The Query That Takes 47 Minutes
A question your CFO asks in one sentence requires your analyst to open nine tabs, wait four refreshes, and paste three times. An agent should resolve it in seconds. Here is how.
The Sentence and the Stack
Your CFO asks, in a one-line Slack message: "Q3 revenue by region versus plan?" The sentence is eight words. The answer lives in four systems. Extracting it takes 47 minutes, three tabs, two exports to CSV, one pivot table, and a manual override for the currency conversion that isn't quite right in the reporting warehouse.
That 47 minutes is the analyst's entire morning. The question itself does not require 47 minutes of thought. It requires 47 minutes of retrieval. The thought is three seconds.
The expensive part of most questions is not answering them. It is fetching what you need to answer them.
The Retrieval Problem
This is the retrieval problem. It is the single largest hidden cost in most mid-sized companies, and it is invisible because it is distributed across every analyst, ops person, and manager. Nobody has 47 minutes of the problem on their calendar. Everyone has 47 minutes of it in their day.
An agent that solves retrieval does not replace the analyst. It eliminates the fetch work so the analyst can spend their morning on questions that actually require thought. This is the part of "AI at work" that is real. The rest is marketing.
[Same question. Same source data. Different architecture. The agent joins four systems, cites every number, and serves the result in sub-second time.]
The Three Requirements
To retire the 47-minute query, three things have to be true. The ledger must be complete, every system of record plugged in, every cell reachable. The orchestration must know where to look, it must have a map of the data, not a dump of it. And the output must cite, every number in the answer must link back to a row in the source.
The citation requirement is non-negotiable. An agent that produces uncited numbers is a hallucination machine. An agent that cites is a research assistant. The difference is architectural, and it is the difference between production and demo.
[Every query walks the stack. The outer layer receives, the middle retrieves, the inner cites. No layer can lie.]
What Changes When the Query Is Free
When retrieval costs approach zero, behavior changes. Leaders ask more questions. Analysts do more analysis. Decisions happen on Monday that used to happen the following Monday. The company's clock speed increases, not because the work got harder, but because the waiting got shorter.
This is the highest-leverage install The Forge does. It is also the one most clients do not ask for by name. They ask for "AI." What they need is the 47-minute query, retired. The architecture to retire it is what we build.
From reading to installing.
Field Notes diagnose the friction. The Sprint and the Install eliminate it.