Internal · CEO & CPO review · July 2026

Who delivers. Proven, not guessed.

Every employee's agent already sits on their real work — code, designs, docs, decisions, deals. People reads that work and answers leadership's blunt question: in an org of 2,014, who ships, who carries others, who's coasting?

One pipeline · every surface renders the same truth Humans decide — the system never sets a rating

The three jobs

What leadership asked for.

1

Find the top performers

Including the quiet ones — the engineer whose runbook three teams depend on.

2

Find who's coasting

Full calendars, nothing shipped. Managers who only forward work.

3

See it at org scale

Company → function → team → person in two clicks.

Why only ai.com can do this: performance tools guess from surveys and self-reports. We read the work itself — every claim verified against two independent sources.

What we measure

Five measures, read where each role works.

A designer is read in Figma, an engineer in the repo, a manager in decisions and handoffs. Same scale, different weights.

1

Output

Finished work, verified by a second source.

Goals · verified done

2

Impact

How many others used or extended it after.

Knowledge graph

3

Progress

Did work change with them — or just pass through?

Threads · docs

4

Speed

How fast things get decided and unblocked.

Approvals · calendar

5

AI Leverage

Shipped work their agent carried. High is good.

Agent activity

Top performer 88+ Strong 72–87 Steady 48–71 Coasting <48 No data always with the reason

Two sources or it doesn't score A broken connector is never a quiet person Chat volume is context, never score Teams under 5 don't roll up Everyone sees their own page first Ratings are set by humans, logged, in calibration

Guardrails

Why the rules exist.

Microsoft, 2020

Removed individual scores from Productivity Score after backlash. Hence n≥5 rollups and a self-view from day one.

EU AI Act

Worker-evaluation AI is high-risk. Hence human-signed reports, logged calibration moves, contestable evidence.

Goodhart's law

Any single number gets gamed. Hence five components visible everywhere the score appears.

Collaborative overload

Top helpers burn out. "Carrying the most" renders as load to manage, not a leaderboard.

Next: Journeys for how it runs day to day · App to use it as each role · Score for the math · Ten for the full prototype portfolio.

How it runs, from first connector to quarter close

The journeys.

Six phases. Most of the system is passive — nobody fills anything in. Click a phase.

Escalation paths

When something's wrong, this is the route.

Every escalation goes to the nearest human first, with evidence attached. Nothing auto-escalates past a manager who's already on it.

One page, four rhythms

Who touches what, how often.

Bands describe visible output; only a human sets a rating. Click a function label to zoom, any dot for the person.

Score = role-weighted visible output, 0–100. Its five components sit beside it everywhere — gaming one shows in the other four.

Each blob is a team; a node is a person, sized by how much others build on them. Cyan strands = work one team builds on in another. Dashed red = a bottleneck holding up work in other teams. Every edge is real work — a citation or handoff, never a message. Drag to spin the globe, scroll to zoom; hover to isolate someone, click for their page.

A flag is a suggested conversation, not a verdict. On-leave and no-data weeks are excluded from rollups.

Q3 cycle · Engineering · senior engineers · 12 people Flagged cards first — that's where the room argues

A band here is a human decision; the score is only evidence. Drag a card — every move requires a reason and lands in the permanent log.

Your brief · computed 07:52

Morning, June.

Two things need you today. The people section has four items — one is an instrument fault, not a person.

People · this week

People items read the same evidence as every other surface. Every action is logged; dismissals keep their reason.

The heart of the system

Five measures × your role's weights, ranked against your peers.

That is the entire formula. Below: what counts in each department, the exact rule behind every number, and a simulator to try it.

1 · IngestRaw events from connectors→ 2 · VerifyTwo sources agree = a claim→ 3 · MeasureFive measures · last 8 weeks→ 4 · CompareAgainst peers at your level→ 5 · ScoreWeighted for your role → 0–100

What counts, by department

Pick a department. Click any row for the exact rule.

"Your peers" means people doing the same job at the same level — never fewer than 20, so nobody is a group of one. New joiners get 90 days before anything counts. Chat volume, meeting hours and online time are never inputs. And when we can't see enough of someone's work, they show as "no data" — never as a zero.

Simulator

Try it. Behavior in, score out.

Runs on the department selected above. Note what the chat slider does — and doesn't.

Connected sources reportingCoverage check runs first

ComponentBehaviorVs peersWeightPoints

Score

Signals that fire

The prototype portfolio

Ten ways in. Six are live here.

The portfolio spans altitude (org → person), bluntness (scores → bands → signals) and host (new surface vs. platform-native). Rows marked ● are working views in this file — open them from App.

Recommendation

Three directions, made to argue.

Direction A

Org Map

The altitude fix — 2,014 people legible in seconds; its person drawer is the drill-down every concept reuses.

Direction B

League Table

The bluntness test. Leadership asked for explicit; this is explicit, built honestly. Forces the scores-vs-bands decision.

Direction C

Monday Brief

The platform-native proof — days of effort, rides shipping infrastructure. ai.com extending itself, not a BI bolt-on.

My Page ships regardless — it answers "can I see mine?", the first question any employee or works council asks. Work Graph is round two: highest ceiling, needs proven citation density first.