84 / AI METRICS FRAMEWORK / USAGE · QUALITY · ECONOMICS · AUTONOMY · VELOCITY
84 / METRIC / EVALS · ANALYTICS · MANAGEMENT

AI METRICS
FRAMEWORK.

AI Metrics Framework — единая система измерения AI-продукта от реального использования до бизнес-эффекта. Она связывает Usage → Quality → Economics → Autonomy → Delivery и не позволяет выдавать рост одной красивой цифры за успех всей системы.

Главная идея: “сколько вызовов LLM было сделано” почти ничего не говорит о ценности. Нужны правильные denominators, eligible population, baseline, cohort, период измерения, quality guardrails, cost of successful operation и доказательство того, что конкретная архитектурная надстройка действительно улучшает результат.
00. ARCHITECTURAL STATUS

MEASUREMENT FRAMEWORK, NOT ANOTHER RUNTIME SERVICE

Canonical plan: №84 = METRIC / DEFAULT N/A / SEPARATE COMPONENT N/A / Evals / Analytics / Management. Это рамка для №85–94 и общая карта KPI: usage, quality, economics, autonomy and velocity.
TYPEMETRICPortfolio / product measurement framework.
DEFAULTN/AMetrics are observed, not “enabled”.
USE WHENMEASURING AI VALUEProduct, process, rollout, architecture investment.
SEPARATE COMPONENTN/AInstrumentation/analytics may be components; framework is not.
LIVES INEVALS / ANALYTICS / MANAGEMENTAcross product and production system.
VALUEONE CAUSAL KPI SYSTEMAvoid vanity metrics and denominator games.
01A. ARCHITECTURE BOUNDARIES & OPERATIONS

WHAT №84 OWNS

A. BOUNDARY WITH NEIGHBORS

№47 Evals measures controlled task quality; №46 Observability records runtime behavior; №84 turns these and product/business events into an integrated management metric system. №85–93 own individual metrics. №94 owns dashboard composition. №92 Ablation Delta owns causal uplift of architecture mechanisms.

B. PREREQUISITES / CROSS-REFERENCES

Strong dependencies: №40 Eval-Driven Optimization, №46 Observability, №47 Evals, №50 Contracts, №61 Provenance, №64 Budgets, №77 Production Architecture. Metrics should resolve to stable task/run/use-case identities and quality outcomes.

C. CONTROL / DATA / RUNTIME / OFFLINE

CONTROL: metric definitions, denominators, eligibility rules, cohorts, thresholds, targets. DATA: user/task/run/outcome/cost/review/deployment events. RUNTIME: instrumentation only—do not put analytical aggregation on critical path. OFFLINE: cohorting, aggregation, causal analysis, executive reporting, eval comparison.

D. FAILURE & OPERATIONS CONTRACT

Success: every KPI has definition, numerator, denominator, eligibility, time window, segmentation, source and owner. Failure: changing denominator, mixing attempts with successful operations, counting waiting time as saved human time, averaging incompatible use cases, or optimizing one KPI while guardrail quality collapses.

E. WHAT THIS TOPIC DOES NOT OWN

№84 does not decide whether an answer is correct, does not implement tracing and does not prove causality from correlation. It owns metric architecture: what should be measured together, how metrics relate, and how to prevent misleading conclusions.

01. AI VALUE TREE

USAGE → QUALITY → ECONOMICS → AUTONOMY → DELIVERY

01 / USAGE

Is it used?

Adoption Rate, active users, habitual usage, eligible population, cohort retention.

If nobody uses the capability, downstream value cannot materialize.

02 / QUALITY

Is it useful?

Acceptance Rate, task eval success, edit depth, verifier outcomes, user correction.

Usage without quality can be novelty or forced adoption.

03 / ECONOMICS

Does it pay?

Time Savings, Cost per Operation, human review cost, failed-run cost, throughput value.

Quality alone does not guarantee positive ROI.

04 / AUTONOMY

Can it complete?

Autonomy Rate, Escalation Rate, approval dependence, eligible risk classes.

Autonomy must be quality- and risk-adjusted.

05 / DELIVERY

Can we improve fast?

AI Velocity, experiment throughput, eval lead time, release cadence, ablation evidence.

A good system also improves economically over time.

No single layer is sufficient. High adoption + low acceptance = people try but don't trust. High autonomy + low quality = dangerous automation. High quality + terrible cost = demo, not scalable product.
02. METRIC LAYERS

HEALTH, VALUE, GUARDRAIL, DIAGNOSTIC

NORTH-STAR / VALUE

Outcome

Useful completed work, verified value, time/cost saved, business throughput.

LEADING

Early signal

Adoption, coverage, activation, weekly usage, eligible task penetration.

GUARDRAIL

Do not break

Quality, safety, privacy, latency, failure/error rates, human override.

DIAGNOSTIC

Why changed?

Escalation reasons, retry rate, model route, use case, cohort, source class, reviewer load.

03. LEADING vs LAGGING

DON'T CONFUSE EARLY ACTIVITY WITH REALIZED VALUE

LayerLeadingLagging
UsageActivation, WAU, task startsHabitual adoption / retained usage
QualityVerifier pass, edit predictionAcceptance as-is, rework, business outcome
EconomicsModel cost/run, active human minutesCost per successful operation, realized time savings
AutonomyEligible automation decisionsAutonomous successful completions
DeliveryExperiments startedIdea → production time, measurable uplift shipped
Leading metrics help steer quickly. Lagging metrics tell whether value actually materialized. A framework needs both.
04. THE CORE METRIC SERIES

№85–93 AT A GLANCE

85 / ADOPTION RATE

Real use

Доля eligible users, реально использующих AI в заданный период.

active eligible users / eligible users
86 / COVERAGE

Work penetration

Доля eligible task volume / task types / process steps, где AI реально участвует.

AI-handled eligible work / eligible work
87 / TIME SAVINGS

Human time saved

Экономия именно активного человеческого времени, с учётом review/rework.

baseline human active time − post-AI human active time
88 / ACCEPTANCE RATE

Output usefulness

Как часто AI-результат принимают as-is, с minor edit, major rewrite или reject.

accepted as-is outputs / eligible AI outputs
89 / ESCALATION RATE

Human fallback

Доля AI-задач, которые требуют передачи человеку; важно разделять policy vs failure escalation.

human escalations / eligible AI tasks
90 / COST PER OPERATION

True unit economics

Models + tools + retrieval + infra + retries + failures + human review per successful completion.

total attributable cost / successful operations
91 / AUTONOMY RATE

Safe completion

Доля eligible задач, завершённых без обязательного человеческого вмешательства.

autonomous successful completions / autonomy-eligible tasks
92 / ABLATION DELTA

Architecture value

Прирост от конкретной надстройки относительно системы без неё.

metric(full system) − metric(ablated system)
93 / AI VELOCITY

Delivery speed

Время от одобренной AI-идеи/гипотезы до production with measurable outcome.

production date − approved idea date
05. WHY №94 IS SEPARATE

DASHBOARD ≠ FRAMEWORK

№84 FRAMEWORK

Definitions and causal map

What metrics mean, how denominators are designed, what dimensions/guardrails are required, how metrics should be read together.

№94 DASHBOARD

Presentation and decision surface

Which cards/charts/alerts/slices are shown to management, product, ops and engineering; diagnostic drill-down and composite views.

06. DENOMINATOR DESIGN

THE MOST DANGEROUS PART OF AI METRICS

BAD:

autonomy_rate =
  autonomous_tasks / all_tasks

Problem:
  many tasks may be legally,
  technically or intentionally
  ineligible for autonomy.

BETTER:

autonomy_rate =
  successful autonomous completions
  / autonomy-eligible tasks

AND record:

total tasks
eligible tasks
attempted autonomous tasks
successful autonomous tasks
policy-blocked tasks
human-required tasks
Denominator determines the story. Every metric definition must name the eligible population explicitly.
07. ELIGIBILITY

DEFINE WHO / WHAT COULD REASONABLY COUNT

USER ELIGIBILITY

Adoption

Users with access + relevant role/use case, not entire company headcount.

TASK ELIGIBILITY

Coverage

Tasks for which AI is allowed and technically capable, not every job in the organization.

AUTONOMY ELIGIBILITY

Risk boundary

Only tasks permitted to run without mandatory approval.

OUTPUT ELIGIBILITY

Acceptance

Comparable output types with a meaningful “accepted as-is” state.

08. BASELINE

“AFTER AI” WITHOUT “BEFORE AI” DOES NOT SHOW VALUE

HISTORICAL

Before rollout

Human process baseline under comparable workload.

CONTROL COHORT

Concurrent comparison

AI vs non-AI cohort where allocation is fair enough for comparison.

ABLATION

Architecture baseline

Same system with one mechanism removed/disabled.

No baseline → you can describe current performance, but causal claims like “AI saved 40%” are unsupported.
09. COHORTS

AVERAGES CAN HIDE BOTH SUCCESS AND FAILURE

USER COHORT

Role / maturity

New users, power users, team, region, role.

USE CASE

Task taxonomy

Writing, research, coding, classification, support, actions.

RISK CLASS

Low / medium / high

Autonomy and escalation rates must not be averaged blindly across risk classes.

TIME COHORT

Rollout wave

Pre/post version, model change, onboarding cohort, release week.

10. PERIOD

DAILY, WEEKLY AND MONTHLY ANSWER DIFFERENT QUESTIONS

WindowGood forRisk
Per-run / per-taskQuality, cost, latency, escalation reasonToo noisy for product adoption/value
DailyOperational anomaly, launch monitoringDay-of-week / low volume volatility
WeeklyWAU adoption, coverage, stable product behaviorCan hide short incidents
MonthlyBusiness/management trend, cohortsSlow feedback loop
Rolling windowSmoothing + ongoing alertsMay hide release boundary unless annotated
11. TASK TAXONOMY

THE UNIT OF VALUE MUST BE CLASSIFIED

task_type:
  summarize_document
  research_question
  draft_marketing_post
  classify_ticket
  create_code_patch
  approve_refund
  browser_action
  ...

dimensions:
  risk_class
  complexity
  expected_human_baseline
  autonomy_eligible
  review_required
  business_value_weight
Without task taxonomy, “average acceptance rate” across code patches, social posts and financial actions is analytically meaningless.
12. USER vs TASK vs RUN

DO NOT MIX UNITS

USER

Adoption unit

Who has actually adopted the system?

TASK / OPERATION

Value unit

Was the real unit of work completed successfully?

RUN / ATTEMPT

Execution unit

How many attempts/retries/model calls occurred to complete one task?

ONE TASK
  may contain:
    1 workflow instance
    4 model calls
    3 tool calls
    2 retries
    1 human review

If cost denominator = model calls,
you measure inference economics.

If denominator = successful task,
you measure product economics.

These are different metrics.
13. QUALITY AS A GUARDRAIL

EVERY VALUE KPI NEEDS QUALITY CONTEXT

TASK EVAL

Controlled quality

Representative benchmark/eval pass rate.

ACCEPTANCE

Real workflow quality

Accepted as-is/minor/major/reject distribution.

ERROR SEVERITY

Not all failures equal

Critical material error should outweigh cosmetic edit.

RISK

Quality by risk class

High-stakes autonomy cannot use same acceptable-error threshold as low-stakes drafting.

Time Savings is not positive value if humans spend less time because they fail to notice errors.
14. METRIC COUPLING

READ METRICS IN PAIRS / TRIADS

Observed patternLikely interpretation
Adoption ↑, Acceptance ↓Usage is growing faster than usefulness; forced rollout / novelty / quality regression possible.
Coverage ↑, Escalation ↑AI is touching more work but may be exceeding capability/risk boundary.
Autonomy ↑, Quality ↓Unsafe automation expansion.
Acceptance ↑, Cost/Operation ↑↑Quality improved but unit economics may be unsustainable.
Time Savings ↑, Human Review ↑Check whether “saved” time excludes review overhead incorrectly.
AI Velocity ↑, Ablation Delta ≈ 0Team ships faster but changes may not create measurable value.
Adoption stable, Coverage ↑Existing users are delegating more task volume / use cases.
15. WEIGHTED METRICS

ONE TASK ≠ ONE TASK

Raw coverage:
  100 AI-handled tasks
  / 200 eligible tasks
  = 50%

But:
  90 AI-handled tasks are trivial
  10 are high-value

Weighted coverage:

sum(task_weight × AI_handled)
--------------------------------
sum(task_weight × eligible)

Weights may reflect:
  baseline human minutes
  business value
  frequency
  strategic importance

CAUTION:
  weights must be governed,
  versioned and transparent.
Weighted metrics are useful but easier to manipulate. Always retain raw unweighted metric beside the weighted version.
16. COST ACCOUNTING

MODEL TOKENS ARE ONLY ONE LINE ITEM

cost_per_successful_operation =

  model inference
+ embedding / rerank
+ web / external tools
+ browser / compute
+ storage / vector / DB
+ queue / workflow infra
+ retries
+ failed runs
+ human review
+ manual recovery
+ amortized platform cost (if desired)
--------------------------------------
successful completed operations
Keep direct variable cost and fully-loaded cost as separate views. Otherwise teams argue because one number includes people/platform overhead and another does not.
17. TIME SAVINGS ACCOUNTING

ELAPSED TIME ≠ HUMAN ACTIVE TIME

BASELINE ACTIVE

Human work before

Typing, reading, searching, calculating, editing, coordination.

POST-AI ACTIVE

Human work after

Prompting, reviewing, correcting, approval, recovery.

WAITING

Track separately

AI processing wait can affect cycle time, but is not necessarily human labor.

time_savings =
  baseline_human_active_minutes
  − post_AI_human_active_minutes

net_value_time =
  time_savings
  − additional coordination/rework
  − manual recovery overhead
18. AUTONOMY IS A LADDER

DON'T FORCE A BINARY “AUTONOMOUS / NOT” LABEL

L0

Assist

AI suggests; human does action.

L1

Draft

AI produces output; human must review/commit.

L2

Conditional

Low-risk auto; higher-risk approval/escalation.

L3

Autonomous

Eligible operation completes without mandatory human.

L4

Closed-loop

System acts, observes outcome, recovers within bounded policy.

№91 should measure autonomy at defined eligibility/risk level—not reward bypassing required human controls.
19. ESCALATION TAXONOMY

HIGH ESCALATION CAN BE HEALTHY OR BROKEN

POLICY

Expected

High-risk task is designed to require human approval.

AMBIGUITY

Clarification

Missing material information; human/user input required.

CAPABILITY

AI limit

System cannot solve reliably with available models/tools.

FAILURE

Operational

Tool error, provider failure, invalid output, retry exhausted.

A single escalation_rate hides whether safety policy works correctly or the system simply fails too often.
20. ABLATION AS CAUSAL DISCIPLINE

“WE ADDED A CRITIC” IS NOT A VALUE CLAIM

FULL SYSTEM:
  RAG
  + reranker
  + critic
  + strong-model escalation

ABLATION:
  same tasks
  same period/model conditions
  remove critic

compare:
  acceptance
  eval pass
  latency
  cost/operation
  escalation
  human review

Ablation Delta:
  full − ablated

If quality delta ≈ 0
but cost/latency ↑,
critic may not justify itself.
№92 is how the 95-topic architecture stays honest: mechanisms are optional until eval/ablation shows they earn their complexity.
21. CAUSALITY vs CORRELATION

METRIC CHANGE AFTER RELEASE IS NOT AUTOMATICALLY CAUSED BY RELEASE

A/B / RANDOMIZED

Strongest practical evidence

Where feasible and ethical.

STAGGERED ROLLOUT

Cohort comparison

Different rollout waves provide concurrent reference.

PRE/POST

Weak to medium

Needs seasonality/workload/composition controls.

ABLATION

Architecture effect

Disable one mechanism under controlled eval/test conditions.

Framework should label evidence strength instead of presenting every trend as causal uplift.
22. CONFIDENCE & SAMPLE SIZE

20 TASKS ARE NOT A STABLE COMPANY KPI

N

Always show sample

Rate without denominator count is easy to misread.

INTERVAL

Uncertainty

Show confidence/credible interval when rates drive decisions.

MIN SAMPLE

Suppress unstable slices

Do not rank teams/use cases from tiny n without warning.

23. MISSINGNESS

“NO REVIEW” ≠ “ACCEPTED”

output_status:

accepted_as_is
minor_edit
major_edit
rejected
not_reviewed
unknown

DO NOT:
  acceptance =
  accepted / (accepted + rejected)

if 60% of outputs
were never reviewed.

Track:
  review_coverage =
  reviewed_outputs / eligible_outputs
Every label-based metric needs label coverage. Otherwise missing data can silently inflate performance.
24. METRIC OWNERSHIP

EVERY KPI NEEDS AN OWNER AND A DEFINITION CONTRACT

FieldExample
Metric IDacceptance_rate_v2
Business definitionShare of reviewed eligible outputs accepted without edits.
Numeratoraccepted_as_is outputs
Denominatorreviewed eligible outputs
Eligibilitytask_type in supported generation tasks
Windowweekly rolling 28d
Required dimensionsuse_case, cohort, model_route, risk_class
Data ownerAI Analytics
Business ownerProduct Lead
Versionv2, effective 2026-09-01
25. CANONICAL EVENT MODEL

MEASURE FROM STABLE EVENTS, NOT AD-HOC DASHBOARD QUERIES

{
  "event_id":"EVT-...",
  "event_type":"task_completed",
  "occurred_at":"...",
  "tenant_id":"...",
  "user_id":"...",
  "task_id":"TASK-...",
  "run_id":"RUN-...",
  "use_case":"research_question",
  "risk_class":"medium",
  "eligible":{
    "ai":true,
    "autonomy":true
  },
  "outcome":{
    "success":true,
    "acceptance":"minor_edit",
    "escalated":false,
    "autonomous":true
  },
  "human":{
    "active_seconds":82,
    "review_seconds":55
  },
  "cost":{
    "model":0.18,
    "tools":0.04,
    "infra":0.02,
    "human_estimated":1.10
  },
  "versions":{
    "product":"2026.09.12",
    "model_route":"router-v8",
    "workflow":"research-v4"
  }
}
Raw event is not the metric. Metric definitions aggregate these events with governed eligibility, cohort and period rules.
26. MINIMUM DATA MODEL

TASK IS THE CENTER OF VALUE MEASUREMENT

metric_task(
  task_id
  tenant_id
  user_id
  use_case
  risk_class
  started_at
  completed_at
  successful
  ai_eligible
  autonomy_eligible
  baseline_class
  product_version
)

metric_run(
  run_id
  task_id
  attempt_no
  mode
  model_route
  cost
  latency
  outcome
)

metric_review(
  review_id
  task_id
  reviewer_type
  acceptance_class
  edit_seconds
  reason_code
)

metric_human_time(
  task_id
  phase
  active_seconds
  measurement_method
)

metric_deployment(
  version
  deployed_at
  cohort
  experiment_id
)
27. IDENTITY & JOIN KEYS

WITHOUT TASK_ID, COST/QUALITY/TIME CANNOT BE CONNECTED

TASK_ID

Value unit

Stable end-to-end unit of work.

RUN_ID

Execution attempt

Links observability spans, retries and model/tool calls.

EXPERIMENT_ID

Causal slice

Links metric outcome to A/B or ablation treatment.

VERSION

Change boundary

Product/workflow/model/policy version for cohort analysis.

28. OBSERVABILITY → METRICS

TRACE DATA IS INPUT, NOT THE MANAGEMENT KPI

OBSERVABILITY

Run/span/tool/model timing, failures, route, retries.

METRIC FACTS

Task outcome, human time, review status, costs, eligibility, cohort.

FRAMEWORK

Adoption, coverage, acceptance, cost, autonomy, velocity + guardrails.

29. EVALS → METRICS

CONTROLLED QUALITY AND REAL-WORLD QUALITY ARE COMPLEMENTARY

OFFLINE EVAL

Reproducible benchmark

Detect regressions, compare models/prompts/architecture under controlled task sets.

ONLINE OUTCOME

Real workflow

Acceptance, escalation, rework, business success and cost under actual users/tasks.

A model can improve offline eval while acceptance drops because workflow/users changed—or vice versa. Dashboard should preserve both views.
30. DATA QUALITY

BAD INSTRUMENTATION CREATES PRECISE NONSENSE

COMPLETENESS

Required events present

Start, completion, review, cost and human time coverage.

UNIQUENESS

No duplicate task events

Idempotent event ingestion / stable event_id.

CONSISTENCY

State transitions valid

Rejected output cannot simultaneously be accepted_as_is.

TIMELINESS

Fresh enough

Operational monitoring and management reporting have different SLA.

31. PRIVACY & GOVERNANCE

ANALYTICS MUST NOT BECOME A SHADOW SURVEILLANCE SYSTEM

MINIMIZE

Measure necessary fields

Prefer task/use-case attributes over raw prompt/output copies.

PSEUDONYMIZE

Identity where possible

Aggregate management metrics need not expose user content.

ACCESS

Role-based analytics

Individual-level slices only when justified and authorized.

RETENTION

Bound event history

Align analytics retention with №76 governance/privacy policy.

32. GOODHART'S LAW

WHEN A MEASURE BECOMES A TARGET, PEOPLE/SYSTEMS LEARN TO GAME IT

MAXIMIZE ADOPTION
Force users to open AI even when it does not help.
ADOPTION + RETENTION + VALUE
MAXIMIZE AUTONOMY
Remove human gates to improve the number.
RISK-ADJUSTED + QUALITY
MINIMIZE ESCALATION
System hides uncertainty and guesses.
EXPECTED vs FAILURE ESCALATION
MAXIMIZE ACCEPTANCE
Only easy tasks are routed to AI.
ACCEPTANCE + COVERAGE / DIFFICULTY
MINIMIZE COST
Cheaper models create more rework and lower total throughput.
COST / SUCCESS + QUALITY
MAXIMIZE VELOCITY
Ship many low-value experiments.
VELOCITY + ABLATION DELTA
33. METRIC HIERARCHY

EXECUTIVE → PRODUCT → OPS → DIAGNOSTIC

AudiencePrimary viewExamples
Executive / ManagementValue + trend + guardrailsAdoption, time saved, cost/success, autonomy, quality, velocity.
ProductUse-case cohortsActivation, habitual adoption, coverage, acceptance, task taxonomy.
AI / EngineeringMechanism and qualityEvals, ablation, model route, retries, latency, tool failures.
OperationsFlow / exceptionsEscalation reason, queue age, approval SLA, failed operations.
Governance / RiskGuardrailsRisk-class autonomy, policy escalations, human override, incidents.
34. METRIC CARD STANDARD

EVERY KPI SHOULD SHIP WITH A CONTRACT

METRIC CARD

Name:
  Autonomy Rate

Purpose:
  measure successful autonomous completion
  among tasks permitted to be autonomous.

Numerator:
  successful autonomous completions

Denominator:
  autonomy-eligible tasks

Exclusions:
  mandatory-HITL risk classes
  test traffic
  duplicate/replayed tasks

Window:
  weekly, rolling 28d

Required slices:
  use_case
  risk_class
  cohort
  product_version

Guardrails:
  acceptance_rate
  eval_pass_rate
  severe_error_rate

Data sources:
  task_completed
  approval_event
  policy_decision

Owner:
  Product + AI Analytics

Version:
  metric_v1
35. TARGETS & THRESHOLDS

DON'T SET A TARGET BEFORE YOU UNDERSTAND THE DISTRIBUTION

BASELINE FIRST

Observe

Measure enough representative history to know normal variability.

TARGET

Desired outcome

Business/product goal, not merely current +10%.

GUARDRAIL

Do-not-cross

Quality/safety/cost/latency boundary.

ALERT

Operational trigger

Statistically/operationally meaningful deviation requiring action.

36. METRIC VERSIONING

DEFINITION CHANGES ARE SCHEMA CHANGES

2026-08:
  adoption_rate_v1 =
    weekly active users / licensed users

2026-09:
  adoption_rate_v2 =
    weekly active eligible users
    / eligible users

Do NOT splice v1 and v2
into one continuous line
without annotation.

Store:
  metric_id
  version
  effective_from
  definition
  denominator_rule
  exclusions
Metric drift is just as dangerous as API drift: the graph can look smooth while the meaning changed underneath.
37. EXPERIMENT READOUT

ONE CHANGE SHOULD HAVE A PRE-REGISTERED METRIC PLAN

CHANGE:
  enable reranker v3

PRIMARY:
  acceptance_rate +3pp

GUARDRAILS:
  cost_per_operation ≤ +8%
  p95 latency ≤ +120ms
  severe_error_rate no increase

DIAGNOSTICS:
  retrieval recall
  evidence coverage
  model route
  task complexity

COHORT:
  research_question
  English + Russian
  medium complexity

BASELINE:
  v2 reranker / control cohort

DECISION:
  ship if primary improves
  and guardrails remain within bounds.
38. FAILURE MODES

HOW AI METRICS LIE

VANITY USAGE
Count prompts/calls and call it adoption/value.
USER + TASK OUTCOME
DENOMINATOR DRIFT
Eligibility changes silently inflate KPI.
VERSION DEFINITIONS
SUCCESS SURVIVORSHIP
Cost measured only on successful attempts; failed runs disappear.
ATTRIBUTE ALL COST
MISSING LABEL = PASS
Unreviewed outputs inflate acceptance.
TRACK REVIEW COVERAGE
GLOBAL AVERAGE
High-performing easy use case hides dangerous low-performing cohort.
SEGMENT
PRE/POST = CAUSE
Seasonality/workload/rollout mix ignored.
CONTROL / ABLATION
TIME SAVED = AI LATENCY
Elapsed and human active time confused.
ACTIVE HUMAN MINUTES
AUTONOMY AT ANY COST
Human controls removed to improve KPI.
RISK + QUALITY GUARDRAILS
39. FRAMEWORK HEALTH METRICS

MEASURE THE MEASUREMENT SYSTEM

DC

Definition Coverage

% executive/product KPIs with formal metric cards.

LC

Label Coverage

% eligible outputs with valid review/outcome labels.

JC

Join Coverage

% successful tasks joinable to runs, costs and reviews.

EL

Eligibility Coverage

% tasks/users with explicit eligibility classification.

BL

Baseline Coverage

% claimed-value metrics with valid baseline/control.

COH

Cohort Stability

Changes in workload mix tracked and visible.

LAT

Data Freshness

Metric data lag relative to reporting SLA.

AB

Ablation Coverage

% expensive architecture mechanisms with measured uplift evidence.

40. MVP IMPLEMENTATION

START SMALL: 7 EVENTS, 6 METRICS, 5 SLICES

MVP EVENTS

1. user_activated
2. task_started
3. task_completed
4. output_reviewed
5. human_escalated
6. external_action_committed
7. deployment_changed

MVP METRICS

1. Adoption Rate
2. Coverage
3. Acceptance Rate
4. Human Active Time Savings
5. Cost per Successful Operation
6. Autonomy / Escalation pair

MVP SLICES

1. use_case
2. risk_class
3. team / cohort
4. product_version
5. model_route

MVP RULE:
  no metric is shown
  without N,
  eligibility rule,
  period
  and definition version.
AI Velocity and Ablation Delta can be added immediately for architecture/product teams, but the first product metric layer should remain small enough to trust.
41. METRIC FRAMEWORK CHECKLIST

BEFORE YOU TRUST A KPI

QuestionRequired
What decision does this metric support?Yes
What is the exact numerator?Yes
What is the exact denominator?Yes
What is eligible / excluded?Yes
What is the unit: user, task, operation, run?Yes
What is the period?Yes
What slices are mandatory?Yes
What is the baseline/control?For value/uplift claims
What quality/risk guardrails pair with it?For adoption/economics/autonomy
What is sample size / label coverage?Yes
Who owns the definition?Yes
What version is this?Yes
42. AI PROMPT TEMPLATE

METRIC DESIGN ASSISTANT

You are designing an AI Metrics Framework.

Do not start from dashboard charts.

1. Define the business/product outcome.
2. Define the unit:
   user, task, operation, run, workflow or experiment.
3. Build the value chain:
   Usage → Quality → Economics → Autonomy → Delivery.
4. For every proposed metric specify:
   - purpose;
   - numerator;
   - denominator;
   - eligibility;
   - exclusions;
   - time window;
   - required cohort dimensions;
   - data source;
   - owner;
   - definition version.
5. Distinguish:
   leading,
   lagging,
   guardrail,
   diagnostic metrics.
6. Never mix:
   attempts with successful operations,
   elapsed time with human active time,
   total tasks with autonomy-eligible tasks.
7. Require baseline/control for claims of improvement.
8. Pair autonomy/economics metrics with quality/risk guardrails.
9. Track missing-label coverage explicitly.
10. If architecture mechanism is expensive,
    require ablation evidence.
11. Show N/sample size for rates.
12. Flag correlation as correlation unless causal design supports more.
13. Prefer a small trustworthy KPI set
    over dozens of ungoverned metrics.
43. PRACTICAL DECISION

WHAT TO BUILD

QuestionAnswer
Separate component?N/A for framework. Instrumentation pipeline, warehouse and dashboard are implementation components.
Minimum 80% value?Task identity + eligibility + adoption + coverage + acceptance + human time + cost/success + autonomy/escalation + version/cohort slices.
When overkill?Tiny prototype before product use exists; still keep basic eval/cost instrumentation.
Trigger?“Is this AI actually useful?”, “Did this architecture improvement pay?”, “Can we safely increase autonomy?”, “What does adoption really mean?”.
How to prove uplift?Baseline/control/ablation + quality guardrails + stable denominators + representative cohorts.
Can rules/code replace LLM?Yes. Metric computation should be deterministic SQL/code. LLM may explain anomalies, suggest slices or draft metric definitions, never calculate authoritative KPIs from prose.
44. DESIGN RULES

THE AI METRICS RULEBOOK

RULE 01

Value tree before dashboard

Usage alone is not value.

RULE 02

Denominator is part of the metric

Never publish a rate without eligibility definition.

RULE 03

Task is the core value unit

Runs/model calls are diagnostics and costs, not business completion.

RULE 04

Baseline before savings

No causal savings claim from current-state observation alone.

RULE 05

Segment by use case and risk

Global averages hide important failures.

RULE 06

Quality guards value metrics

Autonomy, time savings and cost are meaningless if output quality collapses.

RULE 07

Count failed attempts in economics

Cost/success includes retries and failures.

RULE 08

Version metric definitions

A changed denominator is a changed metric.

RULE 09

Ablate complexity

Every expensive architecture layer must earn its cost with measurable delta.

45. FINAL MAP

FROM AI ACTIVITY TO AI VALUE

AI SYSTEM
  ↓
INSTRUMENT TASKS
  user_id
  task_id
  run_id
  use_case
  risk_class
  eligibility
  versions
  ↓
MEASURE USAGE

ADOPTION RATE
  are eligible users active?

COVERAGE
  how much eligible work is AI touching?
  ↓
MEASURE QUALITY

ACCEPTANCE
  accepted as-is?
  minor edit?
  major rewrite?
  reject?

EVALS
  controlled benchmark quality

SEVERE ERROR / GUARDRAILS
  ↓
MEASURE ECONOMICS

TIME SAVINGS
  human active before
  minus
  human active after

COST / SUCCESS
  models
  tools
  infra
  retries
  failures
  human review
  divided by
  successful operations
  ↓
MEASURE AUTONOMY

AUTONOMY RATE
  successful autonomous completions
  / autonomy-eligible tasks

ESCALATION RATE
  human escalations
  / eligible AI tasks

split:
  policy
  ambiguity
  capability
  failure
  ↓
MEASURE ARCHITECTURE VALUE

ABLATION DELTA
  full system
  minus
  system without mechanism
  ↓
MEASURE DELIVERY

AI VELOCITY
  approved idea
  → experiment
  → eval
  → production
  → measurable outcome

═══════════════════════════════════

READ TOGETHER:

ADOPTION ↑
but
ACCEPTANCE ↓
  = usage ≠ value

AUTONOMY ↑
but
QUALITY ↓
  = unsafe automation

COST ↓
but
REWORK ↑
  = local savings, system loss

VELOCITY ↑
but
ABLATION DELTA ≈ 0
  = shipping activity, little value

═══════════════════════════════════

EVERY KPI MUST HAVE:

PURPOSE
NUMERATOR
DENOMINATOR
ELIGIBILITY
EXCLUSIONS
UNIT
PERIOD
COHORTS
BASELINE
GUARDRAILS
DATA SOURCE
OWNER
VERSION
SAMPLE SIZE

═══════════════════════════════════

THE CENTRAL QUESTION IS NOT:

“HOW MUCH AI ARE WE USING?”

IT IS:

“HOW MUCH ELIGIBLE WORK
IS BEING COMPLETED
BETTER,
FASTER,
CHEAPER
AND SAFELY ENOUGH
BECAUSE OF AI —

AND WHICH PART
OF THE ARCHITECTURE
ACTUALLY CREATED
THAT DELTA?”

ECC RETROFIT / PRACTICAL HARNESS INTEGRATION

A. Related ECC ideas. Context-as-cache, scoped memory, lifecycle hooks, selective capabilities, feature flags, deterministic enforcement, provider-neutral adapters and eval-gated learning are applied only where relevant to №84 AI Metrics Framework.

B–E. Existing boundary and placement. The existing conceptual boundary, class METRIC, default N/A and owner Evals / Analytics / Management remain authoritative. Runtime/control/data/offline placement is unchanged; durable state stays outside model context.

F–H. Hooks and contracts. Use bounded PRE_MODEL/POST_MODEL, PRE_TOOL/POST_TOOL, CHECKPOINT and TASK_COMPLETED events as applicable. Illustrative fields and canonical contracts are defined in NEW_CONTRACTS_SPEC.md; no universal schema is implied.

I–J. Security and evaluation. Host-side schema, permission, secret, budget, idempotency and audit checks take precedence over LLM output. Optional mechanisms require a feature flag and WITH/WITHOUT ablation; measure quality, acceptance, correction, latency, cost, escalations and severe errors.

K–L. Task profiles and cross-references. A TaskProfile selects the relevant skill, tool/context slice, memory scope and enforcement profile independently from FAST/STANDARD/DEEP. See cross-reference map, hook spec and ablation plan. Provider adapters remain outside the core.