DC

Data Center Health

Systems-first brief

9–13 minutes

Annual Predictive-Maintenance Maturity Assessment for AI Data Centers: Measure the Operating System, Not the Sensor Count

A predictive-maintenance program should not be judged by how many
sensors were installed, how many dashboards were created or how many
alerts were generated. It should be judged by whether the organization
can repeatedly turn trustworthy condition evidence into safe, timely and
verified maintenance decisions.

That capability changes over time. Staff rotate. Equipment ages.
Models drift. Asset mappings change. Vendors update platforms. New sites
join the portfolio. A control that passed during the pilot can quietly
weaken during production.

An annual maturity assessment gives management and engineering teams
a structured way to test whether the complete operating system still
works—from asset strategy and sensor health to work-order execution,
cybersecurity, verified outcomes and continual improvement.

Abbreviation reference

  • BMS — Building Management System: Supervisory
    platform for building services such as cooling and environmental
    systems.
  • EPMS — Electrical Power Monitoring System: Platform
    for electrical distribution, loading, power quality and related
    alarms.
  • DCIM — Data Center Infrastructure Management:
    Software connecting facility, asset, capacity and service-dependency
    information.
  • CMMS — Computerized Maintenance Management System:
    System for maintenance assets, work orders, job plans, labor, parts and
    history.
  • OT — Operational Technology: Systems that monitor
    or control physical equipment and processes.
  • KPI — Key Performance Indicator: A defined measure
    used to assess performance or control effectiveness.
  • OEM — Original Equipment Manufacturer: The
    equipment manufacturer and source of product-specific requirements.
  • RACI — Responsible, Accountable, Consulted and
    Informed:     A method for assigning roles in a process or
    decision.
  • SOP/MOP/EOP — Standard, Method of Procedure and Emergency
    Operating Procedure:     Controlled instructions for normal,
    planned and emergency work.

This
is a management assessment—not a certification claim

ISO 55001:2024 specifies requirements for establishing, implementing,
maintaining and improving an asset-management system. ISO 17359:2018
provides general procedures for setting up a condition-monitoring
program for machines. The U.S. Department of Energy describes effective
operations and maintenance as an integrated program that balances
reactive, preventive, predictive and reliability-centered
approaches.

These sources support the management-system and condition-monitoring
principles used in this assessment. They do not publish the five-level
predictive-maintenance scale used here, and completing the workbook does
not establish certification or conformity with any standard.

The maturity levels are an operational decision aid. Each
organization should approve its criteria, evidence requirements, scoring
rules and risk gates before using the result for investment, assurance
or portfolio comparison.

Assess capability,
not technology ownership

Buying a vibration platform does not prove that rotating-equipment
degradation will be detected and acted upon. Connecting a battery
monitor does not prove that cell drift will create a usable maintenance
response. Integrating analytics with a Computerized Maintenance
Management System (CMMS) does not prove that the resulting work order
will contain enough evidence or that the original condition will be
retested.

The assessment should therefore examine eight connected
dimensions:

  1. Strategy and scope — Is predictive maintenance tied
    to asset criticality, failure consequences and business objectives?
  2. Asset and failure-mode coverage — Are assets,
    components, measurement points and warning signatures correctly
    mapped?
  3. Measurement and data trust — Are calibration,
    sensor health, timestamps, operating context, lineage and data quality
    controlled?
  4. Analytics and decision assurance — Are rules and
    models validated for their intended use, limitations and failure
    costs?
  5. Work management and verified closure — Do credible
    events become owned work with procedures, retesting and acceptance
    evidence?
  6. People and governance — Are decision rights,
    competence, escalation and change control clear?
  7. Cybersecurity and resilience — Are connected
    sensors, gateways, data flows, remote access and recovery
    protected?
  8. Performance and continual improvement — Are
    outcomes measured, lessons converted into controlled changes and value
    claims supported?

A high score in analytics cannot compensate for weak work-order
execution. A strong sensor-health process cannot compensate for
unmanaged remote access. Maturity belongs to the complete control
loop.

Use five levels with
observable evidence

Level 1 — Reactive

Activity is mainly event-driven and dependent on individuals. Sensors
or dashboards may exist, but scope, ownership and evidence requirements
are inconsistent. Alerts often require manual reconstruction before
anyone can decide what to do.

Typical evidence:

  • Isolated monitoring tools without an approved program charter
  • Incomplete asset or measurement-point mapping
  • Generic thresholds with little operating-context validation
  • Informal escalation through chat or email
  • Work closed without comparable retesting
  • Benefits described through anecdotes rather than traceable
    cases

Level 2 — Defined

The organization has documented processes, named owners and a bounded
asset scope. Baseline, validation, work-order and closure requirements
are defined, but implementation is uneven or still relies heavily on
manual control.

Typical evidence:

  • Approved pilot or program scope
  • Asset and failure-mode register
  • RACI and escalation matrix
  • Data-quality and calibration rules
  • Standard work-order fields and retest expectations
  • Initial KPIs with documented definitions

Level 3 — Controlled

The process operates consistently on the approved scope. Data quality
is monitored, event validation is repeatable, alarm noise is controlled,
work is routed through the CMMS and closure requires evidence.

Typical evidence:

  • Tested suppression, persistence and deduplication rules
  • Traceable baseline and threshold approvals
  • Sensor-health and clock-offset monitoring
  • Work orders linked to source evidence
  • Retest results and acceptance criteria recorded
  • Periodic rule, access and performance reviews

Level 4 — Integrated

Predictive maintenance is integrated with asset strategy, Building
Management System (BMS), Electrical Power Monitoring System (EPMS), Data
Center Infrastructure Management (DCIM), CMMS, cybersecurity, financial
review and site governance. Decisions consider service consequence,
redundancy and operating context.

Typical evidence:

  • Asset-to-service dependencies used in prioritization
  • Common portfolio taxonomy with approved local extensions
  • Model and rule release gates
  • Supplier and remote-access controls integrated with work
    management
  • Comparable KPI reporting across sites and equipment classes
  • Finance-verified realized benefits separated from avoided-risk
    scenarios

Level 5 — Optimizing

The program uses verified outcomes to improve rules, procedures,
training, sensor coverage and investment choices. Cross-site learning is
controlled, performance is compared fairly and the organization can
adapt without weakening safety or accountability.

Typical evidence:

  • Closed-loop analysis of false positives, missed detections and
    repeat alerts
  • Controlled transfer of validated lessons across sites
  • Trend-based maturity targets and investment decisions
  • Independent challenge of models, benefit claims and evidence
    quality
  • Retirement of low-value rules and scaling of proven use cases
  • Demonstrated recovery exercises and continuous improvement
    actions

Level 5 does not mean perfect automation. It means the organization
can learn and adapt while preserving human authority over
high-consequence maintenance decisions.

Score evidence before
scoring maturity

Self-assessments become unreliable when respondents score intent
instead of implementation. A policy draft should not receive the same
credit as an approved process demonstrated through current records.

Each criterion should record both:

  • Implementation score: How consistently is the
    control operating?
  • Evidence strength: How reliable, current and
    representative is the proof?

A practical evidence scale is:

  • 0 — None: No evidence supplied.
  • 1 — Stated: Interview or narrative evidence
    only.
  • 2 — Documented: Approved process, register or
    configuration exists.
  • 3 — Demonstrated: Current operating records show
    the control working across the assessed scope.

The workbook uses evidence strength to adjust the raw maturity
result. This prevents a well-written procedure with no operating records
from appearing equal to a demonstrated control.

Do not let averages hide
critical gaps

An overall average is useful for trend reporting, but it must not
override mandatory safety, integrity or accountability controls.

Critical gates should include requirements such as:

  • High-consequence work requires corroborating evidence and human
    approval.
  • Critical assets and measurement points have accountable owners.
  • Sensor quality and timestamp integrity are checked before
    action.
  • Remote access is controlled, time-bound and reviewable.
  • Completed work is retested against the original condition.
  • Material model, rule and baseline changes are approved and
    traceable.
  • High-risk findings have assigned actions and due dates.

If a critical gate fails, the assessment should hold the program at
its current level even when the weighted average suggests
advancement.

Assess by sampling real
cases

Interviews and procedures are necessary, but the strongest evidence
comes from tracing complete operating cases.

Sample a balanced set that includes:

  • A validated condition that became planned maintenance
  • A high-consequence event that required operational escalation
  • A false positive or invalid-data event
  • A repeated alert after corrective work
  • A suppressed event during approved maintenance
  • A rule or model change
  • A vendor remote-access session
  • A cybersecurity or data-integrity incident exercise
  • A claimed realized saving or avoided-risk scenario
  • A retired monitoring rule or deferred investment decision

For each case, trace the chain from original measurement through
validation, prioritization, work, retest, closure, learning and
financial classification. A process that works only in presentations is
not mature.

Run the
assessment as an annual control cycle

1. Set the scope

Define included sites, systems, asset classes, platforms, vendors and
assessment period. Record material exclusions and why they were
excluded.

2. Confirm assessors and
independence

Use a cross-functional team covering operations, maintenance,
reliability, controls, data, cybersecurity and finance. The owner of a
control can provide evidence, but material conclusions should receive
independent challenge.

3. Freeze the scoring rules

Approve criteria, weights, evidence rules, critical gates and
maturity thresholds before reviewing results. Changing the scoring model
after seeing the score undermines comparability.

4. Collect and sample
evidence

Use current records from the assessed period. Record source, owner,
date, scope, sample size and limitations. Do not accept screenshots
without enough context to establish origin and time.

5. Calibrate ratings

Assessors should compare difficult or borderline criteria together.
The purpose is consistent interpretation—not negotiating a higher
score.

6. Approve gaps and actions

Every material gap should have a risk statement, action, accountable
owner, due date, required evidence and acceptance authority.

7. Set a target state

The target should reflect risk and operating need. Not every
dimension must reach Level 5. A stable Level 3 control may be sufficient
for a low-consequence asset class, while a protected AI load path may
justify Level 4 integration and stronger evidence.

8. Review progress quarterly

The maturity assessment is annual, but corrective actions should not
wait a year. Review overdue high-risk actions, failed gates,
deteriorating KPIs and major program changes quarterly.

Compare sites carefully

Portfolio reporting should show where support or control improvement
is required, not create a simplistic league table.

Before comparing sites:

  • Use the same criterion definitions and evidence scale.
  • Preserve local operating context and approved deviations.
  • Compare equivalent asset classes and monitoring scope.
  • Normalize volume-based KPIs where necessary.
  • Separate missing evidence from genuinely weak performance.
  • Record changes to the scoring model between years.
  • Explain material changes in scope, staffing, technology or
    risk.

A site may score lower because it supplied honest evidence and
identified gaps, while another site appears stronger because its
assessment was superficial. Independent calibration matters.

Measure
improvement, not score movement alone

Useful annual outcome measures include:

  • Percentage of critical criteria with demonstrated evidence
  • Failed critical gates and time to closure
  • High-risk actions overdue
  • Data availability and sensor-health performance
  • Actionable-alert and false-positive rates
  • Detection-to-validation and validation-to-assignment time
  • Work orders with complete evidence
  • Completed work with comparable retesting
  • Repeat-alert rate after intervention
  • Planned versus emergency interventions
  • Verified realized benefits
  • Confidence-adjusted avoided-risk scenarios
  • Rules scaled, redesigned or retired after review

A higher maturity score without better operating evidence is not
improvement. The objective is safer and more reliable decisions.

Put this into practice

Use the AI Data Center Annual Predictive-Maintenance Maturity
Assessment Workbook to score eight capability dimensions,
record evidence strength, test critical gates, compare current and
target states, manage corrective actions, track annual and quarterly
reviews, and present an executive dashboard without allowing averages to
conceal material gaps.

Download the Excel Maturity-Assessment Workbook

A biblical
perspective on knowing the condition

“Be sure you know the condition of your flocks, give careful
attention to your herds.”
— Proverbs
27:23 (NIV)

Responsible stewardship begins with knowing the true condition of
what has been entrusted to us. That requires more than collecting
information. It requires attention, verification and a willingness to
act on what the evidence reveals.

An annual maturity assessment applies the same principle to the
maintenance program itself. The organization should know whether its
controls are functioning, whether people have clear responsibility,
whether decisions are supported by trustworthy evidence and whether
identified weaknesses are being corrected. Honest assessment is not
criticism of the program; it is how the program remains worthy of
trust.

The annual standard

A predictive-maintenance program is mature when it can repeatedly
demonstrate that credible condition evidence reaches the right
decision-maker, produces safe and accountable work, closes through
comparable retesting, survives change and uses verified outcomes to
improve the next decision.

The annual review should answer one practical question:

Can the organization prove that its predictive-maintenance
control loop still works under current people, assets, systems,
suppliers and operating risks?

If the answer is incomplete, the assessment has done its job. It has
made the next improvement visible.

Research references