By John Patzakis
For two years, the enterprise conversation about AI was about capability: What can it do? Is it good enough? In 2026, that conversation changed almost overnight. The question now is the bill. Alexander Embiricos, who leads enterprise at OpenAI, put the shift plainly to TechCrunch in June: “Six months ago, I would have a conversation with a customer, and it would be all about ‘What can it do?’ Our conversations are never about that now. Now the conversations are about, ‘hey, we’re spending so much. What visibility do you have?…What token controls do you have? What is the efficiency of your models?’”
Anthropic — the maker of Claude — identifies inefficient retrieval as a primary culprit in its own engineering writing. In its November 2025 piece on advanced tool use, Anthropic describes the failure mode directly: when an agent fetches records across a data set, “every record accumulates in context regardless of relevance,” and when a large file is retrieved, “the entire file enters its context window.” Independent analysts reach the same conclusion. A June 2026 study measured a 26x per-query token gap between dumping full documents into context and retrieving selectively. A DeployStack analysis traced a routine two-step document workflow that “consumed 120,000 tokens” because a single file passed through the model twice, warning: “Run this 100 times a day across a team, and you’re looking at real money.” And Dennis Pilarinos of Unblocked names the specific culprit without hedging:
“Broad search is the default failure mode, and it’s the most expensive one.”
— Dennis Pilarinos, Unblocked
Enterprise search is the textbook worst case
Nowhere does this dynamic bite harder than search over large, unstructured stores — email inboxes, file shares, chat archives, document repositories. When an employee asks Claude to “find everything about the Henderson matter” or “pull the emails where we discussed pricing,” a single query can drag hundreds of matching messages, long threads, and full attachments into the context window as raw payload. The user pays for all of it, every time — even though only a handful of items actually mattered. The larger the store and the broader the query, the worse the ratio. And broad, unfiltered search over massive stores is exactly how people naturally use these tools.
The fix is architectural: search in place, then bring only what matters
If token cost is driven by data flowing through context, the highest-leverage optimization for enterprise search is not a spending cap — it is a change in architecture. Locate first; retrieve selectively. Search the data where it lives, then bring only the relevant results into the model. A design that returns a ranked result list, targeted snippets, and a pointer to the source document — rather than dumping full inboxes and file bodies into context — attacks the cost problem precisely where the evidence says it lives. This is, notably, the same conclusion Anthropic’s own engineers advocate.
This is exactly what the new X1 Search MCP Connector for Claude does. X1 maintains a local, enriched index over an organization’s actual content — files, emails, attachments, Microsoft 365, Google Workspace, chats — and searches that content in place, on the user’s own machine or behind the corporate firewall. Exposed to Claude through the Model Context Protocol, X1 returns exactly what the model needs and nothing it doesn’t: a compact ranked result list, targeted snippets, and a file location, instead of streaming entire mailboxes and documents through the context window. Claude then reasons over only the specific items that matter, opening full content on demand for the few documents actually under review. Because the matching happens locally, a query across a massive corpus costs roughly the same handful of tokens whether the index holds a thousand items or a million — and because only relevant results ever leave the machine, the data-exposure footprint shrinks at the same time, preserving confidentiality and privilege for legal, compliance, and government teams.
The ROI
The economics are not marginal. At current large-model rates, a single broad search across a year’s worth of emails or a file share can cost roughly $20 in AI tokens when the raw content is streamed into the model — versus a fraction of a cent when only the relevant results are passed to it. Across an organization with thousands of users searching throughout the day, that difference compounds into millions of dollars in avoided token costs each year, while simultaneously improving response speed and reducing data exposure. The savings scale directly with data volume, which means the connector becomes more valuable — not less — as an organization’s data grows.
The takeaway
The token bill has come due, and every fix now being marketed — observability tools, model routers, usage caps — treats the symptom. Each one helps a company watch its spending or throttle it; none of them changes the fact that, by default, enterprise data has to travel to the AI to be searched, at full token price, every single time. The durable answer is to invert that: bring AI to the data, not the data to the AI. Search in place, return only what matters, and let the model reason over the few items that count. That is not a workaround for the cost problem. It is the architecture the evidence — including the AI providers’ own — points to as the way out.
Sources
- Anthropic, Introducing advanced tool use on the Claude Developer Platform (Nov 2025) — anthropic.com/engineering/advanced-tool-use
- Anthropic, Code execution with MCP: building more efficient AI agents (A. Jones & C. Kelly, Nov 2025) — anthropic.com/engineering/code-execution-with-mcp
- TechCrunch, The token bill comes due: Inside the industry scramble to manage AI’s runaway costs (Jun 5, 2026)
- DeployStack, How MCP Servers Use Your Context Window (Jan 2026) — deploystack.io
- Unblocked (D. Pilarinos), Why AI Agents Burn Tokens (Jun 2026) — getunblocked.com
- The Token Tax of Epistemic Accuracy: Comparing RAG and Long-Context Architectures (arXiv, Jun 2026)
Next Gen eDiscovery Law & Tech Blog 