Case studyEnterprise2023-2025
Cut quote errors 85% with an AI quote review agent
The review logic was knowable and rule-shaped. The failure mode was human attention at scale.
- 40+hours a week returned to the team
- 85%fewer quote errors, in production
- 0new tools the team had to learn to adopt it
- Context
- Inbound quotes carry cost, scope, and compliance risk, and review capacity caps how fast any team can move.
- What I built
- A Microsoft Copilot Studio agent, backed by Azure AI Language, that reviews quotes end to end and flags what a human needs to see.
- Headline result
- 40+ hours a week returned to the team and 85% fewer quote errors, in production.
Specification
- Role
- Design, build, rollout
- Timeline
- 2023 to 2025
- Stack
- Microsoft Copilot Studio, Azure AI Language
- Status
- In production
The problem
A quote is a promise with three ways to be wrong: cost against the rules that govern it, scope against the work actually being done, compliance against the standards the work answers to. Checking that at volume is expensive, and an error that slips through costs twice: rework, and credibility with the partners the quotes serve.
The review logic itself was knowable and rule-shaped. The solution had to run inside the tools the team already used and hold up in production, because a quote reviewer that is only mostly right creates more checking work than it removes.
The approach
The agent is built in Microsoft Copilot Studio and lives inside the Microsoft 365 environment the team already worked in. I chose the platform the work already lived on over a custom-coded agent, because adoption then required no new tool, no new login, and no engineering allocation.
The review runs as explicit, auditable checks on cost, scope, and compliance rather than one free-form prompt. Quote review is a compliance task: each rule stays individually testable, and a failed check names the rule that failed.
Azure AI Language handles the language layer, reading quote content and mapping it to the checks.
I designed and deployed it to production myself, end to end, which kept the loop between feedback and agent behavior at days, not quarters.
- 01
Quote in
The quote arrives in the environment the team already works in.
- 02
Explicit checks
Cost, scope, and compliance run as separate, individually testable rules.
- 03
Named failure
A failed check names the rule, so a human sees the reason and not just a flag.
- 04
Human review
Only the flagged quotes reach a person. The rest pass straight through.
A quote reviewer that is only mostly right creates more checking work than it removes.
The outcome
The agent returned 40+ hours a week to the team and reduced quote errors by 85%. The freed hours went back into the customer-facing work the team was hired for, and the error reduction protected the partner relationships that quote accuracy underwrites. It stayed in production in daily use, which is the adoption signal I trust most.
What I would do differently
Build the error taxonomy first, so the 85% could be decomposed by rule and the remaining errors would have a named shape. Write the maintenance handoff documentation at deploy time instead of after.
Some specifics are abstracted for confidentiality. I am glad to go deeper in conversation.