Base Annual SalaryBLS 2023 median for knowledge workers in tech/finance: $75K–$110K. Default $85K is blended midpoint. Mercer Compensation Survey 2023.

Employer BenefitsWorldatWork 2023: median employer benefits = 28–33% of base. Includes health, 401k, dental, vision, disability, life.

Employer Payroll TaxFICA 7.65% + FUTA ~0.6% + state unemployment ~0.3%. Total employer burden: 8.5%. IRS 2024.

Indirect OverheadOffice space, hardware, software licenses, IT support, HR administration. SHRM 2023 median: 20–30% of salary.

Training & DevelopmentATD 2023 State of the Industry: average employer spent $1,280/employee on direct training. Default $1,500 adds onboarding amortization.

Voluntary TurnoverBLS JOLTS 2023: professional/tech sector median = 13–18%. Default 15%.

Turnover Replacement CostBoushey & Glynn (CAP/HBR): replacement = 16%–213% of annual salary. 50% is mid-level knowledge worker median.

Supervisory CostSpan-of-control: managing each FTE consumes ~10% of a manager's time. McKinsey Org Design 2022.

API Pricing (per 1M tokens)Prices as of Q1 2025. Input/output tokens priced separately. Note: agentic mode dramatically amplifies token costs — model selection is the highest-leverage cost decision in agentic deployments.

Daily Task VolumeDistinct workflow invocations per business day. In agentic mode, each task spawns multiple API calls per the loop multiplier above.

Base Input TokensSystem prompt + user message + initial context. In agentic mode this is multiplied by the context growth factor across loop iterations.

Base Output TokensGenerated response per call. In agentic mode, each loop iteration generates output tokens independently.

Residual Human CostMcKinsey 2023: most deployments require 15–25% residual human oversight for review and exception handling.

Error & Rework OverheadStanford HELM 2024: GPT-4 class models have 3–8% error rates on structured tasks. Agentic tasks have higher rates (see Module 6).

Platform OverheadLangChain/LlamaIndex, API gateway, rate limiting, prompt management. In agentic deployments, orchestration middleware costs are significantly higher — consider $500–$1,500/mo.

Peak ConcurrencyMaximum simultaneous users/processes hitting the AI system. At low concurrency (<10), standard pay-per-token is cost-optimal. At medium concurrency (10–50), rate limits and latency SLAs require architectural changes. At high concurrency (50+), Azure PTUs or dedicated capacity becomes necessary.

Provisioned Throughput UnitsAzure OpenAI PTU pricing (2024): ~$2,000–$4,000/month per 100 PTU. AWS Bedrock Reserved Capacity: similar structure. At 50+ concurrent users on GPT-4o class models, ~200 PTUs required. This replaces per-token billing with a semi-fixed monthly cost.

SLA Breach CostAt high concurrency without provisioned capacity, inference latency degrades. If the AI system is customer-facing, SLA breaches trigger penalties or customer churn costs. Internal tools: productivity loss from waiting on slow completions.

Rate Limit OverheadExponential backoff, request queuing, retry logic for 429 errors at scale. Increases effective latency and requires additional orchestration engineering.

GPU CountH100 SXM5 is current LLM inference standard. Llama 70B: ~2 H100s for inference. Production HA deployment: 4–8 GPUs minimum.

H100 Market PriceH100 SXM5: $25K–$35K list; often $30K+ with supply constraints 2024. H100 PCIe: ~$25K. A100: ~$10K.

Depreciation ScheduleIRS MACRS: 3-year for computing equipment. Gartner recommends 3-year for AI hardware due to rapid capability advancement.

Power ConsumptionH100 TDP: 700W. At 24/7 × $0.10/kWh: ~$600/mo. At 50% utilization + datacenter PUE 1.5: ~$3,500/yr.

Cooling & FacilitiesLiquid-cooled H100s require specialized rack infrastructure. Amortized datacenter space + cooling: $1,000–$1,500/GPU/year.

MLOps / Platform EngineeringSenior ML engineer at $150K fully loaded. 0.25 FTE = $37,500/yr. Minimum viable allocation for a production on-prem LLM system.

ComplianceSOC 2 Type II for AI: $15–40K/yr. HIPAA/GDPR AI audits: $8–20K. Default $8K for basic compliance posture.

AI Workload HostingGPU-backed inference endpoints, API servers, orchestration. AWS/Azure/GCP 2024 AI workload pricing. Small = dev/test; Large = production; Enterprise = HA multi-region.

DB & RAG InfrastructureManaged PostgreSQL + pgvector, or Pinecone/Weaviate. RAG workloads are storage-intensive. $400/mo covers managed DB + modest vector store for production.

ObservabilityDatadog/New Relic/Grafana Cloud for AI workloads. Includes LLM-specific monitoring: token usage, latency p99, hallucination detection. 2–3× standard SaaS cost. Gartner 2024.

Snowflake / Databricks ComputeDaily or hourly ETL jobs that extract, clean, chunk, and embed documents into the vector store. Snowflake credits for AI data prep: $200–$2,000/mo depending on data volume. Databricks SQL compute: similar range. Default $600/mo for a mid-scale RAG knowledge base.

Vector Re-embedding CostWhen source documents change, all affected chunks must be re-embedded. At OpenAI text-embedding-3-small ($0.02/1M tokens): 10M tokens/month ≈ $0.20. But Cohere Embed Enterprise / custom embedding models add significant compute cost. Default $150/mo for a continuously-updated knowledge base.

Pipeline OrchestrationAirflow (MWAA), Prefect, or Dagster for scheduling data ingestion pipelines. Includes scheduler infrastructure, alert management, and pipeline monitoring.

Data QA Engineering CostSomeone must validate that the AI's knowledge base is accurate, complete, and not corrupted by bad ingestion. Even with automation, 10–20 hours/month of a data engineer's time (~$100/hr) is required for a production RAG system. Default: 10 hrs × $100 = $1,000/mo.

Stale Data Risk CostAnnualized cost of incidents caused by the AI working from outdated or incorrect knowledge. Examples: quoting discontinued products, citing superseded policies, referencing wrong pricing. Conservative $5,000/yr for internal use; $25,000+ for customer-facing deployments.

Data GovernanceAs source systems evolve (schema changes, new data sources, PII handling requirements), the data pipeline requires maintenance. Annual engineering cost for pipeline schema migrations and governance updates.

AI-Specific SecurityPrompt injection detection, output filtering, model access controls. OWASP LLM Top 10 mitigations: Lakera Guard, Rebuff, LLM Guard. Default $500/mo.

DLP for AI I/OSensitive data detection in prompts/outputs (PII, PHI, credentials). Cloud API deployments: data may be sent to third-party model providers. OWASP LLM Top 10 #2.

Zero-Trust IAM for AI AgentsService account governance, API key rotation, least-privilege tool permissions, audit logging. Note: identity policies alone are insufficient for autonomous agents — see Network Microsegmentation below (MITRE ATLAS AML.T0047).

AI Red-TeamingMITRE ATLAS documents AI-specific attacks (adversarial ML, model inversion, data poisoning). AI red-teaming requires skills distinct from traditional pentesting. $15K/yr = one annual engagement.

AI Regulatory ComplianceGDPR Article 22, CCPA AI disclosures, HIPAA BAAs for AI vendors. EU AI Act compliance adds $20–100K for high-risk systems starting 2025.

Insurance Premium UpliftMarsh & McLennan 2023: AI API deployments face 15–40% increase in cyber insurance premiums. Default $8K ≈ 20% uplift on $40K base premium.

AI Breach IRAI breach response requires expertise in model forensics, prompt log analysis, vector store audit. Mandiant/CrowdStrike retainers: $15–30K. Default $10K minimum viable coverage.

East-West Traffic Control + Runtime Kill SwitchRestricts unauthorized inter-workload communications to limit "blast radius" if an agent is compromised. Includes infrastructure to immediately halt automated lateral movement at the network layer. An agent on EC2 with broad IAM can still traverse VPCs without network-layer controls. Tools: Aviatrix, Illumio, Zscaler Workload Segmentation. Source: MITRE ATLAS AML.T0008 (Network Service Scanning), NIST SP 800-204.

Agentic Loop Behavioral MonitoringStandard DLP and static known-bad signatures fail against agentic lateral movement because agents hypothesize alternative paths when blocked — mimicking a skilled human attacker. This line items funds behavioral anomaly detection that monitors shifting, unpredictable execution loops rather than fixed attack sequences. Tools: Darktrace, Vectra AI, Exabeam for agentic workloads, or custom SIEM rules with ML-based anomaly scoring. Source: MITRE ATLAS AML.T0043 (Craft Adversarial Data).

HashiCorp Vault / AWS Secrets Manager for AgentsAgentic architectures change how secrets are exposed: sensitivity is not always visible at deployment — it emerges through use. Agents accumulate credentials, API keys, approval patterns, and infrastructure decisions in persistent memory. Storing these in environment variables or local files creates a high-value target. This funds short-lived, automatically rotated credentials via a secrets manager. Especially critical for agents executing financial transactions or accessing production databases. Tools: HashiCorp Vault ($200–$500/mo), AWS Secrets Manager, Azure Key Vault. Source: OWASP LLM Top 10 #2 (Insecure Output Handling), OWASP LLM #9 (Overreliance).

Third-Party Skill / Tool VettingAgent skills are often plain-text Markdown or JSON files with no CVEs, no container image scanning, and no SBOMs. A skill designed to exfiltrate environment variables looks like a harmless text document to standard CI/CD scanners. Prompt injections routed through a pluggable tool can trigger unauthorized actions on real infrastructure. This funds continuous auditing and vetting of third-party agent tools: static analysis, sandboxed execution testing, and behavioral review. Tools: custom review pipeline, Snyk (extended to AI skills), manual security review retainer. Source: MITRE ATLAS AML.T0047 (Acquire ML Artifacts), OWASP LLM #3 (Training Data Poisoning analogue for skills).

Agentic Task Failure RateAnthropic & DeepMind: complex multi-step agentic tasks fail 5–15%. WebArena (2023): best agents complete ~14% of complex tasks. Default 8% conservative for business workflows.

MTTD for Silent FailuresGoogle SRE benchmark: MTTD for silent AI failures = 2–8 hours without dedicated monitoring. Default 2 hours. This lag allows bad outputs to propagate downstream.

Cascade ProportionOf the processes that run during the detection lag window, what % are actually impacted by the upstream failure? A single failure rarely corrupts all downstream work. Typical range: 10–30%. Use 50–100% only for tightly-coupled sequential pipelines (e.g., financial data feeds).

Hallucination CostBusiness decisions made on incorrect AI outputs. Finance: bad model inputs. Legal: incorrect citations. Medical: wrong references. Default $20K conservative; adjust to $100K+ for regulated industries.

HeadcountFTE positions this deployment replaces or significantly augments. Scales the employee cost side. Note: partial augmentation should use fractional values or adjust oversight % in Module 2.

NPV Discount RateStandard corporate hurdle rate for IT investments. WACC for US large-cap tech: 8–10%. Use 8% for typical internal ROI analysis.

Volume MultiplierMcKinsey MGI 2023: AI increases throughput 2–4× for knowledge work. Used in cost-per-unit-output calculation only; does NOT affect cost totals.

Risk AdjustmentCost buffer for operational, regulatory, and reputational risks not fully captured above. 5% conservative for greenfield; 10–15% for regulated industries.

Technical IntegrationData prep, API integration, pipeline development, testing. McKinsey 2023: average AI deployment integration = $25K–$250K depending on complexity. Default $15K for a mid-scale internal deployment.

Change Management / Process RedesignConsulting hours, process mapping, SOP rewriting, stakeholder alignment. Often the most underestimated cost. Gartner: 30–50% of AI project failures are due to change management failures, not technical issues.

Productivity J-Curve DurationThe period during which employee productivity is below baseline due to learning new workflows, parallel-running old processes, and debugging AI outputs. Typically 2–4 months. During this window, you pay BOTH the employee cost AND the AI cost without full benefit.

Productivity Dip %How much below baseline employee productivity falls during the transition. A 15% dip means the employee delivers 85% of their normal output. Applied as an additional cost penalty: dip% × employee_cost × (dip_months/12).