Appendix B: Templates and Tools
This appendix collects what a practitioner needs at the desk while applying the five environmental functions of AI Environment: the records of the book as skeletons to copy, the optimisation hierarchy as a checklist, the limits of an agent, the age of evidence and the categories of tools. Appendix A shows the same records filled in for Nordvik Region and Hansa Mutual.
The appendix names no product. Tools are described by category and by what each must provide, because a category can be replaced by whatever the organisation has procured, and a product name in a book dates quickly. Every figure in a record carries its boundary, its year, its source and whether it is vendor-reported or independent (the evidence-boundary form of the authoring guide); a skeleton field left empty means that the figure has not been established, not that it is zero.
How to use this appendix
- Count first (B.1). A use needs a footprint record at a stated boundary before it is justified, designed or reported. The record is the start of the other records.
- Compare before you scale (B.1, B.3). A benefit record compares the use with the non-AI alternative and states the rebound. An unproven benefit does not offset a measured footprint.
- Work down the hierarchy (B.2). Ask whether the use is needed before choosing the model, and choose the model before tuning the serving. A lower level is not a reason to skip a higher one.
- Cap every agent and every scaled use (B.3, B.4). A budget set by someone other than the agent, measured outside it, with a stop path outside the system it stops.
- Date everything (B.5). Every record that holds a result carries an evidence date, and evidence expires.
B.1 The five functions, the tests and the records
Table B.1.1: The five functions, the question each answers, its net-benefit test and its records.
| Function | Question | Net-benefit test | Records |
|---|---|---|---|
| E1 Measure | What does this use cost the environment, at which boundary and in which year? | Count | Footprint record FPR-*; supplier disclosure record SDR-* (as input) |
| E2 Justify | Is it worth its footprint, compared with the non-AI alternative and the rebound? | Compare | Benefit record BEN-*; AI Decision Record |
| E3 Design | Can the same benefit be had for less? | Cap (design the budget) | Environmental budget EBG-*; mandate manifest MND-*-ENV-*; hardware lifecycle record HWL-* |
| E4 Operate | Does it run where and when grid and water allow, within its budget, observed? | Cap (hold the budget) | Placement policy PLC-*; environmental budget; runbook manifest RBK-* |
| E5 Account | Can the organisation comply, report, claim, procure and answer? | All three, with evidence | Environmental claim record CLM-*; supplier disclosure record; the environmental register |
B.2 The optimisation hierarchy as a checklist
Work the levels in order and record, for each, what was done, what it saved, at which boundary and year, and how that was checked. A saving that has not been measured is an estimate and is labelled so.
Table B.2.1: The optimisation hierarchy, level by level, with the evidence to record.
| Level | Question | Evidence to record |
|---|---|---|
| 1 Necessity | Is the use needed? Can it be declined, merged or stopped? Is any service idle? | The decision, the owner, the volume before and after, the date |
| 2 Sufficiency of model | Is this the smallest model, with the lowest reasoning effort and the shortest context that meets the requirement? | The requirement, the evaluation result per candidate, the footprint per unit of each, the choice |
| 3 Efficient serving | Are quantisation, caching, batching, routing and retrieval used where they help? | The measured change in energy per unit, the quality check, the boundary |
| 4 Place and time | Does the work run where and when the grid is cleaner and water is not scarce, and is heat reused where possible? | The placement policy, the intensity data source and method (average or marginal), the measured effect |
| 5 Hardware and lifecycle | Is utilisation high, is embodied impact counted, and is reuse and retirement planned? | The hardware lifecycle record |
B.3 Record templates
Each template is a skeleton to copy into version control. It uses exactly the field names of the records in Appendix A (A.4 to A.13), which the schemas of A.14 check, and the chapters that own the records print the same fields. Replace the placeholder values, keep every field, and write null for what is not established and name a person for it: a field is never deleted. Where a field takes one of a fixed set of values, the comment lists them. Values written null, "" or 0 are placeholders, not results. Appendix A gives the section where each record is explained and filled in for Nordvik Region and Hansa Mutual; the identifiers NR and HM stand for your own organisation's codes.
Footprint record (A.4)
# register/fpr/FPR-NR-0000.yaml (skeleton; copy and fill)
id: FPR-NR-0000
revision: 1
use: "" # the use, as the service record names it
service: SVC-NR-000
owner: "" # the use's owner
accountable: ""
functional_unit: "" # per query | per task | per agent run | per resolved enquiry
period: "" # a month, 2032-09
volume: null # units in the period
boundary:
scope: operational # operational | lifecycle
included: [] # for example accelerator servers, host, facility overhead
excluded: [] # stated, never silent
carbon_basis: location-based # location-based | market-based
grid_g_per_kwh: {value: null, year: null, source: ""}
method:
type: "" # for example metered pool with allocation, bottom-up, top-down
inputs: []
allocation: ""
multipliers: {queries_per_unit: "", reasoning: false}
result:
statistic: mean # mean | median | p95
energy_wh_per_unit: {low: null, central: null, high: null}
carbon_g_per_unit: {low: null, central: null, high: null}
dominating_assumptions: [] # the assumptions that cause most of the range
independence: {produced_by: "", checked_by: "", independent_of_producer: false}
independent_check: {status: none} # none | partial | done
evidence: {measured_on: null, valid_days: 90, result_of_check: ""}
labels: [illustrative] # remove only for a real, checked figure
Benefit record (A.5)
# register/ben/BEN-NR-0000.yaml (skeleton; the number matches the footprint record of the use)
id: BEN-NR-0000
footprint_record: FPR-NR-0000
use: ""
functional_unit: ""
owner: "" # outside the sponsor's line
sponsor: ""
intended_benefit: {unit: "", ledger: B} # ledger A | B | C
baseline: {kind: "", period: "", control: ""} # for example randomised hold-out, paired comparison
alternative: {chosen: "", untested: ""} # the non-AI alternative
measured_effect: # shape of the use; a routing use gives correct_route_pct: {agent: null, rules: null}
calls_per_100_enquiries:
control: null
assistant: null
difference: {low: null, central: null, high: null}
quality: {sampled: null, materially_wrong: null}
cost_eur_per_month: {use: null, calls_avoided: null} # a separate ledger, never netted
comparison:
footprint_g_per_call_avoided: {low: null, central: null, high: null} # footprint_g_per_extra_correct_route for a routing use
footprint_of_alternative: unknown
environmental_effect: not shown # shown | not shown
rebound: {volume_at_start: null, unit: "", recheck: null}
decision: {value: defer, by: "", date: null, record: AIDR-1300, threshold: ""} # proceed | redesign | defer | decline
claims_allowed: []
claims_withheld: []
evidence: {measured_on: null, valid_days: 180, expires: null} # expires = decision date + valid_days
labels: [illustrative]
Environmental budget (A.6)
# budget/EBG-NR-0000.yaml (skeleton)
id: EBG-NR-0000
use: FPR-NR-0000
status: proposed # proposed | active | suspended | retired; active only after a passed drill
period: monthly # monthly | quarterly | yearly
set_by: "" # never the agent and not the use's own service owner
approved_by: ""
unit: ""
budget: {energy_kwh_per_month: null, carbon_g_per_unit: null, cost_eur_per_month: null}
quota: {units_per_month: null, derived_with_wh_per_unit: null} # derived once at approval
conversion: {wh_per_unit: null, measured: null, valid_days: 90} # latest estimate and its date
baseline_units_per_month: null
basis: "" # the footprint record, its date and the headroom
measured_by: {energy: "", volume: "", reviewed_by: ""} # outside the system it limits
stop:
holder: []
path: "" # outside the system it stops
independent_of_assistant: true
last_drill: null # null: not counted as a cap
drill_valid_days: 90
drills: [] # {date, result: pass | fail, minutes}
on_breach: {alert: 0.80, defer: 1.00, stop: 1.10} # shares of the quota
rebound_check: ""
rebound_alert_share: 0.20
resolution_floor: 0.80 # a limit is not met by failing to resolve
Placement policy (A.7)
# PLC-NR-0000 (skeleton)
id: PLC-NR-0000
status: proposed # proposed | active | suspended | retired
owner: ""
reviewed_by: []
approved_by: ""
review_by: ""
applies_to: [] # the classes of work the policy governs
signal:
basis: "" # for example location-based, hourly, average
source: ""
zone_of_site: {} # site: grid zone, null until disclosed
space_shift: false # false until a basin and sovereignty look-up exists
regions: []
classes:
batch_class: # one entry per class of deferrable work
window: {earliest_start: "", latest_start: ""}
deadline: ""
runtime_h: null
time_shift:
enabled: false # on only after a replay record shows a saving above min_saving
min_saving: {relative: 0.10, kg_co2e: 1.0}
replay_record: null
budget: {record: EBG-NR-0000, energy_kwh_per_month: null, measured_by: ""}
stop:
holder: []
path: ""
independent_of_scheduler: true
last_drill: null
drill_valid_days: 90
scheduler_may_not: [edit_policy, add_region, change_zone, raise_budget, disable_meter]
Supplier disclosure record (A.8)
# supplier/SDR-NR-0000.yaml (skeleton)
id: SDR-NR-0000
supplier: ""
holder: "" # the person who owns the relationship
received_on: null
source_document: ""
valid_until: null # the earliest item date
items:
- {item: "", boundary: "", year: null, kind: vendor-reported, check: none, rung: null, valid_until: null}
# kind: metered | calculated | estimated | vendor-reported
not_given:
- {item: "", rung: null, asked_via: ""} # what the supplier does not give, and where it was asked for
independent_check: {status: none, next: ""}
evidence_age: {limit_days: 365, event_triggers: [region change, hardware pool change, method change]}
published_by_region: {allowed: "", basis: ""}
review: {due: null, by: ""}
Environmental claim record (A.9)
# claims/CLM-HM-0000.yaml (skeleton; not publishable until the approver is named and the fields are filled)
id: CLM-HM-0000
revision: 1
supersedes: {revision: null, statement: "", status: withheld} # remove for a first revision
statement: "" # the exact sentence as it is to be published
status: not-publishable # not-publishable | approved | published | withdrawn
claim_kind: footprint # footprint | benefit | commitment; one kind per record
object: FPR-HM-0000
boundary: {phase: "", energy: "", site: "", carbon: ""}
period: ""
baseline: {what: ""}
unit: ""
evidence:
kind: none # measured | estimated | vendor-reported | none
measured_on: null
valid_days: 90
valid_until: null
independent_check: {by: "", result: ""} # by a person who did not make the change
claimed_by: ""
reviewed_by: []
approved_for_publication_by: null # a named person; never an agent
decision: {record: AIDR-1350, approved_on: null}
published: {where: "", status: "", withdraw_if: [evidence expires, model or routing changes]}
withheld: [] # statements the evidence does not support
labels: [illustrative]
Hardware lifecycle record (A.10)
# HWL-HM-0000 (skeleton; kind inventory for a pool, clause_set for a contract)
id: HWL-HM-0000
kind: inventory # inventory | clause_set
pool: ""
owner: ""
facility_owner: ""
reviewer: ""
created: null
valid_until: null # 90 days for measured utilisation
inventory: {servers: null, accelerators: null, installed: "", planned_life_years: 5}
utilisation: {measured: null, target: null, floor_alert: null}
energy: {facility_mwh_per_year: null, location_based_g_per_kwh: null}
embodied:
status: "" # for example estimated from a disclosed baseboard figure
baseboard_kg_per_server: null
host_kg_per_server: null
excluded: []
stress_multiplier: null
options:
- {id: A, what: "keep as is", servers: null, facility_mwh: null, tco2e_year: null}
decision: {aidr: "", chosen: "", then: "", date: null}
stop_path: {who: "", how: "", outside_the_agent: true}
end_of_life: {route: "", erasure_certificate: "", report: []}
Mandate manifest of an environmental agent (A.11)
# MND-NR-ENV-0000 (skeleton)
id: MND-NR-ENV-0000
agent: ""
service: SVC-NR-000
decision: AIDR-1300
footprint: FPR-NR-0000
status: proposed # proposed | active | suspended | retired
resource:
set_by: "" # not the agent and not the team that uses it
approved_by: ""
basis: proxies # proxies (calls, tokens, sandbox time) until a footprint record exists; then energy
per_run: {model_calls: null, tool_calls: null, wall_s: null, energy_wh: null, on_breach: ""}
per_day: {energy_kwh: null, soft_at: null, on_hard: ""}
per_month: {energy_kwh: null}
background_or_scheduled_work: not delegated
measured_by: {calls_and_tokens: "", accelerator_seconds: "", reviewed_by: ""}
delegation: # a Level per action class; an unlisted action is undelegated
start_run_on_registered_event: 3
retry_inside_per_run_limit: 3
spawn_sub_agent: 0
schedule_or_background_work: 0
provision_or_scale_compute: 0
change_own_limits_meter_credential_or_stop: 0
publish_environmental_claim: 0
never_level_4: [provision_or_scale_beyond_budget, disable_meter_budget_or_stop, publish_environmental_claim, decide_about_named_person]
stop:
path: []
holders: []
target_minutes: null
last_drill: null
drill_valid_days: 90
drills: [] # {date, result, minutes, finding}
labels: [illustrative]
Runbook manifest of a budget breach (A.12)
# runbook/RBK-NR-ENV-000.yaml (skeleton)
runbook: RBK-NR-ENV-000
title: ""
version: 1
service: SVC-NR-000
budget: EBG-NR-0000
owner_unit: ""
accountable: ""
decision_record: AIDR-1300
trigger:
- {signal: quota_share, at: 0.80, action: alert} # signal: quota_share | projected_share; action: alert | defer | stop
- {signal: quota_share, at: 1.00, action: defer}
- {signal: quota_share, at: 1.10, action: stop}
delegation:
profile: DLG-NR-ENV-0000
automatic_actions: [alert, defer, stop]
assistant_may: []
steps: {detect: "", stop: "", assess: "", restart: "", review: ""}
fallback: {path: "", staffed: ""}
verification: {meter: "", volume: ""} # outside the assistant
stop:
held_by: []
path: ""
independent_of_assistant: true
drill: {last: null, result_minutes: null, target_minutes: null, valid_days: 90}
review: {due: null, approved_by: ""}
Environmental register and scale-up decision (A.13)
# register/environment-nr.yaml (skeleton; one file per organisation, one entry per AI use)
register: environment-nr
organisation: Nordvik Region # Nordvik Region | Hansa Mutual; your own name
register_owner: ""
as_of: null
limits_days: {footprint: 90, benefit: 180, stop_drill: 90, supplier: 365, hardware: 90}
uses:
- use: FPR-NR-0000
service: SVC-NR-000
owner: ""
evidence:
footprint: {record: null, measured: null}
benefit: {record: null, measured: null} # for a benefit record the clock starts at the decision date
stop_drill: {record: null, measured: null}
supplier: {record: null, measured: null}
hardware: {record: null, measured: null}
tests: {count: open, compare: open, cap: open} # pass | fail | open
decision: AIDR-1300
claim: null
# decision/AIDR-1300.yaml (skeleton; AI Decision Record at the scale-up gate)
id: AIDR-1300
date: null
organisation: Nordvik Region
subject: ""
use: FPR-NR-0000
trigger: scale-up # scale-up | renewal | new-use | breach | claim
tests:
count: {result: open, basis: ""}
compare: {result: open, basis: ""}
cap: {result: open, basis: ""}
decision: ""
conditions:
- {what: "", by: null} # each open test has a condition and a date
decided_by: ""
claim_effect: ""
review: null
labels: [illustrative]
B.4 The agent's limits
These limits apply to every agent that can start work, at every degree of delegation. They are written into the mandate manifest MND-*-ENV-* (A.11) and enforced outside the agent. The ceilings are those of Chapter 4: starting ceilings that a named authority may raise with evidence, and four prohibitions that no evidence lifts.
Table B.4.1: Starting ceilings by action class and what stays outside the agent.
| Action class | Starting ceiling | What stays outside the agent |
|---|---|---|
| Measure | Level 3 Bounded | The meter and the right to change what it reads |
| Recommend | Level 3 Bounded | Acceptance of the recommendation by a named owner |
| Schedule | Level 3 Bounded, within the placement policy | The floors (deadline, sovereignty, resilience) and the policy itself |
| Provision | Level 2 Supervised, within the budget | The budget and its setter |
| Scale | Scale down Level 3 Bounded; scale up Level 2 Supervised | The spend ceiling in the delegation profile |
| Stop | A gateway control, not a delegated act; the agent may request a stop of its own run at Level 3 | The stop path and its holders |
| Publish | Level 1 Assisted (drafting) | The named approver and the evidence |
Level 4 Autonomous is never used for an agent that provisions or scales its own compute, or starts work, beyond a budget set by someone else; for an action that disables the measurement, the budget or the stop path itself; for the publication of an environmental claim; or for a decision about a named person's access to work.
B.5 Evidence age
Evidence expires. The periods below are those of Chapter 4 and the conventions of A.3, which the register of Chapter 24 enforces; an organisation sets its own and records them in the register. A change of model, region, hardware, supplier or contract, or of volume beyond a stated trigger, expires a record at once.
Table B.5.1: Validity of each record, when its clock starts and why.
| Record | Validity | Clock starts at | Reason |
|---|---|---|---|
| Footprint record of a running use | 90 days | The measurement | Volume, model version and grid change |
| Benefit record | 180 days | The decision date (Chapter 9) | The alternative and the rebound change |
| Stop drill | 90 days | The passed drill; a failed drill does not count | The stop path must work when it is needed |
| Supplier disclosure record | 12 months, or the supplier's reporting year | The date received | Suppliers report annually and change method |
| Hardware record | 90 days (measured utilisation) | The measurement | Utilisation changes |
| Environmental claim record | The reporting year at most; a narrow claim carries its own valid_until |
The measurement behind the claim | A claim is true of a period, not for ever |
| Placement policy | Reviewed at review_by; replay record before time shift |
The approval | Grid data and region offers change |
B.6 Tool categories and what each must provide
The organisation needs the categories below. Each row says what the category must provide for the functions to work, not which product to buy. See AI Operations, Chapter 11 for capacity, cost and energy practice (34, Nordqvist, 2026) and AI Platform and Security, Chapter 22 for cost and sustainability at platform level (27, Nordqvist, 2026).
Table B.6.1: Tool categories, the functions they serve and what each must provide.
| Category | Functions | Must provide |
|---|---|---|
| Energy and power telemetry | E1, E4 | Metering at a stated boundary (accelerator, server, rack, facility) with timestamps, from a source outside the workload |
| Usage and cost tagging | E1, E2 | Tags that tie every request, job and service to a use and an owner, so that footprint per use can be computed |
| Carbon and water data services | E1, E4 | Grid carbon intensity by region and hour with the method stated (average or marginal) and water-stress indicators with their source |
| Footprint calculation method | E1 | A published method with a named version and functional unit (for example the Software Carbon Intensity specification) and the boundary it uses |
| Model and evaluation harness | E2, E3 | Quality evaluation per candidate model and per task, with the energy per unit measured on the same hardware |
| Model router and gateway | E3, E4 | Routing by task and requirement, caching, per-use quotas, logs that feed the footprint record |
| Scheduler and job runner | E4 | Deferral and placement by the placement policy, a budget it cannot raise itself, a stop path outside the scheduler |
| Budget and quota enforcement | E3, E4 | Limits set by a person other than the owner of the use, enforced by the platform, with warnings and a stop |
| Observability | E4 | Dashboards and alerts on energy, carbon, water and cost per use, evaluated outside the workload |
| Supplier disclosure intake | E5 | A place for supplier figures with boundary, year and check, and a way to compare them like with like |
| Environmental register | E5 | One row per use with owner, records, budget and evidence age, and a review date |
| Claims review | E5 | A review step that checks every claim against its claim record before it is published |
| Hardware asset management | E3, E5 | Inventory, utilisation, embodied figures where disclosed, reuse and retirement routes |
Sources
- Nordqvist, G. (2026). AI Operations: How Artificial Intelligence Transforms IT Operations in European Enterprises and Public Authorities. The Transformed Architecture, Book 10. (34, Nordqvist, 2026)
- Nordqvist, G. (2026). AI Platform and Security: Building and Securing AI Platforms in European Enterprises and Public Authorities. The Transformed Architecture, Book 3. (27, Nordqvist, 2026)