AI estate glossary
Plain-English definitions of the terms behind governing an AI estate — what each one means, and where to read more.
MCP (Model Context Protocol)
An open protocol that standardizes how AI applications and agents connect to tools, data sources, and each other. Instead of writing a custom integration per tool, an agent speaks MCP to any compliant server, which makes fleets of agents from different vendors composable. MCP is the connective tissue that lets a coordination layer register, address, and govern heterogeneous agents.
A2A (agent-to-agent)
Communication that happens directly between AI agents rather than between an agent and a human. A2A covers task delegation, status updates, and context exchange across agents that may come from different vendors. Without a shared protocol and a shared record, A2A interactions are invisible to the organization that pays for them.
AI FinOps
The practice of applying FinOps discipline — inform, optimize, operate — to AI and LLM spend. It means unifying usage and dollar-cost across providers, attributing spend to the teams that drive it, and operating with budgets, alerts, and anomaly detection. AI FinOps treats model spend as a managed cloud cost rather than a monthly surprise.
Coherence
A measure of how consistently a fleet of agents is working from the same, current understanding of a mission. Coherence drops when agents act on stale context, duplicate each other's work, or write conflicting changes. Tracking it turns "are the agents stepping on each other?" from a feeling into a number you can alert on.
Showback
Reporting each team's share of a shared cost — such as AI spend — without actually moving money between budgets. Showback creates visibility and accountability first, which makes it the usual starting point before an organization commits to chargeback. It only works if the underlying attribution is trusted by the teams being shown their numbers.
Chargeback
Allocating a shared cost back to the consuming teams' actual budgets, so each team pays for the AI usage it incurred. Chargeback demands defensible, auditable attribution and stable historical cost, because the numbers move real money and get disputed. Most organizations run showback first and graduate to chargeback once attribution is trusted.
Cost per mission
The total spend attributable to one multi-agent workflow (a "mission"), reconciled from the per-run costs of every agent step inside it. It answers what an outcome cost — not just what a provider billed — by joining coordination data (which runs made up the mission) with spend data (what each run cost). Cost per mission is the unit economics of agentic work.
Agent handoff
The controlled transfer of a task from one agent to another, carrying enough context that the receiving agent can continue without redoing or contradicting prior work. A clean handoff records what was done, what remains, and which version of the shared context it was based on. Unrecorded handoffs are where multi-agent workflows silently lose work.
Audit trail for agents
An append-only record of every consequential agent action — runs started, context written, handoffs, conflicts, approvals — kept so questions can be answered in retrospect. Because agents act autonomously and at machine speed, the trail must be immutable, searchable by mission and agent, and exportable for auditors and incident reviews. It is the difference between "we think the agent did X" and evidence.
Cost attribution
Mapping raw provider usage and its dollar-cost to the team, product, or cost center that incurred it. Provider invoices report by model and time bucket, so attribution needs a mapping layer — commonly per-connection (each team gets its own key) or per-request tags. Attribution recorded at ingestion, with denormalized historical cost, is what makes showback and chargeback defensible.
Spend anomaly detection
Automatically flagging AI spend that deviates from its expected pattern — typically by comparing each day's cost against a trailing baseline per team, model, or provider. Anomaly detection catches the failure modes budgets miss: a retry loop, a misconfigured agent, or a model swap that quietly multiplies unit cost while staying under the monthly limit.