Anthropic began charging Claude Enterprise subscribers full compute costs in April 2026, and GitHub shifted Copilot to usage-based billing weeks later. The shift to token-based pricing exposes inefficiency in enterprise AI systems that was previously absorbed by flat-rate pricing. Agentic AI costs have turned into an architectural question: When an agent answers a query, how many places does it have to look, and what does it carry back?
A federated agent query spanning Salesforce, Snowflake, and ServiceNow consumed 183,541 tokens at $0.596 per query in a recent CData benchmark. The same question, routed through a scoped and pre-joined data layer, consumed 4,427 tokens at $0.027 — a 97.6% reduction.
The model was held constant across all 56 test runs: Claude Sonnet 4.6 at temperature zero. Only the data architecture changed. CData’s results come from structured systems, but the same principle exists in unstructured data retrieval.
What the Benchmark Measured
CData connected an AI agent to three enterprise systems through MCP and ran a realistic prompt asking for open support tickets, related accounts, and product telemetry. The baseline path required 22 tool calls. The agent traversed the catalog, retrieved schemas, enumerated columns, ran three separate queries, then synthesized an answer.
Four categories of overhead drove that cost.
- Tool definition schemas loaded into the request before the conversation began
- Each retrieval pass added context blocks that re-sent on every subsequent turn
- A default Salesforce Account tool exposed more than 70 fields when the query only needed six
- Discovery started from scratch on every run, paying again for exploration
The scale of tool definition overhead alone is substantial. Anthropic’s engineering team measured a five-server MCP setup of 58 tools consuming roughly 55,000 tokens before a single message was sent.
Where Token Cost Concentrates in Unstructured Data
Structured queries return rows while unstructured queries return documents, threads, and conversations, and the retrieval payload dominates the bill.
Consider what an agent receives when it retrieves an email thread. Headers, signatures, legal disclaimers, and quoted replies stacked from every prior message. The answer the agent was looking for only occupies a fraction of the input tokens.
Several cost drivers appear only on the unstructured side:
- Duplicate content: the same attachment lives in a mailbox, a SharePoint library, and a Teams channel, and an agent querying all three pays for it three times
- ROT content: redundant, obsolete, and trivial material accumulates in file shares and mailboxes at volumes structured tables rarely reach
- Whole-file transfer: agents receive complete documents when passages would answer the query
- Format overhead: formatting metadata, embedded objects, and structural noise inflate counts well beyond the readable content
The Silo Multiplier
Over 80% of enterprise data sits in unstructured formats spread across email platforms, document repositories, chat tools, and collaboration systems. An agent assembling a complete answer has to reach each one separately.
Every connector adds its own tool definitions to the request, and every source runs its own discovery sequence. Each retrieval returns content that overlaps with the other sources, and the redundancy piles up inside the context window.
Agentic workflows compound the cost further, since token consumption in agent loops grows exponentially and one early misstep wastes tokens downstream.
An Illustrative Model
The following calculation uses public inputs only: Anthropic list pricing at $3.00 per million input tokens, and the standard ratio of roughly 135 tokens per 100 words.
Picture an agent answering one customer-history question that touches four repositories.
- Four-connector path: tool definitions across four MCP servers, ~20,000 tokens. Twelve email threads at 1,200 words, ~19,000 tokens. Six documents at 3,000 words, ~24,000 tokens. Two hundred chat messages, ~11,000 tokens. Accumulated round trips, ~15,000 tokens. Total: ~89,000 tokens, $0.27 per query.
- Unified index path: one tool definition, ~5,000 tokens. Twenty retrieved passages at 150 words each, ~4,000 tokens. Duplicates removed from the corpus before retrieval, ROT excluded at the governance layer. Total: ~9,000 tokens, $0.03 per query.
- At 100,000 queries per month: roughly $27,000 against roughly $3,000.
The illustration produces an 89% reduction.
What a Unified Index Changes
In-place data management extracts and indexes the essence of every item, its metadata and full content, while original files remain in their source systems. Email stays in the email platform. Documents stay in SharePoint. Chat history stays in Teams and Slack.
The agent connects once. One tool definition, one query path, no cross-source orchestration.
The reduction comes from four mechanisms:
- Deduplication resolves identical content across repositories to a single indexed instance
- Classification and ROT remediation remove low-value material before it reaches any pipeline
- Passage-level retrieval returns the relevant excerpt with its metadata rather than the source file
- Policy enforcement keeps regulated content out of agent context without a separate filtering step
Each mechanism removes a category of token-inflating overhead.
Accuracy as the Second Return
Relevant information competes with irrelevant information across the context window, and the ratio determines the quality of the output. CData’s scoped path answered correctly on the first attempt across every run.
Curated context improves reliability alongside cost. Research on model degradation shows that sustained exposure to redundant and contradictory content measurably weakens reasoning, and enterprise repositories carry that material in bulk.
Governance decisions made at the data layer set both the cost structure and the accuracy ceiling for every agent built above it.
