EM Developer AI Usage Dashboard

Step 4 canonical deliverable · 2026-06-23 · detail-layered from 03-Narrative-Wireframe + full 6/22 transcript (the CEO, Trent Mitchell, the Head of Product, the Dev Lead; a team member relayed by the CEO). Primary audience: engineering manager. Dashboard paradigm — interactive, in-app, progressive disclosure. [Red brackets] = unknown data or decision pending; gaps trace to 02-Data-Relevance.

Feedback sources — hover for verbatim quote

CEO-n the CEO, 6/22 transcript

JK-n the Head of Product, 6/22 transcript

DL-n the Dev Lead, 6/22 transcript

MK-n a team member (relayed by the CEO)

DR Data-relevance finding (02)

Decisions & status

TR-n Trent design decision

P2 Out of scope — other surface or deferred

FLAG Open question or data gap

Review annotations — internal only

✦ manager takeaway what the EM should leave with

⚠ weak takeaway fails the test: would the EM care?

Built for the engineering manager SINGLE PRIMARY USER

Every panel answers: how well is my team leveraging AI, and where should I intervene? the CEO explicitly split the EM dashboard into two: a workflow and work-volume surface (EM Workflow Dashboard, Step 4 done) and this one — AI-specific depth for the line manager. CEO-6 CEO-1

the CEO: “The dashboard should be fairly streamlined.” CEO-5 the Dev Lead: “super detail can be overwhelming.” DL-3 This dashboard lives inside a Work Area tab CEO-7 and can be disabled per work-area template JK1. Developers don’t see this — their drill-down target is the Developer Profile. DL-2

Built for the engineering manager

Every panel answers: how well is my team leveraging AI, and where should I intervene? This dashboard lives inside a Work Area tab and can be disabled per work-area template. Individual developer detail lives on the Developer Profile (linked from every name).

Surface boundaries (reconciliation with parallel deliverables):
EM Workflow Dashboard (Step 4 done): owns activity patterns, work volume (commits, branches, active time), calendar heat map, busiest times, stale work. This dashboard does NOT replicate those. Overlap allowed: AI adoption rate (summary KPI on Workflow; funnel here), Team AI Spend (summary KPI on Workflow; per-tool breakdown here), VLOOKU trend (one-line on Workflow; per-platform/model here). CEO-6
AI Adoption Report (not yet built, exec audience): owns acceleration factor, capacity realization, org-wide adoption funnel, work-area breakdown table. This dashboard does NOT show acceleration factor or capacity realized. CEO-3 CEO-4
AI Tooling Dashboard (sources only, not yet built): will own tool comparison, tool demographics, model-level breakdown. This dashboard shows per-tool cost/sessions as context but defers detailed tool comparison. CEO-9

the Dev Lead’s Team — Backend Services Work Area · 8 engineers TR1

This week (Jun 16–22) | Month | Custom Data as of Jun 22, 11:40 PM TR2 MK1

1AI Value SummaryCEO-1 CEO-2 MK1

Four hero KPIs orient the manager, each with a sparkline and prior-period delta (MK-1: trends over snapshots). The adoption funnel below shows depth of AI usage across the team. CEO-15

✦ Manager takeaway · internalIn one glance: team AI output is growing, spend is managed, and I know how deeply my team has adopted AI.

Team VLOOKU

342

+12% vs 305 prior week

spendAllocation.valueUnitsPerDollar, team-scope (Partial/estimated). DR1 CEO-11

VLOOKU per Dollar

$0.42

+8% vs prior

Denominator = engineer salary + AI token cost. CEO-11

AI-Code Share

72%

+3 pp vs prior

contributionLineage.contributionMix.aiAssistedShare (Complete at 99.7% company scope; team-scope filtered). Collapsed AI share (2-way). DR2

AI Spend

$4,180

-5% vs prior

rollup_subject_tool_period.session_cost_usd_sum, summed across subjects and tools (Partial/estimated). DR3 CEO-8

AI Adoption Funnel DR4 TR3

87.5%

Adopted (7/8)

→

75%

Daily active (6/8)

→

50%

Power user (4/8)

workActivity.aiUsage: adoption 75%, daily 68.8%, power 62.5% at company scope. Team-scoped here. Funnel is monotonically decreasing by definition. JK2

Data basis: Hero KPIs from rollup_subject_period and spendAllocation aggregated across Work Area subjects. VLOOKU per Dollar = Partial/estimated; AI-Code Share = Complete (2-way); AI Spend = Partial/estimated. Funnel from workActivity.aiUsage (adoption/daily/power rates, all Available or Partial). DR1

2Platform / Tool BreakdownCEO-9 CEO-12 CEO-14

the CEO: “VLOOKU per tool… the developer breakdown underneath.” CEO-12 the CEO: “probably color code each platform so they’re consistent across the charts.” CEO-9 the CEO: “value contribution by work model… agent intake versus AI assisted and non AI.” CEO-14

✦ Manager takeaway · internalI know which AI tools my team relies on, what each costs, and which ones are delivering the most efficient results.

AI Spend by Platform — 6-Week Trend TR4

Claude Code GitHub Copilot Codex CLI Codex CLI RS

Chart shows cost by platform (available) rather than VLOOKU by platform (G1: not served). Cost is the honest lens; outcome attribution remains aggregate. TR4

This Week by Platform

Claude Code

$2,170 (52%)

Codex CLI

$1,050 (25%)

Copilot

$750 (18%)

Codex RS

$210 (5%)

Sum: $4,180 = AI Spend KPI above. Bars + stacked area share a consistent color system per CEO-9.

Sessions by Platform

Claude Code892

Copilot614

Codex CLI410

Codex RS102

Total: 2,018 sessions

Tool Efficiency Comparison AGGREGATE / ESTIMATED G1

Outcome efficiency by platform — outcome attribution is aggregate; per-tool VLOOKU pending [G1]

Claude Code

0.48

+6% vs prior

Copilot

0.39

— flat

Codex CLI

0.44

+3% vs prior

Codex RS

0.36

+11% vs prior

Per-tool VLOOKU breakdown will replace this panel when G1 is resolved. spendAllocation.toolEfficiency returns 4 rows (codex_cli_rs, github-copilot, codex-cli, claude-code); “outcome attribution remains aggregate.” DR5

Data basis: Cost and sessions from rollup_subject_tool_period (tool_key × subject × period). Available, no estimation needed. Tool efficiency from spendAllocation.toolEfficiency (4 tool rows, Partial — aggregate outcome attribution). Per-tool VLOOKU requires G1 (fact_contribution_session × tool aggregation). DR5

3Work-Model BreakdownCEO-2 JK2

the CEO: “human only, human w/AI and agentic and let’s just do that.” CEO-2 the Head of Product: “as long as we define how we are saying it and people understand it then the report can make sense.” JK2 the Dev Lead: Copilot-in-IDE vs agentic CLI confusion underscores the need for clear definitions. DL-4

✦ Manager takeaway · internalMy team is steadily shifting toward agentic work — 57% this week, up from 42% six weeks ago. The Human Only share is shrinking, which is what I want to see.

Nomenclature (decided 6/22): Human Only IDE with no AI · Human w/ AI co-pilot in editor · Agentic supervised CLI + cloud/autonomous CEO-2

Contribution Mix — 6-Week Trend G5 TR5

% of retained output by work model [3-way requires G5; V1 shows 2-way]

May 12

May 19

May 26

Jun 2

Jun 9

Jun 16

Agentic Human w/ AI Human Only

Bars sum to 100% each week. Agentic share rose 15pp over 6 weeks; Human Only fell from 18% to 8%. Each bar = % of retained output attributed to that work model.

This Week

Agentic57%

Human w/ AI35%

Human Only8%

IDE sessions614

Agent sessions1,404

Total: 2,018 (matches §2 session sum)

IDE vs agent session split: workActivity.sessions.ideSessions / .agentSessions — both Available. DR6

Data basis: Contribution mix from contributionLineage.contributionMix (human/AI 2-way: Complete; 3-way: G5 pending). IDE vs agent from workActivity.sessions (Available). Period rollups from rollup_subject_period. Donut percentages sum to 100%. DR6

4AI Efficiency SignalsJK3 JK4 TR6

the Head of Product: “if they’re doing agentic are they having to redo a lot of that work meaning they need help with their prompts.” JK3 the Head of Product: “if I’m not doing good prompts and I could be using a lot more tokens than I should be using and that affects my budget.” JK4 Framed as “AI efficiency signals” not “prompt quality” — no true prompt-quality or rework metric exists. TR6

✦ Manager takeaway · internalI can spot developers who might be struggling with AI — Marcus generates a lot of prompts but his code survival is low. I should check in on his workflow.

AI Code Survival Rate G2

% of AI-generated code retained (higher is better) [rate pending G2 approval]

90%

88%

86%

85%

84%

78%

72%

Team avg: 83.3% · Names link to developer profile DL-2

abandoned_ai_sum / (retained_ai_sum + abandoned_ai_sum) per developer from fact_commit components. Mart components exist; rate intentionally not served. DR7

Prompt Volume & Iteration Depth G6

High prompts + low survival = possible prompt-quality issue JK4

Dots sized by session count. Marcus: high prompts, low survival — investigate. David: low prompts, low survival — different issue (limited AI adoption). Dashed line = team trend.

Prompt count from workActivity.aiUsage.promptCount (Complete, 43,729 company). No true prompt-quality metric exists — G6. This is a proxy only. DR8

Waste Cost (Abandoned AI) ESTIMATED DR9

$680

-8% vs prior · 16.3% of total AI spend

Cost per Outcome DR10

$2.45

-11% vs prior (lower is better)

Waste = spendAllocation.wasteCost ($123K company, Partial/estimated). Cost per outcome = spendAllocation.costPerOutcome ($2.75 company). Both team-scoped. Waste% = $680 / $4,180 = 16.3%.

⚠ Weak takeaway · internalWhy weak: Prompt volume and survival rate are proxies, not true rework or prompt-quality measures. The scatter suggests correlation but no causal signal exists in the data. Fix: Keep the visualization but badge it “efficiency proxy.” Do not label it “prompt quality.” The honest framing is “signals worth investigating,” not “diagnosis.” G6

Data basis: AI code survival from fact_commit.abandoned_ai_sum / retained_ai_sum (mart components Available per-subject; rate serving needs G2 approval). Prompt volume from workActivity.aiUsage.promptCount (Complete). Waste cost from spendAllocation.wasteCost (Partial/estimated). Cost per outcome from spendAllocation.costPerOutcome (Partial). DR7

5Spend & Cohort ViewCEO-8 CEO-10 CEO-13 MK2

the CEO: “knowing the AI spend going up or down… how many people are using 80% of their budget and that’s a line over time.” CEO-8 the CEO: “we could map budgets in platforms like Anthropic… regular user, premium user, max user… showing a breakdown of budgets.” CEO-10 a team member (relayed): “if bottom 10% were using 5% of the code, totally fine. Less than 1%, I probably want to fire all of them.” MK2

✦ Manager takeaway · internalI can see that higher AI spend does correlate with more output on my team — top-quartile developers produce 3.3× the VLOOKU of the bottom. That’s ammunition for broader AI allocation.

AI Spend by Developer — This Week G3 G4

Ranked by total token spend. Stacked by platform. Budget tier labels [require G3 + G4]

$820

$710

$680

$590

$520

$480

$280

$100

Sum: $820+710+680+590+520+480+280+100 = $4,180 (reconciles with §1 AI Spend KPI). Stacked by platform · hover for per-tool detail · names link to developer profile.

Budget-tier framing fully data-blocked (G3 + G4). Using spend-based dynamic quartiles instead. TR7

Spend-Based Cohorts DYNAMIC QUARTILES TR8

Developers grouped by AI spend quartile. [Budget tier labels deferred — G3]

Top Quartile

avg VLOOKU

the Dev Lead, Sarah
$765/wk avg spend

2nd Quartile

avg VLOOKU

Marcus, Priya
$635/wk avg spend

3rd Quartile

avg VLOOKU

Elena, James
$500/wk avg spend

Bottom Quartile

avg VLOOKU

David, Alex
$190/wk avg spend

Cohort VLOOKU: (60+50+43+18)/4 = 42.75 avg across quartiles. Team total 342 / 8 devs = 42.75 avg per dev (reconciles). Top quartile 3.3× bottom quartile output. MK2 CEO-13

a team member threshold test: bottom quartile produces 18/342 = 5.3% of team output. Above MK-2’s “less than 1%” trigger. David + Alex together: 36 VLOOKU = 10.5% of team output — low but not the fire-them threshold. MK2

Data basis: Per-developer AI spend from rollup_subject_tool_period.session_cost_usd_sum (Partial/estimated). Cohort VLOOKU from rollup_subject_period (Partial/estimated). Budget facts and platform tier absent (G3, G4). Dynamic quartiles computed from spend rank, not configured tiers. DR11

6Developer-Level SignalsDL-2 DL-3 CEO-15

the Dev Lead: “I can drill down on users on developers.” DL-2 the Dev Lead: “super detail can be overwhelming sometimes.” DL-3 Kept streamlined: one overview table with signal flags. Developer name links to full profile. TR9

✦ Manager takeaway · internalI know exactly which developers to talk to: David needs help with his AI workflow (low survival suggests prompt issues), and Alex hasn’t adopted AI tools at all — time for a 1:1.

Developer	Adoption	VLOOKU	AI Spend	Survival [G2]	Signal
the Dev Lead	Power User	62 +15%	$820	88%	—
Sarah Chen	Power User	58 +10%	$710	86%	—
Priya Patel	Power User	52 +8%	$590	90%	—
Marcus Johnson	Daily Active	48 +4%	$680	78%	Low survival
James Wilson	Daily Active	44 —	$480	84%	—
Elena Rodriguez	Power User	42 +6%	$520	85%	—
David Kim	Adopted	22 -3%	$280	72%	Low output + survival
Alex Thompson	Not adopted	14 -5%	$100	—	No AI adoption

VLOOKU column sums: 62+58+52+48+44+42+22+14 = 342 (matches §1 Team VLOOKU). AI Spend column sums: $4,180 (matches §1 + §5). Adoption levels drawn from funnel: 4 Power Users, 2 Daily Active, 1 Adopted, 1 Not Adopted → 87.5% adopted, 75% daily, 50% power (matches §1 funnel).

Signal logic: > 1 SD below team mean on survival rate, or bottom-quartile VLOOKU + top-quartile spend, or not adopted. Threshold needs EM validation — see open call #5. TR9

Data basis: Developer detail from rollup_subject_period (VLOOKU, active status) and rollup_subject_tool_period (AI spend). Adoption tier from workActivity.aiUsage funnel stages. Work-model bars from contributionLineage.contributionMix per-subject (2-way). Subject dimension: dim_subject (SCD-versioned). All names link to /design/dashboards/trace-developer. DR12

Not in scope / Deferred: Individual PR-level drill-down (lives on developer profile) · True prompt-quality analysis (privacy-bounded; no path with current capture — G6) · Budget-tier cohorts (requires platform tier mapping G3 + budget facts G4) · Cross-session rework detection (no mechanism — G7) · Model-level breakdown within tools (model identity inconsistently captured) · Acceleration factor (belongs on AI Adoption Report CEO-3) · Capacity realized framing (report-level concern CEO-4 JK5) · Activity patterns / work volume / heat map (belongs on EM Workflow Dashboard CEO-6) · Organizational breakdown by work areas (belongs on AI Adoption Report MK3) · Tool demographics deep-dive (belongs on AI Tooling Dashboard CEO-9)

Coverage Cross-Check

Item	Feedback	Where answered	Status
CEO-1	Split: AI adoption report for execs, team-level AI depth for EMs	Entire dashboard; boundary box	Covered
CEO-2	Nomenclature: Human Only / Human w/ AI / Agentic	§3 nomenclature bar + all work-model charts	Covered
CEO-3	Acceleration factor framing	Deferred — belongs on AI Adoption Report	P2
CEO-4	“Capacity realized” not “savings”	Deferred — report-level concern, not this dashboard	P2
CEO-5	Dashboards “fairly streamlined”	6 sections with progressive disclosure; no embedded prose	Covered
CEO-6	Split EM dashboard: workflow vs AI usage	Charter of this dashboard; boundary box	Covered
CEO-7	Work areas have templates; dashboards on/off	Dashboard chrome + audience note	Covered
CEO-8	Budget utilization progression over time	§5 per-developer spend + cohorts; §1 AI Spend KPI sparkline	Covered (spend proxy; budget facts blocked G3+G4)
CEO-9	Tooling demographics; color code platforms	§2 consistent color system across all charts	Covered
CEO-10	Budget cohorts by platform tier	§5 spend-based quartiles (tier labels deferred G3)	Partial — G3
CEO-11	VLOOKU per dollar math	§1 VLOOKU per Dollar KPI	Covered
CEO-12	VLOOKU per tool, developer breakdown	§2 tool efficiency (aggregate proxy, G1 pending); §6 developer table	Partial — G1
CEO-13	Top 10% cohort analysis	§5 spend-based cohorts (quartile granularity)	Covered
CEO-14	Value contribution by work model	§3 contribution mix trend + donut	Covered
CEO-15	Detailed EM page for nuanced metrics to raise VLOOKU	Entire dashboard — this IS that page	Covered
JK1	Can dashboards be turned off if not using AI?	Audience note: disableable per work-area template	Covered
JK2	Nomenclature must be defined upfront	§3 nomenclature bar at section top	Covered
JK3	Prompt quality / rework patterns	§4 AI Efficiency Signals (proxy: survival + prompt volume)	Partial — proxies only (G2, G6)
JK4	Poor prompts → token waste	§4 scatter plot + waste cost callout	Partial — proxy (G6)
JK5	Capacity framing “3 additional developers”	Deferred — report-level framing, not this dashboard	P2
DL-1	Time on AI vs human coding, PRs, branches	§3 work-model breakdown (AI vs human); PRs/branches → EM Workflow Dashboard	Covered (AI/human split); P2 (PRs/branches)
DL-2	Drill down on developers	§6 developer table, all names link to Developer Profile	Covered
DL-3	Super detail can be overwhelming	6-section streamlined design with progressive disclosure	Covered
DL-4	Copilot-in-IDE vs agentic confusion	§3 nomenclature definitions	Covered
DL-5	Acceleration as key investment metric	Deferred — belongs on AI Adoption Report	P2
MK1	Progression / trend lines most important	Every KPI sparkline; §2 6-week stacked area; §3 6-week bars	Covered
MK2	Bottom 10% cohort threshold	§5 cohort cards + MK-2 threshold test in note	Covered
MK3	Organizational breakdown by work areas	Deferred — belongs on AI Adoption Report (exec audience)	P2

All 28 transcript items addressed. 20 covered, 4 partially served (data gaps), 4 correctly scoped to other surfaces (P2). No items dropped.

Design Decisions (TR Log)

ID	Decision	Rationale
TR1	Dedicated surface, scoped to Work Area	the CEO split the EM dashboard in two (CEO-6). This is the AI-depth half. Dashboard inherits team scope from Work Area context.
TR2	Default weekly + “data as of” timestamp	MK-1: trends over snapshots. Trust: data-basis visible at all times.
TR3	Adoption funnel added as sub-KPI	Not in original objectives but directly serves OBJ-1 and is data-ready (DR-4).
TR4	Platform breakdown shows cost, not VLOOKU	Per-tool VLOOKU is G1 (not served). Cost by platform is available and honest.
TR5	3-way work model is target; V1 falls back to 2-way	3-way requires G5 (data population). 2-way (human vs all-AI) works today.
TR6	Section framed “AI efficiency signals” not “prompt quality”	No true prompt-quality or rework metric exists (G2, G6). Honest proxy framing.
TR7	Spend-based quartiles replace budget tiers	Budget tiers fully blocked (G3 + G4). Spend quartiles answer a related question with available data.
TR8	Dynamic cohort cards	Same as TR7 — cohorts from spend rank, not configured tiers.
TR9	Signal logic for developer table	> 1 SD below team mean on survival, bottom-quartile VLOOKU + top-quartile spend, or not adopted. Needs EM validation (open call #5).
TR10	Scope boundary enforced	the CEO’s two-dashboard split (CEO-6). Documented in boundary box and coverage table.

Open QuestionsFLAG

G1 — Per-tool VLOOKU aggregation. Can engineering prioritize a GraphQL aggregation over fact_contribution_session × tool? Without it, the platform breakdown (§2) is cost-only; tool efficiency uses aggregate outcome attribution. G1
G2 — AI retention/abandonment rate approval. Mart components exist (abandoned_ai_sum, retained_ai_sum). Will product approve serving the rate? the Head of Product’s rework ask depends on it (§4). JK3
G3 + G4 — Budget tier mapping & budget facts. Is there a plan to capture platform tier/SKU or per-developer budget/quota data? If not, spend-quartile cohorts (§5) are the permanent alternative. the CEO’s vision: “regular user, premium user, max user.” CEO-10
G5 — Session-aware contribution population. When will work_branch_session_scores populate so the 3-way work-model split (§3) activates? Mart and GraphQL are structurally ready.
Outlier thresholds. What defines “low survival” or “high spend / low output” for §6 signal column? Proposal: > 1 SD below team mean on survival, or bottom-quartile VLOOKU + top-quartile spend. Needs EM validation. TR9
~~Companion surface boundaries.~~ RESOLVED — Trent, 6/23. Boundary box added: this dashboard does NOT show activity patterns (EM Workflow Dashboard), acceleration factor (AI Adoption Report), or detailed tool comparison (AI Tooling Dashboard). Overlap limited to summary KPIs. TR10
Alex Thompson — “Not adopted” signal. Alex shows $100 AI spend but “Not adopted” tier. Decision: is $100/wk with 92% human-only work “adopted” or not? The funnel says adopted = at least one AI session per period. Alex’s 8% AI share may be incidental IDE autocomplete, not intentional adoption. Needs definition clarification.