AI FinOps Standards

A licensed AI seat assigned to a user with zero activity in the billing period. Typically represents 15–35% of enterprise AI subscriptions.

Business outcome: Industry median ghost rate ~22%; PromptKing flags inactive seats within 90 seconds of connector sync.

Ghost Seat Detection Engine →

The discipline of tracking, optimising, and governing enterprise AI subscription costs across multiple vendors. Applies FinOps financial operations principles to AI seat licences, token consumption, and compliance risk.

Business outcome: Multi-vendor sprawl can hide 30–55% recoverable spend without seat-level telemetry.

AI FinOps dashboard →

The breakdown of AI spend into input tokens, output tokens, and reasoning tokens as separate billing line items. Critical since June 2026 vendor billing changes increased apparent costs 20–45% without usage growth.

Business outcome: Programmatic API lanes often bill separately from interactive chat — model both before June 2026 cliffs.

Exposure calculator →

The percentage of licensed AI seats with meaningful activity above a 30% engagement threshold. Industry average: 61%. PromptKing customers reach 84% within 90 days.

Business outcome: Low utilisation is the leading driver of ghost seat waste in FinOps audits.

KPI Health — USR metric →

A score (0–100) measuring the ratio of context tokens consumed to output tokens produced. High CDL indicates runaway agent loops or over-padded system prompts consuming budget without delivering value.

Business outcome: Scores above 70 often correlate with agent runaway spend before monthly caps trigger.

CDL Analytics — agent compute leaks →

AI tools used by employees without IT approval or budget visibility. Estimated at £47K/year per 100-person team in untracked spend.

Business outcome: Shadow tools bypass procurement — FinOps needs connector-level discovery, not surveys.

Overview — shadow IT signals →

The EU AI Act's list of 8 high-risk AI application areas requiring conformity assessment before deployment. August 2, 2026 is the enforcement deadline.

Business outcome: High-risk systems need inventory, confidence scoring, and audit-ready documentation before enforcement.

EU AI Act compliance hub →

ServiceNow's per-'assist' charge for external AI agents (Claude, Copilot, etc.) accessing data or executing workflows through ServiceNow. Variable billing makes costs difficult to predict.

Business outcome: Per-assist fees stack on top of seat waste — model both with flat-rate FinOps tooling.

Pricing — vs ServiceNow AICT →

The process of converting heterogeneous AI vendor billing units (tokens, credits, Resource Units, flat seats) into a unified per-seat-per-day USD metric for cross-vendor comparison. Without normalization, a $200/month Claude Max 20× seat and a $19/month GitHub Copilot Business seat are not directly comparable.

Business outcome: Enables apples-to-apples ROI comparison across all AI vendors in a single dashboard view.

Seat Intelligence — normalized spend view →

A dedicated credit pool in Anthropic's billing model (effective June 15, 2026) that separates automated API workflow costs from interactive chat usage. Each Claude subscription plan includes a monthly programmatic credit allocation billed at API token rates when exhausted.

Business outcome: Teams running automated Claude workflows need to model programmatic credit exhaustion separately from chat usage — one agentic pipeline can exhaust a month's credit in hours.

Anthropic Credits Tracker →

The process of matching an AI subscription plan tier to actual token consumption. A seat consuming less than 20% of plan capacity for two consecutive periods is a downgrade candidate. Rightsizing recovers budget without any reduction in capability.

Business outcome: Industry median: 35% of AI seats are candidates for downgrade. PromptKing surfaces recommendations with confidence scores and one-click acceptance.

Savings Engine — rightsizing recommendations →

An Anthropic API feature that stores processed portions of a prompt for reuse across requests, offering up to 90% discount on cached input tokens. Most teams leave this discount unclaimed because it requires explicit implementation in API integrations.

Business outcome: Enabling prompt caching on document analysis or code review workloads can reduce Anthropic API costs by 50–90% with no change to output quality.

Anthropic Credits Tracker →

A per-model cost factor in GitHub Copilot's AI Credits billing system (effective June 1, 2026). Claude Opus 4.7 carries a 27× multiplier; GPT-5.4 Mini carries a 0.33× multiplier. A Business plan user ($19/month = ~190 credits) can exhaust their entire monthly allocation with approximately 7 heavy Opus 4.7 sessions.

Business outcome: Model selection within GitHub Copilot directly controls monthly spend. Defaulting to Opus 4.7 for routine tasks is the single largest source of GitHub Credits overage.

GitHub Credits Tracker →

A silent cost increase that occurs when an AI vendor ships a new tokenizer that generates more tokens for identical input text. Anthropic's Opus 4.7 tokenizer generates up to 35% more tokens than 4.6 for the same prompts — with no announcement and no change to listed token prices.

Business outcome: Average cost per request increases by the tokenizer delta without any usage growth. PromptKing detects tokenizer drift by monitoring average tokens-per-request trend lines.

Intelligence Centre — tokenizer drift detection →

In GitHub Copilot's AI Credits model, credits are allocated per individual user and do not pool across team members. A shared automation pipeline drawing from a single user account can exhaust that user's monthly credits while adjacent seats remain unused.

Business outcome: Teams running shared automations under a single Copilot seat face unexpected credit exhaustion mid-month. PromptKing flags non-pooled credit risk automatically.

GitHub Credits — non-pooled risk alerts →

A spend spike caused by an AI agent executing dozens or hundreds of API calls in a single session, each consuming tokens. Unlike conversational AI (one request, one response), agentic workflows can consume a month's token budget in a single hour. Microsoft reported week-over-week GitHub Copilot cost doubles driven entirely by agentic workflows in early 2026.

Business outcome: A single misconfigured agent loop can exhaust monthly budget before cost alerts fire. PromptKing's Compute Density Leak score detects runaway loops before they complete.

Intelligence Centre — agent loop detection →

The total monthly spend attributable to licensed AI seats with zero or near-zero usage. Calculated as: ghost seat count × plan monthly price. Industry median GLW is 22% of total AI subscription spend. PromptKing customers reduce GLW to under 4% within 90 days.

Business outcome: GLW is the most recoverable category of AI spend — 100% savings with zero capability impact.

Seat Intelligence — ghost seat detection →

A measure of how closely projected AI spend matches actual end-of-month spend, expressed as a percentage deviation. Target: ±10% accuracy on a 30-day forecast. Only 20% of finance teams can forecast AI spend within ±10% in 2026 — meaning 80% are effectively guessing.

Business outcome: Finance teams cannot budget for AI if spend is unpredictable. Real-time burn rate data is the prerequisite for forecast accuracy.

CFO Report — spend forecast →

A PromptKing executive report that quantifies the financial exposure of delaying AI FinOps action. Calculates monthly waste accumulation, annualized overspend, and the dollar value of unacted rightsizing recommendations.

Business outcome: The Red Report converts FinOps recommendations into CFO-language risk exposure. One report typically closes procurement approval for PromptKing in the same meeting.

Red Report — cost of inaction →

The highest maturity level of AI FinOps governance, in which AI subscription costs are allocated and billed back to the business units or teams that consumed them. Replaces showback (visible but non-binding) with financial accountability.

Business outcome: Organizations that implement AI chargeback reduce discretionary AI usage by 15–30% within 90 days — not by restricting access, but by making cost visible to decision-makers.

Reports — chargeback allocation →

The FinOps Foundation's open standard for cloud and AI billing data normalization. Version 1.4 (ratified at FinOps X, June 2026) extends the standard to AI token economics — consumption quantities, host providers, and token-based billing. FOCUS-conformant exports plug directly into Tableau, Power BI, and any BI tool without ETL.

Business outcome: PromptKing exports native FOCUS 1.4 conformant datasets across every connected AI vendor, with a 1.3 legacy export retained — enabling direct integration with existing enterprise BI infrastructure.

Reports — FOCUS 1.4 export →

A mandatory ICT asset inventory required under the EU Digital Operational Resilience Act (DORA) for financial institutions. AI systems are ICT assets and must be catalogued with function, criticality, and third-party provider details.

Business outcome: Financial services firms deploying AI must complete DORA Register entries for every AI vendor relationship. PromptKing auto-generates vendor inventory compatible with DORA Register format.

Europe Vertical — DORA Register →

Compliance with Canada's Personal Information Protection and Electronic Documents Act. For AI systems, PIPEDA requires purpose limitation (data used only for stated purpose), consent management, and breach notification within 72 hours of discovery.

Business outcome: Canadian organizations deploying AI must document data flows through each AI vendor. PromptKing's vendor inventory supports PIPEDA data flow mapping.

Security & Compliance →

A machine learning score (0–1) measuring how anomalous a seat's AI spend pattern is relative to its historical baseline. Generated by an isolation forest model trained on an 8-dimensional feature vector: spend slope, token efficiency, model-switch frequency, burst patterns, plan tier, and seat age.

Business outcome: Scores above 0.7 are flagged for review. The isolation forest catches anomalies that z-score thresholds miss — particularly gradual spend drift that never triggers a single-day alert.

Intelligence Centre — ML anomaly scoring →

A privacy-preserving aggregate that requires a minimum number of contributing organizations (k≥3) before any benchmark statistic is returned. Prevents reverse-engineering of individual organization data from aggregate metrics.

Business outcome: PromptKing's industry benchmarks (median seat utilization, P75 spend per seat, ghost rate distributions) are k-anonymized at the database layer — not just the application layer.

Intelligence Centre — industry benchmarks →

The fully-loaded cost per meaningful AI-driven business outcome — a resolved ticket, a completed code review, a processed invoice — normalized across models and seat tiers. AI unit economics measures the real cost to generate one unit of business value from AI investment. Unlike per-token pricing, unit economics accounts for model selection efficiency, prompt quality, retry rates, caching performance, and seat amortization. It is the primary CFO-facing metric for connecting AI spend to business outcomes. PromptKing's Effective Cost Per Task KPI operationalizes AI unit economics at the individual seat level.

Business outcome: Connects AI spend to business outcomes at the seat level — the primary CFO-facing unit economics metric.

Unit Economics — cost per task →

The fully-loaded dollar cost of a single model forward pass, including input tokens, output tokens, cache miss penalties, and per-request vendor surcharges. Cost per inference is the atomic billing unit in API-based AI deployments. In agentic workflows, a single user action may trigger 50–200 inferences sequentially — each billed at the per-token rate. The GitHub Copilot 27× Opus 4.7 credit multiplier is fundamentally a cost-per-inference problem. Monitoring cost-per-inference trajectories is the earliest warning signal for runaway agentic cost escalation.

Business outcome: Per-inference cost trajectories are the earliest warning signal for runaway agentic spend.

GitHub Credits — model multiplier tracking →

The audit process for identifying unauthorized or untracked AI subscriptions used by employees — the AI-era equivalent of shadow IT. Shadow AI occurs when employees subscribe to AI tools personally, expense them informally, or use free tiers outside IT procurement records. In 2026, the average enterprise has 3–5 AI tools deployed officially and an unknown number running outside oversight. Shadow AI discovery audits expense reports, corporate card statements, browser extension inventories, and app store registrations to surface the true organizational AI footprint. PromptKing's seat fleet visibility surfaces shadow AI adjacent to known authorized deployments.

Business outcome: Surfaces unauthorized AI spend adjacent to known seat deployments — closing the shadow IT gap for FinOps.

Seat Fleet — shadow AI visibility →

Dynamically selecting the optimal AI model for each request at inference time, based on task complexity, cost targets, and latency requirements. LLM routing directs each inference request to the most cost-effective model capable of satisfying the quality requirement. Simple tasks route to Haiku 4.5 ($1.00/MTok input; legacy Haiku 3.5 bills $0.80/MTok); standard code generation routes to Sonnet ($3.00/MTok); complex reasoning routes to Opus ($5.00/MTok). Intelligent routing can reduce total inference spend by 40–70% with minimal quality degradation. PromptKing's Token Efficiency Ratio is the primary diagnostic signal for routing misconfiguration.

Business outcome: Intelligent model routing can reduce inference spend 40–70% — Token Efficiency Ratio flags misconfiguration.

Intelligence Centre — routing diagnostics →

A FinOps discipline purpose-built for flat-rate AI subscription billing — seat tiers, credit pools, plan utilization — as distinct from infrastructure-based pay-per-token billing. Traditional cloud FinOps was designed for pay-as-you-go infrastructure. Enterprise AI in 2026 runs on two fundamentally different billing models. Infrastructure-native (Anthropic API, AWS Bedrock, Google Vertex) is pay-per-token. Subscription-native (Claude Pro/Max seats, M365 Copilot, GitHub Copilot) is flat monthly fees with capacity limits. Infrastructure tools like Vantage and Finout cover API-native billing. PromptKing is subscription-native — a different buyer, a different discipline, and a different ROI conversation.

Business outcome: Separates subscription seat economics from infrastructure FinOps — the core PromptKing positioning.

Why PromptKing — subscription-native FinOps →

The estimated CO2 equivalent (CO2e) emissions attributable to an organization's AI inference workloads, measured per model, vendor, and business unit. At millions of monthly inferences, AI compute becomes a reportable Scope 3 emissions source under GHG Protocol accounting. AI carbon footprint tracking maps inference volume to energy consumption to CO2e, enabling sustainability teams to include AI in annual ESG disclosures. Under the EU AI Act (enforcement: August 2, 2026), CSRD requirements in the EU, and evolving SEC climate disclosure guidance, AI carbon accounting is transitioning from voluntary to material for regulated industries. PromptKing's governance suite surfaces AI carbon metrics alongside financial cost data.

Business outcome: Maps AI inference to Scope 3 CO2e for ESG disclosures — financial and environmental FinOps in one view.

ESG AI Impact — carbon per token →

The financial and governance obligations imposed by the EU Artificial Intelligence Act on organizations deploying AI systems. Enforcement of initial obligations began August 2, 2026. The EU AI Act classifies AI systems into four risk tiers: Unacceptable Risk (prohibited), High Risk (mandatory conformity assessments, audit trails, bias monitoring), Limited Risk (transparency obligations), and Minimal Risk (largely unregulated). For FinOps practitioners, the Act introduces direct cost implications: conformity assessment fees, mandatory incident reporting infrastructure, model documentation requirements, and ongoing bias and drift monitoring overhead. Organizations using AI in HR, credit, insurance, or medical diagnosis face the most stringent High Risk obligations. PromptKing's governance scoring, audit log, and bias monitoring features support EU AI Act compliance documentation.

Business outcome: Quantifies compliance cost overhead for high-risk AI systems before August 2026 enforcement.

EU AI Act — compliance hub →

xAI's consumer subscription plan at $30/month providing unlimited Grok 4 access, DeepSearch, Big Brain mode, and Grok Imagine. Distinct from the xAI API which bills per token. SuperGrok seats require utilization tracking — PromptKing surfaces seats below 20% utilization as downgrade candidates on the Grok dashboard.

Business outcome: Separates seat subscription waste from API token spend — typical recoverable SuperGrok waste $30–60/seat/mo.

Grok seat fleet →

xAI charges $5 per 1,000 successful tool calls (Web Search, X Search, Code Execution, Document Search) on top of token costs. A single DeepSearch session can trigger 3–5 tool calls adding $0.015–$0.025 per query. PromptKing tracks tool call costs separately from token costs in the Grok dashboard.

Business outcome: Tool-heavy workflows can add 5–15% to monthly Grok API spend if not routed to lighter models.

Grok tool call breakdown →

A separate key from inference API keys, obtained at console.x.ai → Settings → Management Keys, used to access the xAI Management API (management-api.x.ai). Enables programmatic access to billing history, credit balances, usage deductions, and team management — required for PromptKing's full Grok connector.

Business outcome: Without a Management key, PromptKing can validate models but cannot load credit runway or historical usage.

Grok connector setup →

SuperGrok ($30/month) is xAI's consumer subscription for individuals — unlimited Grok 4 access, DeepSearch, Big Brain mode, Voice, and Grok Imagine. Grok Business ($30/user/month) is the enterprise tier with centralized billing, workspace admin controls, and a data privacy guarantee that training data is excluded. Both cost $30 but bill through different surfaces. PromptKing tracks both plan types under the Grok connector, surfaces seat utilization per user, and flags ghost seats consuming less than 20% of session capacity.

Business outcome: Prevents paying enterprise Business rates for seats that only need consumer SuperGrok — or vice versa.

SuperGrok vs Business comparison →

Grok is the only enterprise AI model with native real-time access to X (formerly Twitter) posts, trends, and social signals via the X Search tool call. Each successful X Search call costs $5 per 1,000 calls — the same rate as Web Search and Code Execution tool calls. A research workflow triggering 3-5 tool calls per query adds $0.015-$0.025 per query on top of token costs. PromptKing tracks X Search, Web Search, and Code Execution call volumes separately in the Grok dashboard, surfacing tool call cost as its own line item distinct from token costs.

Business outcome: Social-signal workflows can double apparent per-query cost if tool calls are not budgeted separately.

X Search tool call tracking →