Home · Work · Case study
Case Study

A Production AI Agent, Used Daily by 150+ People

An enterprise team’s experts were spending their days answering the same questions — questions their own documentation already answered, across thousands of documents nobody could search effectively. The fix was an agent. The hard part was making it one that security would approve and users would trust.

IndustryEnterprise services
ServicesAI Agents & Assistants, Knowledge Systems
EngagementDesign & build, then optimization
150+daily users in production
10xdata-operations productivity
90%lower token cost after optimization
Overview

The situation.

The organization’s knowledge lived in thousands of internal documents — policies, procedures, technical references — spread across systems accumulated over years. A public chatbot was ruled out on data-security grounds; an internal search tool existed and was quietly ignored, because finding the document is only half the problem: someone still has to read it, judge it, and answer. Data operations staff had become the human search engine, fielding repeat questions at expert salaries.

Result

Where it landed.

The agent moved from pilot to a daily tool for more than 150 people, and data operations measured a tenfold productivity gain — the queue of repeat questions collapsed, and expert time went back to expert work. Just as important: it stayed up, stayed accurate, and stayed affordable as usage grew, because evaluation and cost controls were part of the original design rather than a rescue mission.

Our Solution

What we built.

  1. Document-grounded retrieval, with citations

    The agent answers only from the internal corpus, and every answer carries its source. Verification takes seconds, which is why users stayed.

  2. PII sanitization layer

    Sensitive content is detected and sanitized before anything reaches the model — the control that carried the design through enterprise security review.

  3. Specialist sub-agents

    Rather than one generalist, the system routes to domain-specialized agents — each with its own corpus slice, tools, and evaluation set.

  4. Evaluation suite & monitoring

    A golden set of real questions, scored on every change; production monitoring on answer quality, latency, and cost from day one.

  5. Cost engineering

    Context caching and compaction designed in early — the groundwork for the later 90% token reduction (see the companion case).

Key Achievements

The numbers.

150+users relying on it daily
10xproductivity gain for data operations
0security findings in review
90%token reduction in follow-on optimization

Start with a fit call.

Thirty minutes. Bring the problem as it actually is — the stalled pilot, the bill, the spreadsheet everyone quietly maintains. We’ll tell you honestly whether we’re the right builder for it.

You leave with a straight answer on fit — and if it fits, a free Blueprint to follow.