Methodology · The Execution Paradox · Q3 2026

N=603 responses were collected; one was a junk response, so all results are computed over the 602 analysable ones, which were collected between 2026-01-18 and 2026-03-07. Recruitment was targeted rather than open: 517 respondents (86%) came through the Prolific participant panel against a screen described in full below, 33 through launch partners, 14 through an organization and 3 through other channels, with 35 carrying no recorded source. The sample is side-builders, founders, freelancers, business owners and employees (30.6% / 20.3% / 17.6% / 15.1% / 12.1% respectively, plus 4.2% on early startup teams and one respondent who described themselves differently). Data-quality checks flagged 4 of 602 possible straightliners (0.7%, within-respondent SD below 0.5 across 11 numeric items); they were retained, and removing them moves no published mean by more than 0.004. The burnout composite (8 items, each normalized to 0 to 1 and the average rescaled to 0 to 5) showed acceptable internal consistency, Cronbach's α = 0.797. Bootstrap 95% percentile intervals (2,000 resamples) were used for means and Wilson score intervals for proportions. Analysis is complete-case per metric with no imputation. Results describe this specific screened sample, not the general working population.

Sample

The panel recruitment was screened, not open. It applies to the 517 respondents recruited through the participant panel, who had to meet every one of these criteria:

• fluent in English
• in full-time employment
• within an age band (23 to 45, and 25 to 43 on some batches)
• holding a job position, with a stated number of years of work experience
• using AI at work weekly or more often
• doing at least a quarter of their work on a computer
• a stated level of education
• a stated employer or organization type
• currently running something entrepreneurial
• a panel approval rate of 99 to 100
• not a respondent to any earlier batch, so nobody answered twice

The remaining 85 respondents came through Flowdealer's own channels and were not screened on these criteria.

Cohort cross-tabs

Instruments

Direction is the scale's direction, not its coding. It says which end of the answer is the better outcome, and nothing about reverse-scoring: the burnout composite reverses productive best, quality of life and plan completion when it combines them, and all three are listed here as higher is better, because they are.

The pain screener above was collected at intake to size an audience, and no figure in this edition uses it. The analysis runs on the burnout composite and the profiles it assigns.

These bands are failure types read off the profile assignment, not rungs on a severity ladder: they name how execution breaks, not how badly. There is no score threshold behind them. The severity ladder is in section 2 of the report.

Analysis

The flagged respondents were kept. The table above is why the count is worth stating and not worth arguing about: on the pipeline's own convention it is four people, on the other common convention it is one, and dropping all four moves no published mean by more than 0.004.

Specific LLM extraction prompts, confidence thresholds, and profile-assignment cutoffs are proprietary methodology and not publicly released. Profiles are assigned by rule, not clustering. The aggregate data behind every figure in this report is published under CC-BY-NC 4.0, and the files are linked above.

• Aggregate data (CSV)
• Aggregate data (JSON)
• Codebook (JSON)
• Data license

Limitations

This is the one known-groups check we ran, and it did not discriminate. If the burnout composite measured strain from juggling roles, people carrying four or more roles should score higher than people carrying fewer; they score 0.05 higher on a 0-to-5 scale, and the interval on that difference straddles zero. Read it as a null result, not as validation. It is also a weak test of the composite: section 10 of the report publishes role count as a non-predictor of almost everything, so a null here is what that finding predicts. A contrast that should separate is the better test, and this edition does not publish one.

This edition has limits, and they are worth stating plainly. The sample is a convenience sample of high-agency working people, N=602, and it does not represent the general working population. Most of it was recruited through a paid participant panel against the screen set out above, so the people here are, by construction, in full-time work, inside an age band, doing most of their work on a computer, already using AI at work weekly, and running something of their own. Every figure should be read as describing people like that. The design is cross-sectional: every figure in the report is an association measured at a single point in time, and nothing here establishes cause. The instrument collected no physiological, sleep, time-tracking or buffer-time measures, and it followed nobody over time, so any claim that would need those is attributed external research or it is not made. Neurodivergence is one self-report item asking how much ADHD, autism, dyslexia or similar impacts daily work, answered on a four-point scale from not applicable to significantly. It measures impact, not diagnosis.

Three groups are small enough to name. The burnout taxonomy defines nine profiles; the ninth, Quiet Struggle, matched a single respondent and is excluded from every published figure. Productive but Paying holds 15 respondents, so its numbers are directional rather than precise. So is the planning inversion: only 41 respondents never plan at all, and the gap between them and the inconsistent planners does not reach statistical significance.

Two definitions carry more weight than their wording suggests. The severity groups in section 2 are cumulative rather than a partition, and each one is a threshold rather than a label. In crisis means a burnout composite of 3.0 or above. In distress means severe on any one of three measures: a burnout composite of 3.0 or above, a quality of life of 1 or below, or daily overwhelm. Under strain means a quality of life of 2 or less on the 0 to 5 scale, or overwhelm almost daily. Each group contains the one before it, so the shares do not sum. Thriving is the apex profile, outside crisis and distress, though two of its 37 respondents fall under strain. The pain screener in the segment-definitions table above was collected at intake and is not used in this edition’s analysis, which runs on the burnout composite and the profiles it assigns.

Profile detail