# Where AI saves time: research record

Built: 4 November 2026. Article: `/journal/where-ai-saves-time`.

## Question

On one repeated task in sales, customer support and bookkeeping, how much of the by-hand time does an AI tool save once every result is checked and the unusable ones are redone, and how many tasks does a test need before it can say whether the tool saves time? This is a worked example. There is no outside data; every number is arithmetic on the assumptions below, and the assumptions are printed on each chart. The studies cited in the article are not charted.

## Files

- `request.json`: the dataset declaration (no series; rights as declared).
- `build.mjs`: offline rebuild. Holds the assumptions, computes every plotted row, writes `where-ai-saves-time.json`, `validation.json`, the four chart SVGs and `sha256.json`. `node build.mjs` in this directory; no network, no dependencies.
- `where-ai-saves-time.json`: every row of both charts, the assumptions and the formulas.
- `scripts/research/chart.mjs`: the chart renderer. `scripts/research/rasterize.py`: SVG to PNG (2x) and WebP with the site font.
- Tests: `scripts/research/where-ai-saves-time.test.mjs` (in `npm test`).

## Assumptions

| Task | By hand | Check of one AI result | Chart |
| --- | --- | --- | --- |
| Sales: a follow-up email after a sales call | 10 minutes | 2 minutes | 1 |
| Support: a reply to a customer question | 3 minutes | 1 minute | 1 |
| Bookkeeping: one bank line matched to its receipt and a category | 60 seconds | 45 seconds | 1 |

Chart 2 assumes a test in which four in five results came back usable, at test sizes from 10 to 100 tasks in steps of 5.

The minutes are chosen to show the arithmetic, not to describe any business or product. Replace them with your own.

## Formulas

- Time with the tool, per task: `check + (1 - usable share) * by-hand time`. Every result is checked; a result that is not usable as it comes is redone by hand at the full by-hand time.
- Share of the by-hand time saved: `usable share - check time / by-hand time`.
- Break-even usable share: `check time / by-hand time`. At this share the time with the tool equals the time by hand.
- Likely range of the true usable share after a test: the Wilson score interval at z = 1.96 (95 percent) for `usable` out of `tasks tested` (E. B. Wilson, Journal of the American Statistical Association, June 1927, https://doi.org/10.1080/01621459.1927.10502953).

## Headline numbers (from `validation.json`)

- Break-even: sales email 20% usable, support reply 33%, bank match 75%.
- Share of the by-hand time saved at 80% usable: sales email 60%, support reply 47%, bank match 5%. At 100% usable: 80%, 67% and 25%.
- Four in five usable: after 20 tasks the likely range is 58% to 92%; after 100 tasks it is 71% to 87%. The support reply's 33% break-even sits below the range at every test size from 10; the bank match's 75% break-even sits inside it at every test size up to 100.

## Limits

- A wrong result that is not caught is outside the formula. If the check is skipped, the chart is a best case, not an estimate.
- The check time is the average over all results. Checking selectively (a glance for routine results, a close look for unusual ones) lowers it; that is a change of assumption, not of formula.
- Setup, learning time and the price of the tool are not in the formula.
- The interval assumes each tested task is independent of the others and like the tasks to come.

## Reproduce

```
cd assets/research/where-ai-saves-time-2026-11-04 && node build.mjs
python3 -I ../../../scripts/research/rasterize.py *.svg
```

`sha256.json` lists the hash of every input and output after a build. Code version: the git commit that last touched this directory.
