Where the Gain Goes: the practitioner survey

Eight anonymous questions testing the forecasts in the Where the Gain Goes report. No email, no account, and results are public and live.

D-03Measurement & Evidence D-01Enterprise AI D-04Governance & Trust

The report this survey serves reads twenty-six sources and reaches a conclusion none of them measured directly. Not one of them recorded how much AI-generated code a developer actually reads before approving it, or how many organisations have written down what an agent must prove before it may act.

These eight questions exist to settle that. Each one names the claim or forecast it tests, so you can see what your answer is for.

What happens to your answers

This survey is anonymous. It asks for no email address, no name and no account, and the database has no column and no table that could hold one. What is recorded is your answers, a two-letter country code, and how you arrived at the page. No IP address, no cookie and no tracking is stored. Your network address and browser are hashed with a secret salt so one person cannot answer fifty times; the hash is kept out of the published dataset, and so is the optional free-text answer. A Cloudflare Turnstile check runs before answers are accepted, to keep automated submissions out.

Results are public and update as responses arrive, so nothing is held back for a publication. The anonymised answers are published live under CC BY 4.0, for anyone to re-analyse, and a country code appears only once at least five responses share it.

Nothing collected identifies you directly, and there is no list to leave. Please do not name people or employers in the optional answer.

How the data will be graded

The survey does not close. It is pre-registered at Tier B at best, never A: a self-selected online sample is a large survey, and the report grades large surveys at B whoever runs them. Below 400 responses, or with the sample mostly from one channel, the data stays on the results page and is not quoted in an edition.

Publishing results live has a cost worth naming: later respondents can see earlier answers before giving their own, so the sample may be anchored in a way a closed survey would not be. Responses carry a day stamp in the published dataset so anyone can test for drift.

Q01tests: adoption saturationIn the last 30 days, how often did you delegate a task to an AI coding agent?

Q02tests: review rationedOf the AI-generated code you accepted last month, how much did you read line by line?

Only one published source attempts this as a behaviour rather than an attitude, and it asks whether developers verify at all rather than how much they read [S39]. This is the question this report most wants answered.

Q03tests: review rationedDoes your organisation route code review by risk class, or attempt to review everything?

Q04tests: review rationedRoughly what share of changes merged in the last month had no human review at all?

Q05tests: delivery gap widensCompared with 12 months ago, how has the time from change written to change in production moved?

Q06tests: evidence completenessFor a typical change in your main repository, which of these are produced automatically and linked to the change?Select all that apply.

This produces an evidence-completeness distribution, a number that does not currently exist anywhere.

Q07tests: authority policyIs there a written policy tying how much an agent may do to what must be proven first?

Q08tests: the job changesOver the last 12 months, how has your own time split moved across writing code, reviewing it, specifying work, and accounting for it?

OptionalDescribe one thing an agent did in your codebase that you did not expect.

Please do not name people or employers.

No email, no name, no account. There is nowhere to enter one because there is nowhere to store one.

See results first