Outcome
By the end of this lesson, you will be able to read environment and noStop predictions from the auto-tagger, act on confidence-tiered suggestions, and close the tag-coverage gap with auto-tagging.
| Tier | Engineer |
| JTBD | ”Close the tag-coverage gap automatically: auto-tagger handles what IaC + cloud policy missed.” |
| Personas | Platform Engineer · FinOps Lead · DevOps Engineer |
| Prerequisites | T5.M5.1 (tagging strategy) |
| Time | 9 minutes |
| Bloom verb | Read (Apply), Act (Apply), Close (Apply) |
1. Concept
The auto-tagger makes two kinds of predictions per resource:
PREDICTION 1: ENVIRONMENT Values: dev / test / stage / prod
PREDICTION 2: NOSTOP Boolean: can this resource be safely stopped?
EACH WITH CONFIDENCE SCORE (0-100%)Customers accept or rejectEach prediction has a confidence score. Customers accept or reject.
Why these two predictions
ENVIRONMENT: Most important tag in any cost-attribution scheme Drives reporting, showback, scheduling policy Most-missed tag in practice
NOSTOP: Critical "is this safe to act on" signal Prevents accidental production downtime Especially important for auto-remediation
TOGETHER: These two answer 80% of the "what is this resource?" question Other tags (team, cost-center, owner) follow naturally Once env + noStop are clean, the rest is easierThe two-prediction approach focuses on what matters most.
Where predictions appear
INSIGHTS → Auto-Tagging page─────────────────────────────────────────────────────────────────────RESOURCE CURRENT TAGS PREDICTION CONFIDENCE─────────────────────────────────────────────────────────────────────dev-platform-eu-1 (none) env=dev 87% noStop=false 82%
staging-postgres (none) env=staging 91% noStop=true 78%
prod-payment-api (none) env=prod 94% noStop=true 96%
i-0xyz123 (none) env=dev 45% ← low noStop=true 62%For each prediction with high confidence, customer can accept. Low-confidence predictions (<70%) get a “review” suggestion.
Acting on predictions
ACCEPT Tag is applied to the resource in ZopNight's database Synced to the cloud if configured (cloud-tag sync) Resource participates in cost reports going forward
REJECT Tag suggestion ignored Resource remains untagged Auto-tagger won't re-suggest this specific tag
REVIEW Show evidence panel User decides based on evidence Can accept after reading
SNOOZE Re-evaluate later Predictions can change as more data arrives Useful when uncertain right nowThe four actions cover all decision paths.
How confidence affects defaults
CONFIDENCE RECOMMENDED ACTION──────────────────────────────────────────────────────────────≥ 95% HIGH CONFIDENCE Likely safe to accept Bulk-accept is reasonable
85-95% REASONABLE CONFIDENCE Review evidence first Accept after quick verify
70-85% MODERATE CONFIDENCE Definitely review Don't bulk-accept
< 70% DON'T AUTO-ACT Prediction is uncertain Manual investigation required May need more discovery cyclesA team that auto-accepts at 95% threshold sees most predictions land cleanly. Below that: manual review.
Why predictions matter
AUTO-TAGGING CLOSES THE TAG-COVERAGE GAP automatically
TYPICAL JOURNEY: Without auto-tagger: 60-70% tag coverage With auto-tagger: 90+% tag coverage
The remaining 5-10%: edge cases that need manual ID
ONCE TAGGED: Resources participate in cost reports Showback works Recommendations apply Scheduling can use tag-based selection
THE 30% GAP IS SIGNIFICANT: Untagged resources = orphaned cost Hard to attribute; hard to optimize Auto-tagger fills this gapThe gap closure compounds: every tagged resource enables more downstream value.
How predictions are generated
SOURCES OF PREDICTION:
Resource name patterns: "dev-*" → env=dev "prod-*" → env=prod "staging-*" → env=stage Deterministic pattern matching (no ML)
Parent context: Resource in account "acme-dev" → env=dev Resource in account "acme-prod" → env=prod
Sibling tagging: Other resources in same VPC/cluster Inheritance of common tags
Resource type heuristics: RDS Multi-AZ → likely prod EBS encrypted → likely prod Smaller instance types → likely dev
Existing tags + instance config: Other tags already on the resource Instance type / size (large = likely prod) Encryption / Multi-AZ config
COMPOSITE SCORE COMBINES all of these Higher when multiple rules agree Lower when rules conflictThe prediction is composite rule-based: a fixed set of signals (naming patterns, existing tags, instance config, group and account context) is scored deterministically. Same inputs always produce the same prediction. There is no ML model, no ensemble, and no behavioral signals like CloudTrail-user or time-of-creation.
Cost-coverage impact
EXAMPLE: untagged resource at $500/mo
WITHOUT AUTO-TAGGER: Resource in "untagged" bucket No team attribution No environment-specific policy $500/mo allocated to "shared overhead"
WITH AUTO-TAGGER: Predicted env=dev, noStop=false After accept: tagged appropriately Cost reports show under correct team's spend Scheduling can apply if non-prod policy exists Cost attribution: clean
ACROSS THE ESTATE: Every untagged resource is a coverage gap Auto-tagger systematically closes gaps Per-team chargeback accuracy improves Recommendation accuracy improvesThe tagging is the enabler; auto-tagging is what makes it tractable at scale.
2. Demo
A team’s auto-tagging adoption journey:
T+0 Before: tag coverage at 62% (untagged accumulated over years) 47 resources with no team or env tag
T+5 min Open Insights → Auto-Tagging 47 resources have predictions
T+15 min Review predictions: 18 high-confidence (≥95%): Mostly env=prod tags Bulk-accept all
14 moderate (85-95%): Mostly env=dev Review evidence for each Accept 12 of 14 Reject 2 (one mistakenly tagged dev that's actually staging)
10 low-moderate (70-85%): Mixed predictions Accept 4 (high signal alignment) Reject 6 (signals don't align)
5 low (<70%): Manual investigation needed Skip for now
T+45 min Accepted: 34 predictions Tag coverage improves: 62% → 76%
T+30 DAYS: Re-run auto-tagger as new resources discovered Many new resources have predictions Tag coverage climbs: 76% → 91%
T+90 DAYS: Stable at ~93% tag coverage Remaining 7%: edge cases requiring manual ID (Old resources with cryptic names; orphans from old projects)
OUTCOME: 62% → 93% tag coverage in 3 months Effort: ~3 hours total across the period Cost attribution accuracy: dramatically improved Per-team chargeback: trustedThree months from 62% → 93% tag coverage via auto-tagging.
3. Hands-on (5 min)
Use the auto-tagger:
□ STEP 1: Open Insights → Auto-Tagging Total predictions: _____ Resources to evaluate: _____
□ STEP 2: Review by confidence High (≥95%): _____ predictions Moderate (85-95%): _____ predictions Low-moderate (70-85%): _____ predictions Low (<70%): _____ predictions
□ STEP 3: Bulk-accept high-confidence Accept all ≥95% predictions Confirm Tag coverage improves: _____% → _____%
□ STEP 4: Review moderate For each: Read evidence Accept or reject Note: how many felt obvious vs uncertain
□ STEP 5: Plan for low-confidence Mark for manual investigation Owner: __________ Quarterly cleanup targetA 30-minute session accepts most actionable predictions. Quarterly cadence keeps coverage high.
4. Knowledge check
Q1
A prediction at 92% confidence:
A. Always accept
B. Reasonable confidence: review the evidence and accept if it matches your knowledge of the resource. Moderate-high confidence still warrants quick review. Bulk-accept reserved for ≥95%.
C. Reject
D. Random
Show answer
Correct: B. Moderate-high confidence still warrants quick review.
Q2
The two predictions per resource are:
A. Cost and savings
B. Environment (dev/test/stage/prod) and noStop (boolean: safe to stop). These two cover most of the “what is this resource?” question. Other tags follow naturally once env + noStop are clean.
C. Region and account
D. Random
Show answer
Correct: B. Env + noStop are the two predictions.
Q3
A prediction at 55% confidence:
A. Auto-accept
B. Don’t act automatically: manually investigate. Confidence is too low to safely act. May need more discovery cycles to gather signals. Or: resource is genuinely ambiguous.
C. Reject by default
D. Random
Show answer
Correct: B. Low confidence = manual investigation.
5. Apply
Insights → Auto-Tagging is the canonical surface. Predictions update as new data arrives.
For your team: monthly auto-tagging review; quarterly deep-dive on remaining low-confidence resources.
Related lessons
- L2: Inference signals (next)
- L3: Accept / reject patterns
- L4: Drift detection
- T5.M5.1.L2: MVT (Minimum Viable Tags)
Glossary terms touched
Auto-tagger · Environment prediction · noStop prediction · Confidence score