Which office roles can AI take today? The evidence by role

On published evidence, AI can take most of the volume in three office roles today: tier-1 support by chat and email, document processing and data entry, and bookkeeping preparation. A person still handles about 10 to 35 percent of the work in each, phone work has thinner evidence, and for outbound sales, software, legal and analysis work the measured results say assist, not replace.

Which roles have published automation rates?

The table lists only numbers a vendor, customer or research group has put in writing. They come from mature deployments, so read them as ceilings, not as a first month.

A caution on the first row: in May 2025 Klarna's chief executive said cost had been "a too predominant evaluation factor" and the result was "lower quality", and Klarna hired human agents again (Fortune). The volume stayed with the AI; the human path was rebuilt.

Which roles are assist only on the evidence?

These are drafting and research aids. Nobody publishes a rate at which they finish the job alone; the error rates above are the reason.

Which roles have no evidence yet?

Front desk and phone intake is the biggest gap. AI receptionist vendors publish prices, answer rates and cost per call, but not the share of calls they finish without a transfer, and Smith.ai says most firms start on its hybrid plan with people behind the AI (Smith.ai). Content drafting has no containment figure because an editor approves every piece. Recruiting screening tools make result claims, but recruitment and selection sit on the EU AI Act's high-risk list for employment, so a person decides about a candidate whatever the tool can do (Annex III). For these roles the honest number is the one your pilot measures.

How should you read the task-exposure studies?

Three kinds of numbers get mixed up. Exposure studies measure what could be done. The 2023 paper by Eloundou, Manning, Mishkin and Rock found about 80 percent of US workers could have at least 10 percent of their tasks affected by language models, about 19 percent at least half, and that 15 percent of all tasks could be done significantly faster by a model alone, rising to 47 to 56 percent with software built on it (arXiv).

Usage studies measure what people hand over. Anthropic's March 2026 analysis found observed coverage of 75 percent of tasks for computer programmers and 67 percent for data entry keyers, zero coverage for 30 percent of workers, and no systematic rise in unemployment among exposed workers since late 2022 (Anthropic). The Microsoft team that scored 785 occupations from 200,000 Copilot conversations says its metrics "should not be misconstrued or misrepresented as measuring the ability of AI to replace jobs" (GitHub).

Outcome studies measure what happened. In the New York Fed's September 2026 survey, 61 percent of service firms used AI, 4 percent had laid anyone off because of it, about 15 percent had hired fewer people and about 13 percent more (New York Fed). The Stanford payroll study found employment of 22 to 25 year olds in the most exposed jobs down about 11 percent from November 2022 to June 2026 while the least exposed grew about 10 percent, mostly through fewer hires, not dismissals (Stanford Digital Economy Lab).

Put the usage number, measured in your own company, in front of a payroll decision. The exposure number tells you where to look, not what to cut.