Skip to main content

Application B — Operations plan

Objective

In this application, you will produce a complete operations plan for your AI system — the document that ensures the system stays reliable, trustworthy, and improving after launch day.

If Application A was the launch plan (getting the system into production), this application is the sustainability plan (keeping the system healthy and growing). Together, they form the complete product lifecycle strategy.


Why this matters

Launch day is not the finish line. It is the starting line.

The vast majority of an AI system's life is spent in production — being used, being monitored, encountering problems, being updated, and growing. A system that launches brilliantly but degrades over time is worse than a system that launches modestly but improves continuously, because the first one destroys trust while the second one builds it.

Everyday analogy: think about buying a car. The showroom experience matters — the test drive, the features, the price. But what truly determines your satisfaction is what happens after you drive off the lot. Does the car start reliably every morning? Is the maintenance affordable? Does the dealer respond when something breaks? Can the car handle a road trip, not just a commute? Are there mechanics who know how to fix it?

Your operations plan is the "ownership manual" for your AI system. It tells the team how to keep the system running, how to handle problems, how to grow capacity, and how to evolve the system as the world changes around it.


The assignment

Produce a complete operations plan for the same AI system you used in Application A. This plan covers four areas: monitoring and health, incident response, scaling strategy, and evolution roadmap.


Document structure

Your operations plan should contain the following six sections.

Section 1: System health dashboard (one page)

Define the complete monitoring framework for your system. This is the control panel that operators use to understand system health at a glance.

1a. SLO definitions

Define at least 10 SLOs across these categories:

Infrastructure SLOs:

SLOTargetSLI (how measured)Alert thresholdError budget
Availability[X]% uptimeHealth check endpoint, 30-second intervalsBelow target for 5 minutes[X] minutes/month
Latencyp95 under [X]sResponse time percentile trackingp95 exceeds [X]s for 10 minutes[X] violations/month
Error rateBelow [X]%HTTP 5xx responses / totalAbove [X]% for 5 minutes[X]% budget

Semantic SLOs:

SLOTargetSLI (how measured)Alert thresholdError budget
Classification accuracyAbove [X]%Weekly human-reviewed sample of [X] itemsBelow target in weekly review[X]% budget
Hallucination rateBelow [X]%Automated fact-check + random auditAbove threshold in any audit window[X]% budget
Confidence calibrationWithin [X]%Comparing confidence to actuals weeklyCalibration gap exceeds [X]%[X] points
Safety violation rateZero criticalGuardrail monitoring + incident reportsAny critical violationZero tolerance
Fallback rateBelow [X]%Requests requiring human intervention / totalAbove threshold for 24 hours[X]% budget

Business SLOs:

SLOTargetSLI (how measured)Alert thresholdError budget
Cost per outcomeBelow $[X]Total cost / completed outcomes, dailyExceeds target by 20%[X] budget
User satisfactionAbove [X]/5Post-interaction surveyBelow target in weekly average[X] points
Task completion rateAbove [X]%Outcomes completed without escalationBelow target for 1 week[X]% budget

1b. Dashboard layout

Describe how the information is organized for operators. Think of it like the instrument panel in an airplane cockpit — the most critical information in the center, secondary information on the sides, historical trends below.

Specify:

  • What is visible at a glance (top-level health indicators: green/yellow/red)
  • What requires one click to access (detailed metrics per SLO)
  • What requires deeper investigation (historical trends, drill-down to individual requests)

1c. Alert routing

Define where alerts go based on severity and type:

Alert typeSeverityFirst responderEscalation pathChannel
System downS1On-call engineerEngineering manager, then VPPage + phone
Accuracy below SLOS2On-call engineerDomain expert, then product leadPage + chat
Cost spikeS3On-call engineerEngineering managerChat notification
Drift detectedS4On-call engineerLogged for weekly reviewEmail

Section 2: Ownership model and on-call (one page)

Define who is responsible for what, and how the team is organized for operational support.

2a. Ownership matrix

ComponentPrimary ownerBackupResponsibilities
Infrastructure[Role/Name][Role/Name]Uptime, deployment, capacity
AI model performance[Role/Name][Role/Name]Accuracy, drift, retraining
Domain/ontology[Role/Name][Role/Name]Ontology updates, contract revisions
Evaluation suite[Role/Name][Role/Name]Test maintenance, evaluation runs
Governance/compliance[Role/Name][Role/Name]Guardrails, audits, regulatory
User experience[Role/Name][Role/Name]Feedback, satisfaction, training

2b. On-call rotation

Specify:

  • Rotation schedule (weekly? bi-weekly?)
  • Number of people in the rotation
  • Skills required for on-call duty
  • Handoff protocol (what the outgoing on-call communicates to the incoming)
  • Compensation and sustainability (how you prevent burnout)

2c. Escalation matrix

ScenarioTier 1 (first 15 min)Tier 2 (within 1 hour)Tier 3 (within 4 hours)
System outageOn-call engineerEngineering managerVP Engineering + business stakeholders
Accuracy degradationOn-call engineerDomain expert + ML engineerProduct lead + business stakeholders
Safety violationOn-call engineer + domain expertGovernance lead + legalExecutive team
Cost overrunOn-call engineerEngineering managerFinance + business stakeholders
Data breachOn-call engineer + securityCISO + legalExecutive team + regulators

Section 3: Incident response and learning (one and a half pages)

Define the complete incident lifecycle: detection, response, resolution, and learning.

3a. Incident classification

Provide your complete severity taxonomy with specific criteria for both traditional and semantic failures. Include examples for each severity level to help the on-call engineer classify quickly.

3b. Response procedures

For each severity level, specify:

  • Response time commitment
  • Communication requirements (who is notified, how, how often)
  • Mitigation options available at each level
  • Documentation requirements during the incident

3c. Post-mortem process

Specify:

  • Which incidents require a post-mortem (recommendation: all S1 and S2, selected S3)
  • Timeline for completing the post-mortem (recommendation: within 5 business days of resolution)
  • Post-mortem template (use the template from Lesson 2 or create your own)
  • Review process (who reads and approves the post-mortem)
  • Action item tracking (how you ensure follow-up actions are completed)
  • Learning sharing (how lessons are communicated to the broader team)

3d. Runbook library

List the runbooks your team needs and write at least 2 in detail. Recommended runbooks:

RunbookScenarioPriority
RB-01Complete AI service failure and fallback activationCritical
RB-02Accuracy drop below SLO thresholdCritical
RB-03Safety guardrail triggered at elevated rateCritical
RB-04Model drift detectedHigh
RB-05Hallucination reported by userHigh
RB-06Cost per outcome exceeding budgetMedium
RB-07User satisfaction dropMedium
RB-08Ontology update deploymentMedium

Write at least 2 complete runbooks with: scenario description, symptoms, pre-checks, step-by-step response, escalation criteria, rollback procedure, and communication template.

Section 4: Scaling strategy (one page)

Define how the system will grow in the next 12-18 months.

4a. Growth projections

DimensionCurrent+6 months+12 months+18 months
Daily requests
Active users
Domains served
Languages supported
Data volume (monthly)

4b. Model-tier strategy

Define your routing tiers:

TierModel typeRequest criteriaExpected volumeCost/requestAccuracy target
Tier 1[Type][Criteria][X]% of traffic$[X][X]%
Tier 2[Type][Criteria][X]% of traffic$[X][X]%
Tier 3[Type][Criteria][X]% of traffic$[X][X]%

4c. Bottleneck analysis

For your 12-month growth target, identify the top 5 bottlenecks and your solution for each:

BottleneckWhen it hitsImpactSolutionInvestment needed
[Bottleneck]At [X] requests/day[Impact][Solution][Cost/effort]

4d. Cost projection

Model your cost per completed outcome at each growth stage:

StageAI processingHuman reviewInfrastructureOperations overheadTotal cost/outcome
Current$[X]$[X]$[X]$[X]$[X]
+6 months$[X]$[X]$[X]$[X]$[X]
+12 months$[X]$[X]$[X]$[X]$[X]

Section 5: Evolution roadmap (one page)

Define how the system will evolve beyond just handling more volume.

5a. Ontology evolution plan

Describe anticipated ontology changes over the next 12 months:

ChangeTriggerTimelineImpactMigration approach
[New entity or category][Why needed][When][What is affected][How to migrate]
[Category refinement][Why needed][When][What is affected][How to migrate]

5b. Capability expansion

List new capabilities planned for the next 12 months, with dependencies and readiness criteria:

CapabilityTarget quarterDependenciesReadiness criteria
[Capability][Quarter][What must be true first][How you know it is ready]

5c. Technical debt management

List your top 5 items of semantic technical debt, classified by risk:

Debt itemRisk levelCurrent impactPaydown planTimeline
[Item]High/Med/Low[Current cost][Specific plan][When]

Include a debt budget: what percentage of team capacity is allocated to debt paydown each quarter?

5d. Version management policy

Define your versioning approach:

  • What is versioned (ontology, contracts, evaluation data, guardrails, models)
  • Version numbering scheme
  • Changelog requirements
  • Rollback policy (how long are old versions retained?)
  • Compatibility rules (how many versions back must the system support?)

Section 6: Governance and compliance operations (half page)

Define the ongoing governance practices that keep the system safe and compliant.

6a. Regular audits

Audit typeFrequencyScopeResponsibleOutput
Accuracy auditWeeklyRandom sample of [X] outputs[Role]Accuracy report
Safety auditMonthlyAll guardrail triggers + flagged outputs[Role]Safety report
Bias auditQuarterlyDemographic analysis of outcomes[Role]Fairness report
Compliance auditQuarterlyFull regulatory checklist[Role]Compliance attestation
Cost auditMonthlyPer-outcome cost analysis[Role]Cost report

6b. Governance review cadence

ReviewFrequencyParticipantsAgenda
Operational reviewWeeklyOn-call team + engineering leadSLO status, open incidents, upcoming changes
Product reviewBi-weeklyProduct + engineering + domainFeature performance, user feedback, roadmap progress
Governance reviewMonthlyCompliance + legal + product + engineeringAudit results, policy updates, risk posture
Executive reviewQuarterlyExecutive sponsor + product lead + engineering leadValue evidence, risk summary, investment decisions

Deliverable

A document (6-10 pages) containing all six sections:

  • System health dashboard with 10+ SLOs, alert routing, and dashboard layout
  • Ownership model with responsibility matrix, on-call rotation, and escalation paths
  • Incident response with classification, procedures, post-mortem process, and 2+ complete runbooks
  • Scaling strategy with growth projections, model-tier strategy, bottleneck analysis, and cost projection
  • Evolution roadmap with ontology plan, capability expansion, technical debt register, and version management
  • Governance and compliance operations with audit schedule and review cadence

Evaluation criteria

CriterionWhat "good" looks like
Operational completenessThe plan covers prevention, detection, response, and improvement. No major operational scenario is unaddressed. A team receiving this plan could begin production operations immediately.
SLO rigorSLOs are specific, measurable, and include both traditional and semantic metrics. Error budgets are defined. Alert thresholds are calibrated. Targets are realistic, not aspirational.
Incident preparednessThe incident classification captures both infrastructure and semantic failures. Runbooks are detailed enough for someone unfamiliar with the system to follow at 3 AM. Post-mortem process is blameless and focused on systemic improvement.
Scale realismGrowth projections are grounded in evidence or reasonable assumptions. Bottleneck analysis identifies real constraints, not hypothetical ones. Cost projections include all cost categories, not just API fees.
Evolution planningOntology changes are anticipated with migration plans. Technical debt is acknowledged and budgeted for. Version management is systematic. The system is designed to get better over time, not just bigger.

Tips for success

  • Make it actionable on day one. This is not a strategic document — it is an operations manual. Every section should be immediately usable by the team responsible for running the system.
  • Be specific about numbers. "Good availability" is meaningless. "99.5% uptime with a 3.6-hour error budget per month" is specific. "Fast response" is meaningless. "p95 latency under 2 seconds" is specific. Operations lives and dies by precision.
  • Design for the worst day, not the average day. Your operations plan should handle the day when three things go wrong simultaneously, the lead engineer is on vacation, and a stakeholder is asking for an urgent status update. If the plan only works on a calm Tuesday, it is not a plan.
  • Connect to your earlier work. Your SLOs should reflect the quality metrics from Module 6. Your governance operations should implement the controls from Module 7. Your scaling strategy should honor the architectural decisions from Modules 2-5. This plan is the operational expression of everything you have designed.
  • Include the human side. On-call burnout is real. Runbook fatigue is real. Alert fatigue is real. Your plan should acknowledge these human factors and include sustainability measures: rotation limits, alert tuning to reduce noise, clear escalation paths so no one person carries too much weight.