Why the hard part is deciding what to ask
Survey research looks like data work. Most of the difficulty sits earlier, in choosing what to measure and who to ask.
A questionnaire is a measuring instrument. Wording, question order and answer options all move the result. Writing and testing question wording is a judgment task, not a drafting task: you are guessing how a stranger will read a sentence, then checking that guess in a pretest. Choosing a sample and setting weights carries the same weight of judgment. If the people who answer differ from the people who don’t, the numbers drift, and someone has to decide how far to trust them.
The second half of the job is defending the result. Survey researchers brief clients, explain limits, and answer challenges from reviewers, reporters and regulators. A model can draft that explanation. It cannot sign it. That accountability is why the question “will AI replace survey researchers” lands differently here than in jobs where output is checked by a machine rather than by a reputation.
Work volume matters too. BLS counts about 8,290 survey researchers in the US with median pay of $69,460, and projects employment falling 4.8% between 2025 and 2035 (BLS, 2025). Fewer roles is a different story from vanishing work, and it mostly hits the entry-level rungs where coding and tabulation used to live.
What AI drafts, what it assists with, and what stays with people
Some tasks already move cleanly to software. Coding written-in answers into categories and producing first-pass tabulations and summary tables are the clearest cases; a model can do both in minutes and repeat the rules consistently. That group accounts for 24% of measured task time. Our overall coverage score for this job, which asks how much of the work AI can handle today, reads 43 on a 0-to-100 scale; the Can AI Do It score page explains how that is built.
A second group is assisted rather than handled. Drafting candidate question wording and running routine analysis of the finished dataset both go faster with a model in the loop, and both still need a researcher to check the logic, catch a leading question and decide which cut of the data is honest. Assisted work covers 44% of task time.
The rest stays with a person. Deciding the research design with a client, judging whether a low-response survey can carry the claim being made, and presenting and defending findings to people who may dispute them sit in that group, which is 32% of the work. None of it needs a robot: the physical share of this job is zero, so adoption here is pure software, which also means there is no hardware cost slowing it down.
What the evidence does and does not show
There is no direct head-to-head test of AI against survey researchers in our evidence yet. Our quality parity grade for this job is D, and a D grade means not measured, so we publish no parity number for it. We would rather say that plainly than imply a comparison nobody has run.
What would settle it is specific. One test: AI-designed questionnaires and AI-set weights run against a known benchmark, such as an administrative total or an election result, next to the same survey built by experienced researchers. Another: give models and researchers the same raw dataset, then have independent reviewers judge the writeups blind, without knowing which is which. Until something like that is published, the honest answer is uncertainty, not confidence. How we grade evidence, and why a missing test never becomes a score, is set out in our scoring methodology.
The headline figure here, Still needs a human, reads 63 out of 100 (higher is safer).
When the balance could shift
Most likely between 2037 and 2048 (8 in 10 of our scenarios). The replacement year method explains what that window is measuring and how wide the uncertainty is.
Two things could pull it earlier. The first is cost: the spend comparison above puts annual tool spend well under a salaried researcher, and with no equipment to buy, a small team can test a swap cheaply. The second is budget pressure. With employment projected to shrink (BLS, 2025), the routine coding and tabulation layer is the first place a research shop trims, and that work is already the most automatable part of the job.
Two things hold it back. Response rates keep falling, which makes weighting and nonresponse decisions harder, not easier, and those are the judgment calls a model cannot be held to. And published survey numbers carry accountability: clients, journals and regulators want a named person who stands behind the method, the consent process and the handling of respondent data.
What to do: get fluent at auditing model output, because checking an AI-coded dataset for miscategorized answers is becoming part of the job rather than a threat to it.
How to stay needed in survey research
Lean into the three tasks that stay human. Own the design conversation with clients, so you are the person defining what the study can and cannot answer. Own the credibility call on weak samples and low response, and write it down in plain language. Own the presentation, including the hostile questions after it.
Two skills compound. One is applied statistics you can explain without jargon, especially weighting and margin of error. The other is prompt-and-verify practice: using models for drafting and coding, then documenting the checks you ran. Both make you the reviewer rather than the output.
If you are weighing adjacent moves, the closest work sits in the same family: Sociologists, Social Science Research Assistants and Market Research Analysts and Marketing Specialists, which share the design-and-interpret core but attach to different employers. You can also browse the wider social scientists job family, see how research roles sit inside professional services, or check the list of jobs expected to shrink before you commit. To weigh two options directly, put them side by side in compare jobs.