.6% between 2025 and 2035 (BLS, 2025). That decline started long before generative AI, driven by self-service portals and shared service centers. The honest reading is task erosion plus fewer new openings, not an occupation switching off. Coverage, our measure of how much task time AI can handle today, sits at 50 out of 100. How coverage is measured explains what goes into that figure.
What software handles, what it drafts, what people keep
Start with the tasks already done by machine. Completing form letters in response to routine requests is template work, and so is sorting incoming mail by subject and routing it to the right department. Those are the clearest cases, and they account for 24% of task time on this page’s split.
Next, the assisted group. Composing replies about incorrect billing or delinquent accounts is a drafting job, and models are fast first-drafters. Same with compiling data from records into periodic reports: the pull and the summary are quick, the sense-check is not. Work in that middle band comes to 64% of task time, and it is where most clerks are already spending their day.
Then the residue. Reviewing gathered records for accuracy, and preparing the correspondence that follows a contested damage claim or a complaint about service, is work we score as needing a person: 12% of task time. It is a narrower base than clerks would like, which is the real pressure in this job. For the wider picture, see the information and record clerks family page and the administrative support sector.
What has actually been tested
Not much, specific to this job. The quality-parity grade shown above is D. In plain terms: no published study has put AI head to head with correspondence clerks on their own correspondence, so we give no parity number rather than guess one. The general-purpose benchmarks that exist test writing and summarizing in the abstract, not whether a reply to a damage claim survives a supervisor’s review.
What would settle it is a narrow test with a clear design: a batch of real incoming letters, replies produced by clerks and by software, and blind scoring by the people who normally approve outgoing mail, on accuracy, tone and whether the records were read correctly. Error rates on cases where the file was incomplete would be the most useful number of all. Until something like that is published, treat confident parity claims about clerical writing with care. How we grade quality parity sets out the A to D scale.
When the picture could shift
Most likely between 2035 and 2045 (8 in 10 of our scenarios). Two things could pull that earlier. Correspondence already arrives as text, so there is no sensing problem to solve; the physical share of this job is small and sits in the fixed-automation tier, meaning mail handling and printing, not robots. And the cost gap shown in the cost panel above is wide enough that a vendor pitch is easy to make to a shared service center.
Two things hold it back. Liability is the big one: letters about claims, credit and billing create a record, and employers keep a named person in the approval loop for a reason. The second is messy source data. Records split across old systems, scanned attachments and partial case notes are where automated replies go wrong, and fixing that plumbing is slower and less exciting than buying a model. How the replacement-year range works explains what the window does and does not claim.
What to do: If your week is mostly form letters, ask now to be moved onto disputed cases and record quality checks, where the judgment sits.
How to stay needed in this job
Lean into the three tasks that hold up. Review records for completeness and accuracy, and get known as the person who catches a bad file before it becomes a bad letter. Own the contested correspondence: claims, credit disputes, service complaints where the first reply failed. And take responsibility for approving outgoing mail, including drafts that software produced.
Two skills are worth real effort. First, editing machine drafts fast and knowing when to throw one away, which is different from writing from scratch. Second, basic records and data hygiene, including how your case systems connect, because that is what decides whether automation helps or embarrasses the company. The guide on AI skills employers want covers the first; future-proofing your career covers the move after this one.
If you are weighing a sideways step, the nearest work is worth comparing rather than assuming. Look at File Clerks, Receptionists and Information Clerks and Data Entry Keyers, then put two of them side by side on our compare tool. Our headline figure for this job is 59 out of 100 (higher is safer); the methodology shows how it is built from open data, and the jobs most at risk list and full rankings put it in context.