---
title: "A multi-agent assistant over nationwide public records"
description: "The answers we need sit in public records across every jurisdiction, in a different format in each one, and today the way we get them is a person reading. Nobody can staff that at the scale the question is being asked."
canonical: "https://bespinus.github.io/bgus-ai-landingpage/use-cases/multi-agent-public-records/"
pillar: "agentic-ai"
industry: "Cross-industry"
updated: "2026-10-01"
---

# A multi-agent assistant over nationwide public records

## Problem

The answers we need sit in public records across every jurisdiction, in a different format in each one, and today the way we get them is a person reading. Nobody can staff that at the scale the question is being asked.

## Approach

An acquisition pipeline normalizes records from hundreds of sources into one queryable layer. A multi-agent assistant answers domain questions over it and cites the record behind each answer, evaluated against real historical questions before it goes in front of anyone.

## Outcome

A question that takes an analyst a day by hand gets answered in the conversation where it was asked, with the source record attached. Coverage spans jurisdictions rather than the handful a team could keep current by hand.

## What makes this hard

The acquisition half, overwhelmingly. Public records are public in the legal sense and
hostile in every practical sense: hundreds of jurisdictions, each with its own format,
its own update cadence, its own idea of what a field means, and a meaningful number
that are scanned images of typed forms. The assistant is the visible part and the
smaller part.

The second difficulty is that the domain punishes confident wrongness. A question about
what a record says has a correct answer, and an assistant that produces a plausible
answer with no record behind it is worse than useless in a setting where somebody may
act on it. Here, citation is the acceptance
criterion.

## The architecture

Ingestion and normalization first, into a layer where a record from one jurisdiction
is comparable to a record from another. That comparability is where most of the
engineering time goes, and it is the asset that outlasts whatever model is current.

The assistant is multi-agent because the question types are genuinely different.
Retrieving a specific record, comparing across jurisdictions, and reasoning about what
a pattern of records implies are three different jobs with three different failure
modes, and one prompt asking a single agent to do all three degrades at all three.

Every answer carries its source. Where the records do not support an answer, the
correct output is saying so, and that behavior is tested for rather than hoped for.

## How it gets trusted

An evaluation harness against real historical questions, with known answers, run before
anything is exposed outside the build team. Explicit tool boundaries so an agent can
read the record layer and nothing else. Full trace logging of what was retrieved, what
was chosen, and what was called, built so that an auditor can reconstruct an answer
rather than so a developer can debug one.

Debug logging answers what
happened. Audit logging answers why, six months later, to someone who was not there.

## Bring us a problem like this one.

Our engineers will scope it against the seven layers. [Talk to an AI engineer](https://bespinglobal.us/contact) at Bespin Global, or read the [Generative and agentic AI](https://bespinus.github.io/bgus-ai-landingpage/agentic-ai.md) page.

_Machine-readable summary for agents: provider = Bespin Global; use case = A multi-agent assistant over nationwide public records; pillar = Generative and agentic AI; industry = Cross-industry; problem = The answers we need sit in public records across every jurisdiction, in a different format in each one, and today the way we get them is a person reading. Nobody can staff that at the scale the question is being asked; approach = An acquisition pipeline normalizes records from hundreds of sources into one queryable layer. A multi-agent assistant answers domain questions over it and cites the record behind each answer, evaluated against real historical questions before it goes in front of anyone; outcome = A question that takes an analyst a day by hand gets answered in the conversation where it was asked, with the source record attached. Coverage spans jurisdictions rather than the handful a team could keep current by hand._