fromeach

The Metrics

What an engagement looks like measured. A redacted view of a working KPI dashboard, with five years of anonymized client data. The structure, the visualizations, and the numbers are real — only the client and project specifics have been removed.

Knowledge work has always been hard to measure objectively. Story points get gamed; subjective reviews carry bias; manager intuition drifts. The AI Bench instruments output continuously — per agent, per engineer, per ticket, per hour. For the first time, engineering productivity is as observable as physical work.

Two-week sprint, four engineers

Same team, same sprint, with and without a tuned AI Bench.

Without AI Bench

40
story points delivered
$750
per story point
~5
features shipped
$30K
sprint labor cost

With AI Bench

1,000
story points delivered
$31
per story point
~125
features shipped
$31K
labor plus compute

Same team, same sprint, twenty-five times the output. Cost per story point drops from $750 to about $31 — the kind of number leadership can actually do something with. The team chooses how to spend the gain: more features, or the same features in days instead of weeks.

Time to ship — 9,550 tickets, five years

Distribution of ticket resolution time before and after the AI Bench was introduced. Log-spaced bins. Median lines marked.

Histogram of ticket resolution time, comparing pre-AI and AI-era distributions across nine thousand five hundred fifty tickets over five years 0 500 1,000 1,500 3+mo 2–3mo 1–2mo 2–4w 1–2w 3–7d 1–3d 4–24h 1–4h <1h Pre-AI median: 63 days AI-era median: 6.4 hours Pre-AI AI-era

Quality mix over time

Share of total effort by ticket category, by year. Stable until the AI Bench is introduced — then bug-fix shrinks, and feature, refactor, exploration, and documentation work all grow.

Stacked area chart showing percentage share of work types over five years; the mix stays stable until the AI Bench is introduced between Year 2 and Year 3, then bug-fix shrinks while quality-oriented work grows AI Bench introduced Year 1 Year 2 Year 3 Year 4 Year 5 feature bug-fix docs refactor exploration

1 day, 1 engineer, 7 agents in the Bench

145h human estimate → 7h 29m human hands-on (plus 13h of Bench time) · 19× less human time · 39 tasks closed.

Agent Tasks Human est Hands-on AI actual Ratio Tokens (M) Commits
A2063h20m1.2h53.2×84.024
B14.0h5m16m15.0×18.231
C14.0h5m18m13.3×7.351
D1052h37m5.2h10.1×201.729
E14.0h10m47m5.1×37.048
F517h1h 35m4.4h3.9×285.9813
G11.0h12m32m1.9×20.331

Actual engagement dashboards go further: per-agent trends over time, share-of-attention breakdowns, tokens-per-hour comparisons, weekly aggregates with prior-day deltas. This is the daily headline.

Weekly trend — AI hours versus human hands-on

Daily breakdown across one work week, with weekly means shown as dashed horizontal references.

Four-line trend chart showing daily AI hours and daily human hands-on hours across one work week, with means for each series 0h 3h 6h 9h 12h Mon Tue Wed Thu Fri AI mean: 6.2h human mean: 4.5h AI hours, daily human hands-on, daily

The day's economics

What the same day would have cost in traditional engineering versus what it actually cost. Adjust the engineer rate to your own cost basis — offshore hourly, onshore annual, or anywhere in between.

Engineer rate
$150/hour
$21,750
traditional equivalent labor
$1,225
actual cost (steering + compute)
$20,525
labor cost avoided, this day
17.8×
cost ratio
Side-by-side cost comparison: traditional equivalent labor cost versus actual cost for this single day Traditional $21,750 With Bench $1,225 ~$20,525 of labor cost avoided in this single day.

Engineer time, redirected

The Bench takes the routine work, and the work that always slipped — documentation, cleanup, the refactors no one got to. Your engineers spend their day on vision, strategy, and the tough problems you hired them to solve.

~6h
engineer hours per day freed
~$900
value of that time per day
~$225K
redirected capacity, per engineer per year
~$2.25M
across a ten-engineer team, per year

The numbers above the bar chart are labor cost avoided. The numbers below it are the opportunity created on top — engineer hours redirected from writing routine code to the higher-leverage work expensive engineers were hired to do.

Compute and tokens

50M
tokens total
3.3M
average per day
6M
peak day
~$320
compute cost, total

Ticket throughput

5.3
tickets per day average
8
peak day throughput
1.5
average cycle time (hours)
100%
tickets passing review
§

This is what an engagement looks like measured. Real numbers, in your repository, available to your team and to whoever needs them.

If you want this kind of visibility over your own team's work, write us.

hello@fromeach.com

Source: data anonymized from five years of client engagements using Jira with Tempo for timesheet collection. We do not make up numbers. We put them on charts and teach your team how to collect the right data to tell the right story. No vibes, no feelings — this is what AI-first development actually looks like when you measure it.