Generative AI benefits at work: which ones go together?

The questionnaire asks workers whether they agree, neither agree nor disagree, or disagree with statements about generative AI saving time, improving the quality of their work and enabling new tasks. Our analysis includes people who reported working for pay or profit and using generative AI at their job in the previous month. Merely seeing an AI-generated result in a search engine does not qualify as use under the questionnaire.^[2]

The Federal Reserve already reported the separate agreement figures among these users: 81% for saving time, 52% for quality and 55% for new tasks. Those figures establish that the three statements receive different levels of agreement. They do not tell us how many of the same people agree with all three, or how the other answers combine.^[3]

The wording matters. Agreement about quality is not a quality score assigned by an independent reviewer. Agreement about time does not say how many hours were saved. And new tasks can mean something different from doing familiar work faster. We retain those distinctions throughout the article instead of combining the responses into a single productivity score.

The largest of the eight agreement profiles includes all three statements: 34.0% of eligible users. The chart keeps every profile visible. A statement omitted from a row received either a neutral answer or disagreement; omission does not mean the worker experienced no benefit.

Three other profiles show why the combined view is useful. Agreement about time saving without agreement about quality or new tasks accounted for 17.0% of users. Agreement about time and new tasks, without quality agreement, accounted for 16.9%. Agreement about time and quality, without new-task agreement, accounted for 13.1%.

These describe different combinations of answers among people who all agreed about time saving. The two figures near 17% are descriptively similar; their tiny difference is not a basis for ranking one kind of AI use above another. More usefully, adding the new-task question separates people whom a time-and-quality comparison alone would place together.

A further 10.6% did not agree with any of the three statements. That group includes neutral answers as well as disagreement. Calling it a group that received no value from AI would go beyond the questions. People can use a tool without endorsing these particular claims, and the survey does not explain their reasons.

The small quality-and-new-tasks profile without time agreement contains 20 respondents. It remains in the complete display, but it should not become a counterintuitive headline. Keeping a category visible is different from building a strong interpretation around it. The figures below preserve more of the response detail.

Neutral answers account for much of the difference between agreement about time and agreement about quality. The following table retains the three original response choices. Every percentage uses all 1,850 eligible users, so a row does not add to 100%.

An estimated 47.1% of all eligible users agreed about both time saving and quality. Another 26.9% agreed about time but were neutral about quality, while 7.0% agreed about time and disagreed about quality. These are separate responses with different meanings.

If we look only at people who agreed that AI saved time, the denominator changes to 1,489 respondents. Within that group, 58.2% agreed about quality, 33.2% were neutral and 8.6% disagreed. These are the same underlying answers expressed within the time-saving group, not another sample or a later measurement.

A neutral answer does not tell us that a person is uncertain, that quality stayed unchanged or that the person lacks experience. The response is simply neither agreement nor disagreement. Likewise, disagreeing that AI improves quality does not establish that it makes work worse.

For the bicycle shop, this distinction is useful when interpreting its own feedback. If a colleague says drafting feels quicker but does not endorse the quality of the final explanation, the owner has a clear next step: review the explanation against the technician’s notes. That turns a broad impression into a concrete discussion about the work the customer will see.

New-task answers add information even after we focus on people who agree about time saving. In this next table, all percentages use the 1,489 people who agreed that AI saved time. The nine cells describe how their quality and new-task responses combine.

Within this group, 42.0% agreed with both other statements. These same 617 respondents make up the 34.0% all-three profile among all 1,850 eligible users. The two percentages differ because one denominator includes every eligible user and the other includes only the 1,489 who agreed about time saving.

Another 17.8% of people who agreed that AI saved time agreed about new tasks while remaining neutral about quality. A different 13.4% agreed about quality while remaining neutral about new tasks. A further 11.9% were neutral on both. Each of these cells is supported by more than 180 respondents; all nine cells in this conditional table have at least 40.

This is more specific than saying AI has several possible benefits. Some respondents endorse broader task capability alongside time saving without endorsing better quality. Others endorse quality alongside time saving without endorsing new tasks. Their answers should not be turned into stages of adoption or a story about how skills develop. We observe combinations, not a progression.

A business can use that distinction to decide what evidence it needs. A tool adopted to speed up a familiar task should be judged against that task. A proposed new use needs its own review of whether the resulting work is useful and acceptable. Neither decision requires promising that every kind of benefit will arrive together.

Nor does a useful application need to deliver every benefit. If the aim is to make an established task less time-consuming while maintaining an acceptable standard, new task capability may be irrelevant to that decision. If the aim is to offer something the business could not previously produce, speed alone would not establish success. The profiles make room for these different goals without telling a business which goal to choose.

Return to the fictional bicycle shop. The owner wants posts that explain which service a customer should book. A technician supplies the facts, and generative AI helps draft the explanation. The team can review this bounded task before making it a routine part of production. The following approach is our suggested application, not a process tested by SHED.

First, define the finished piece of work. Here it is an approved post that describes the service accurately and gives the customer a clear next step. Time spent preparing notes, checking the draft, correcting mistakes and approving the post belongs in the review. The moment a first draft appears is not the same as the moment the work is ready.

Next, examine the final explanation against the technician’s source notes. Does it distinguish the service options correctly? Does it add a price, promise or technical claim that the notes do not support? Can a reader tell what to do next? These checks give the team evidence about this post. They do not require treating a worker’s general satisfaction with AI as a quality assessment.

For new capability, the owner might consider a separate question-and-answer outline that turns recurring customer queries into a useful resource. The relevant question is whether the team can now produce something useful that it could not previously produce, rather than whether it generated more variations of an already routine post. A new format still needs factual and editorial review.

The eventual decision could be narrower than “use AI more.” If the shop finds a practical advantage in drafting service explanations, it could retain that use while separately reviewing the proposed question-and-answer resource. If review and correction consume the apparent time saving, that would be a reason to reconsider this workflow. The decision rests on what happens in the shop’s own task, including the work needed to make a draft ready to publish.

A creator can apply the same distinction to a video outline, a sponsor brief or a subscriber resource. Start with a specific piece of work and an existing standard for accepting it. The national profiles do not set a target for the creator’s own results, and this survey does not identify a separate creator or freelancer population.

We used the public 2025 Survey of Household Economics and Decisionmaking, conducted online for the Federal Reserve from October 17 to 28, 2025. Its target population is civilian, noninstitutionalized U.S. residents aged 18 or older. The full survey contains 12,934 respondents. Our analysis concerns the 1,850 who reported both working and using generative AI at work in the preceding month; all had released answers to the three benefit questions.^[1][4]

Eligibility is important to the interpretation. This is a study of people already using generative AI for work, not all workers and not a comparison with nonusers. It cannot predict the benefits a business will obtain by adopting a tool. The responses also describe the surveyed work period in 2025; translating the article does not turn them into current estimates for another country.

Every percentage uses the released cross-sectional survey weights. The tables also show unweighted respondent counts. Those counts describe the records behind an estimate, not the number of people in the country. The complete three-answer table retains all 27 combinations; the eight-profile chart combines neutral and disagreement only when identifying which statements receive agreement.

We also repeated the analysis after excluding records with an imputation flag on any of the five eligibility, use or benefit fields. This left 1,841 respondents. Across the prespecified joint tables, agreement profiles and benefit-response shares, the largest absolute change was about 0.2 percentage points. That check addresses this particular exclusion; it does not remove nonresponse, coverage or measurement error. The reduced group retains the original weights and is not a corrected national estimate.^[5]

We report descriptive estimates without confidence intervals or significance tests. We did not establish a valid survey-design variance estimate for these conditional tables. Small differences should not be treated as reliable rankings. Fifteen of the 27 complete-response cells contain fewer than 30 respondents and are marked accordingly; the full counts remain available.

The joint question, response definitions, weighting, sensitivity checks and publication criteria were fixed internally before we calculated these combinations. The published marginal figures were already known. This was not a public preregistration. A source-label clarification changed the input mapping, not the planned analysis, and is recorded in the download. Two separate calculations reproduced the aggregate results.

The source data and wording belong to the Federal Reserve’s survey; the joint calculations and interpretation are Influence IQ’s. They do not imply agency endorsement. The download provides complete aggregate results, the method and code, and instructions for obtaining the pinned public source. Respondent records and identifiers are not included.^[6]