The short answer
The Paperwork Reduction Act does not regulate asking. It regulates asking the same thing twice. The trigger in 5 CFR 1320.3(c) is "identical questions posed to... ten or more persons," and every meaningful exemption in the regulation is an exemption for questions that are not identical: single-person requests, nonstandardized oral communication, nonstandardized follow-ups, open public meetings.
That means the usual advice to public sector teams, which is to restructure research until it falls outside the Act, has a price that nobody quotes. You do not buy your way out of clearance with paperwork. You buy your way out with comparability. The exemptions are available exactly to the extent that your data cannot be pooled, and a research program optimized for exemption produces faster fieldwork, higher volume, and findings that cannot be added together.
This piece maps the four doors out of the regime, prices each one in methodological terms, and describes the architecture that gets you depth and defensibility at the same time. For the threshold question of whether you are covered at all, start with our guide to the ten-person rule.
The trigger word is "identical"
Read the definition at 5 CFR 1320.3(c) with an eye on the verb. A collection of information is the soliciting of information "by means of identical questions posed to, or identical reporting, recordkeeping, or disclosure requirements imposed on, ten or more persons."
Standardization is the regulated property. Not sensitivity, not volume, not burden, not whether the subject is controversial. If ten people get the same question, a federal review process attaches. If ten people get different questions, it does not.
This is close to an exact inversion of how research methodology grades quality. Asking every participant the same thing in the same order is the defining feature of a rigorous instrument. It is what makes responses commensurable, what allows a scale item to be averaged, what makes two waves comparable across time, and what lets you say a difference between segments is a difference in the world rather than a difference in what you asked. Under the PRA, that property is the jurisdictional hook.
The regulation is not being perverse. Standardized instruments are the ones that scale to millions of respondents and therefore the ones that generate the burden the statute exists to control. CRS, reporting OMB's own figures, records federal paperwork burden of 10.34 billion hours in FY2022 against 9.97 billion in FY2021. Bespoke conversations do not produce numbers like that. But the practical effect on a research team is that the property you would defend in a methods review is the property that triggers the process.
The four doors out, and the price of each
Section 1320.3(h) sets out categories that are generally not treated as "information." Four of them are the exits research teams actually use.
Door one: the single person. Under 1320.3(h)(6), "a request for facts or opinions addressed to a single person" is not information. One interview with one person is unregulated, full stop.
The price: n equals one, forever. This door does not widen. Ten separate single-person requests asking the same thing are ten instances of an identical question posed to ten persons, which is the definition you were trying to avoid.
Door two: nonstandardized oral communication. Under 1320.3(h)(3), facts or opinions obtained "through direct observation by an employee or agent of the sponsoring agency or through nonstandardized oral communication in connection with such direct observations" are excluded.
The price: the exclusion is tied to direct observation and to the communication being nonstandardized. The moment you write a discussion guide that every observer follows, you have manufactured the identical questions the exemption was conditioned on not having. This door stays open only while the fieldwork stays improvisational, which is to say only while you cannot audit what was asked.
Door three: nonstandardized follow-ups. Under 1320.3(h)(9), "nonstandardized follow-up questions designed to clarify responses to approved collections of information" are not information.
The price: none worth mentioning, and this is the important one. Note the word "approved." This door is not an alternative to clearance. It is a privilege that clearance buys you. Get an instrument approved and you may probe freely on top of it without re-clearing every probe.
Door four: public meetings and general solicitations. Under 1320.3(h)(8), facts or opinions obtained at or in connection with public hearings or meetings are excluded, and under 1320.3(h)(4) so are responses to general solicitations of comments published in the Federal Register, provided no respondent is required to supply information beyond self-identification.
The price: self-selection, and total loss of sample control. Whoever shows up to the meeting is your sample. You cannot screen, quota, or weight your way out of it, because imposing a screener on ten or more people is itself a collection.
The inversion: what optimizing for exemption actually does
Put the four doors together and the shape of an exemption-optimized program becomes clear. It runs one-on-one, improvises its questions, never re-uses an instrument, and takes whoever turns up to open sessions. Every one of those choices is defensible individually. Together they describe a program that has traded away the ability to aggregate.
The metrics that a research leader typically reports all move in the right direction. Fieldwork starts in weeks instead of quarters, because nothing waits on a 60-day notice. Volume rises, because there is no burden budget to spend down. Participant counts look healthy. Nothing in the compliance posture is wrong, and nothing in the raw data is fabricated.
What collapses is the denominator. Consider a routine question: does satisfaction differ between two service channels? Using a standardized scale item, this is a two-sample comparison, and detecting a moderate difference of half a standard deviation at 80% power and the conventional 5% significance level takes roughly 63 respondents per group by the normal approximation, 64 by the exact t-test. Those are ordinary numbers, reachable in a week of fielding a cleared instrument.
Now run the same question through door two. Every conversation is nonstandardized, so no two respondents answered the same item. There is no group of 63 to compare against another group of 63, because there are no groups. There are 126 individual accounts, each valid, none commensurable. You cannot reach the threshold by running more waves, because the thing that is missing is not sample size. It is a shared question.
This is the sign inversion at the heart of the regime. Under the PRA, the compliance-minimizing move and the evidence-maximizing move point in opposite directions, and the compliance-minimizing move is the one that is invisible in a status report. A program can look like it is doing more research every quarter while steadily losing the ability to answer any question that requires two numbers to be put side by side.
Two second-order effects follow, and both are worth naming to stakeholders early:
- No time series. Comparing this year to last year requires the same instrument in both years. An exemption-optimized program cannot trend anything, so it can never demonstrate that a service improvement worked.
- No segment analysis. A statement such as veterans report lower satisfaction than the general population requires identical questions across both populations, which is the trigger. The findings most likely to drive equitable service delivery are the findings most likely to need clearance.
The architecture that is actually both legal and rigorous
The way through is door three, and it is the only door that does not charge you comparability.
Clear one broad, generic instrument. Then use nonstandardized follow-ups under 1320.3(h)(9) to get the depth. That is not a loophole; it is the structure the regulation explicitly contemplates, and it is how the government's own customer experience program is built. The National Science Foundation's renewal notice of 20 August 2026, Federal Register document 2026-16984 under OMB Clearance Number 3145-0254, covers a single umbrella collection spanning "interviews, questionnaires, surveys, and focus groups" for up to 2,001,550 respondents, with outputs including "the creation of personas, customer journey maps and reports."
Three design rules make that architecture work in practice.
Scope the cleared core generically and widely. You are living with it for up to three years, because 5 CFR 1320.10(b) provides that OMB "shall not approve any collection of information for a period longer than three years." Narrow instruments require change requests; generic ones absorb new questions.
Budget burden hours for depth at submission. The burden estimate is a cap you set on yourself. NSF's notice pairs 2,001,550 respondents with 101,125 total annual burden hours, an average of just over three minutes each. Once that ratio is filed, long-form interviews compete against it for room.
Put the standardization where it earns its keep, and the improvisation where it is free. Fixed, identical, cleared items for anything you intend to count, trend, or compare. Adaptive probing for the "why" behind each answer. The first gives you the denominator; the second gives you the explanation.
Where Koji fits
This architecture is precisely what an AI-moderated interview is: a fixed instrument that every participant receives identically, with conversational probing layered on top.
A Koji study is defined by structured questions in six explicit types (open_ended, scale, single_choice, multiple_choice, ranking, and yes_no). That set is the cleared core. It is written down, reviewable by your paperwork clearance officer, and delivered to every participant the same way, which is exactly what a PRA supporting statement has to describe and what a legacy survey tool also gives you. What legacy tools do not give you is the second half.
On top of that fixed core, the AI interviewer probes conversationally, in the participant's own words, following whatever the answer opens up. Those probes are nonstandardized follow-ups clarifying responses to an approved collection, which is the 1320.3(h)(9) category. You get scale-item data you can trend and segment, plus interview-grade explanation, from a single cleared instrument.
The operational difference matters just as much under a burden cap. Because sessions are asynchronous and moderated by AI, a five-minute voice interview costs the same three minutes of respondent burden as a five-minute form while returning vastly more, and analysis is automatic rather than a moderator-hour problem. Thematic analysis and reports are generated in one click, with no moderator bias and no scheduling. Traditional platforms like Qualtrics or SurveyMonkey give you the standardized half and leave the depth to a separate, unaffordable research operation. UserTesting and dscout give you depth but not an instrument you would file. Koji is the only shape that files as one collection and delivers both.
Further reading in our documentation: structured questions in AI interviews, how AI-moderated interviews work, structured vs. unstructured interviews, semi-structured interviews, and intake forms and consent.
Frequently asked questions
What exactly triggers the Paperwork Reduction Act?
Posing identical questions to ten or more persons within a twelve-month period, on behalf of a federal agency. 5 CFR 1320.3(c) defines a collection of information by reference to "identical questions posed to, or identical reporting, recordkeeping, or disclosure requirements imposed on, ten or more persons," and 1320.3(c)(4) sets the twelve-month window. Standardization is the trigger, not sample size alone.
Can I run unstructured interviews without OMB clearance?
Often yes, under 5 CFR 1320.3(h)(3), which excludes facts or opinions obtained through direct observation or "nonstandardized oral communication in connection with such direct observations," and 1320.3(h)(6), which excludes a request addressed to a single person. The cost is that genuinely nonstandardized responses cannot be pooled, trended, or compared across segments.
What is the difference between an exemption and a generic clearance?
An exemption means the activity is not a collection of information at all, so no approval is needed and no aggregation is possible. A generic or umbrella clearance is an approval covering a broad family of collections, which lets you field many activities under one control number and then probe freely using the nonstandardized follow-up category at 1320.3(h)(9).
Do adaptive AI follow-up questions need separate clearance?
Follow-ups that clarify responses to an already-approved collection fall within 5 CFR 1320.3(h)(9), which excludes "nonstandardized follow-up questions designed to clarify responses to approved collections of information." The controlling word is "approved": the exclusion presupposes a cleared underlying instrument, so it supports adaptive probing on top of clearance rather than instead of it. Confirm scope with your agency's paperwork clearance officer.
How long does a PRA approval last?
No longer than three years. 5 CFR 1320.10(b) states that OMB "shall not approve any collection of information for a period longer than three years," and renewals go back through Federal Register notice. Emergency processing under 1320.13 yields a control number valid for a maximum of 90 days.
Does staying under ten respondents solve the problem?
Only for genuinely one-off work. The count under 1320.3(c)(4) is of persons the instrument is addressed to across any twelve-month period, not per wave, so repeated small batches of the same questions aggregate into a covered collection. It also counts invitations rather than completed responses.
Design the cleared core once, then go as deep as you like
The Paperwork Reduction Act asks you to write down the questions you intend to count and defend them in public. That is a reasonable bargain, and it is a much better bargain than the alternative most teams drift into, which is a research program that never triggers the Act and never produces a comparable number either.
Koji is built for the shape the regulation rewards: an explicit set of structured questions that every participant answers identically, plus an AI interviewer that probes conversationally on top of it. One instrument to clear, interview depth at survey scale, automatic thematic analysis, one-click reports, no moderator bias, and 10x faster insights than a traditional moderated program.
Start a free Koji study and build a core instrument worth clearing.
