Two-Ledger Model Attributes Multi-Agent AI Costs to Business Transactions

Summary:Enterprise AI billing is typically aggregated at the SKU and billing-period level, making it difficult to determine what an individual order, claim, or case actually cost across models and agents. The paper proposes a Two-Ledger Model that keeps a FOCUS-conformant settlement ledger and joins it with an invocation-grain attribution ledger built from OpenTelemetry GenAI telemetry. Trace context, A2A/MCP metadata, business-key correlation, and probabilistic matching form the correlation spine. A v

---

Core Idea: Separate Settlement from Invocation-Level Attribution

Enterprise AI bills are generally aggregated by SKU, model or service, and billing period. That level of detail cannot answer the operational question of what an individual order, claim, or case actually cost across models, agents, and tools.

The Two-Ledger Model maintains two complementary sources of truth:

• A settlement ledger containing FOCUS-conformant provider billing data used to verify actual charges.
• An attribution ledger containing invocation-grain usage records collected through OpenTelemetry GenAI telemetry and mapped to business transactions.

Telemetry does not replace the invoice. It creates an independently estimable cost trail that can be reconciled against the settlement ledger.

Correlation Spine: Four Layers Across an Agent Workflow

A single order may trigger multiple agents, model calls, and MCP or A2A tool interactions. Providers do not possess enterprise transaction boundaries, so request identifiers alone are insufficient. The proposed correlation spine combines:

1. Trace context for relationships within distributed execution chains.
2. A2A/MCP metadata for agent-to-agent and tool-invocation relationships.
3. Business-key correlation using order IDs, claim IDs, and other deterministic identifiers.
4. Probabilistic matching when structured identifiers are missing, with the method and confidence recorded for every decision.

The engineering objective is not to make one layer perfect. It is to preserve evidence and confidence so that finance teams can distinguish deterministic attribution, heuristic attribution, and unexplained residuals. Production implementations should avoid exposing sensitive business keys directly to models or cross-domain telemetry; tokenized mappings and minimal attribute propagation are safer defaults.

Pricing, Reconciliation, and Shadow AI

A versioned rate card converts model usage, token consumption, tool costs, and related services into near-real-time estimates. Each valuation must bind the applicable rate version, effective date, and currency so that later price changes do not rewrite historical cost.

At the end of a billing period, the attribution ledger is reconciled against the FOCUS settlement ledger and invoice. Differences may arise from rounding, taxes, discounts, missing attributes, or provider latency. A persistent unexplained residual can instead indicate shadow AI—unregistered model or tool usage—or provider billing errors such as missing charges, duplicates, or incorrect rates.

An unexplained residual should not be treated as proof of misuse. It is an investigation signal that must be checked against contracts, invoice details, and business records.

Simulation Results and Technical Boundaries

In a simulation with known ground truth, the correlation spine maintained cost coverage above 99% even when trace propagation fell to zero. The reconciliation residual recovered planted unmetered spend within 0.24 percentage points and flagged provider billing errors. Mid-flight gating reduced p99 transaction cost by up to 61%, with a false-pause rate of 2–4%.

These findings support the approach, but they do not establish production validity. The abstract does not disclose real enterprise deployments, cross-provider billing samples, telemetry sampling rates, false-match or missed-match rates for probabilistic correlation, or the operational cost of maintaining the rate card. The meaning of the 99% coverage metric also needs scrutiny: it must be clear whether the denominator represents invocation counts, transaction counts, or monetary cost, and whether weakly correlated calls are being counted as attributed merely to preserve coverage.

Engineering Trade-offs and Deployment Guidance

The two-ledger approach moves AI FinOps from retrospective provider-spend analysis toward transaction- and outcome-level cost control. It also introduces three new costs: propagating trace context across orchestration layers, standardizing telemetry semantics, and maintaining auditable business-correlation services. If an agent framework cannot propagate structured context, the trace layer degrades quickly. Differences in provider token-accounting conventions also require model-specific normalization rules beyond the FOCUS billing schema.

A practical rollout should proceed in stages:

1. Adopt the OpenTelemetry GenAI semantic conventions for model, token, requester, agent, and tool-call fields.
2. Map order and claim identifiers through enterprise-side hashes or tokens instead of writing sensitive values into general-purpose telemetry.
3. Report coverage and confidence separately for deterministic tracing, business-key correlation, and probabilistic matching.
4. Treat shadow-AI detection as a ledger-difference alert rather than automatic evidence of misconduct.
5. Pilot mid-flight gating on high-value workflows while independently measuring false pauses, task completion, and p99 cost.

The paper proposes an invocation-grain FOCUS profile for the attribution ledger. If adopted across cloud providers, model platforms, and orchestration frameworks, it could shift AI FinOps from a provider-billing center to a business-outcome cost center. Its success depends on stable call semantics and consistently propagated business-correlation fields across the entire stack.

---

Source: ArXiv AI (cs.AI) (2026-11-06)

— END —

Tags: #Multi-Agent Cost Governance Follow #OpenTelemetry Follow #FOCUS Follow #Observability Follow #FinOps Follow