“Al-Daftaran”: model links cloud invoices with agent transactions and cuts peak costs by 61 percent

Enterprises that run multi-agent AI systems across clouds and various language models face a silent accounting dilemma, as finance departments receive aggregated consumption invoices at the level of cloud storage units and monthly billing periods, without any link to operational transaction limits. The answer to the core business question, what it costs to process a specific request or settle a particular insurance claim through smart agents, remains completely absent from traditional billing data because cloud providers do not see the internal transaction boundaries within the organization.

Even with the adoption of the open-source cost-and-usage specification known as FOCUS, which standardises billing schemas across providers, the data remain trapped at the overall financial settlement level, despite the addition of consumption attribution attributes and call identifiers. To address this structural mismatch between billing level and transaction level, researcher Fefik Asharia proposed, in a study published on the ArchiVe platform, an accounting model based on a “dual-ledger” architecture aimed at matching instantaneous consumption with official invoices with extreme precision.

The solution separates accounting into two synchronized ledgers: a settlement ledger that conforms to FOCUS specifications, and an attribution ledger that records every instantaneous call using the Open Telemetry standards for generative AI.A multi-layer backbone for inference and linking connects the two ledgers, relying on trace context, inter-agent communication data via protocols such as the Model Context Protocol (MCP), matching of operational keys, and probabilistic linking. This enables each call to be attributed to its specific corporate transaction, recording the linking mechanism and confidence score, while a multi-version pricing table evaluates cost in real time.

Experimental simulation results demonstrated high system efficiency, with the linking column maintaining accounting coverage above 99 % even in cases where the entire trace context propagation was interrupted, thanks to alternative matching layers. The remaining unexplained portion of periodic matching recovered unmeasured spend embedded in the experiment with a deviation of no more than 0.24 percentage points, while also detecting billing errors from providers and uncovering untracked AI systems.Instant objection gates during execution also succeeded in lowering anomalous costs for the 99th percentile by up to 61 %, with a false-stop margin that did not exceed two to four percent.Based on these results, the study put forward the attribution ledger as a standard specification file at the call level for formal adoption within the international FOCUS standard.

This transformation carries a decisive operational impact for engineering and finance teams in enterprises, banks and telecom operators across the Gulf, Egypt and the region, which are racing to build automation agents for customer service, credit and supply-chain functions. Moving from receiving blind cloud invoices at month-end to measuring the profitability of each individual transaction completely changes oversight rules; instead of absorbing budget shocks caused by agents looping in excessive inference cycles and repeatedly calling costly models, transaction-level accounting allows immediate mechanical brakes that protect profit margins. It also gives technology leaders a precise tool to detect parallel systems operating outside official governance gateways and to assess the actual viability of each automated workflow before scaling its deployment.