02AI Operations Copilot

An intelligence layer that watches operations, answers in natural language and automates the busywork for a fast-scaling enterprise.

Industry
Enterprise SaaS
Year
2025
Duration
16 weeks
Services
AI Solutions · Web Applications · Business Automation

-71%

Time to answer

Median across the twenty most common operational questions.

3

Tools replaced

Two reporting subscriptions and an internal wiki retired.

94%

Answer accuracy

Scored against the golden evaluation set at release.

<2s

First token

Streamed response latency at the 95th percentile.

The Challenge

Operational knowledge lived in three systems and a decade of documents. Answering a routine account question meant a twenty-minute hunt across tools, and the same three analysts had become a bottleneck for the entire company.

The Approach

Four moves that shaped the build, in the order they mattered.

  1. 01

    Map the questions, not the data

    We logged the questions people actually asked for three weeks, then designed retrieval around those, rather than indexing everything and hoping.

  2. 02

    Ground every answer

    Hybrid search with reranking over a governed corpus, with inline citations so any claim can be checked in one click.

  3. 03

    Give the model real tools

    Typed, permissioned tools for live system queries, so the copilot reads current state instead of guessing from stale documents.

  4. 04

    Score it continuously

    A golden dataset of 400 real questions runs in CI, so no prompt or model change ships without evidence it did not regress.

What shipped

  • Hybrid retrieval with reranking
  • Inline citations on every answer
  • Typed, permissioned tool calls
  • 400-question evaluation suite in CI
  • Full prompt and response audit log
  • Per-team cost attribution

Built with

  • Next.js
  • TypeScript
  • Python
  • pgvector
  • LangGraph
  • Claude API
  • AWS

Every repository, cloud account and piece of documentation was handed over in the client's own organisation, with a walkthrough recording for the internal team.

The platform CITS engineered replaced three internal tools. The automation alone paid for the project within months.

Chief Operating Officer/Enterprise SaaS Company
Let's Talk

Most engagements start with a brief that has more ambition than certainty. Send it over and we'll tell you honestly what we would do.