Case study · Agentic document intelligence for regulated teams

Akurion: ask your documents, get answers you can check

Akurion reads your files, plans its own searches and cites the page behind every statement, so regulated teams and the agents they build get governed answers they can verify.

Visit akurion.ai

Why RAG demos break on real documents, and how Akurion handles itNarrated explainer that starts with why RAG prototypes fail on large, changing document sets, then shows Akurion product screens: connectors, auto-generated metadata keys, chat filters, the Graph RAG relationship map, cited source pages opened in chat and the MCP playground.

Documents shown in the Akurion videos are fictional sample data.

The problem

Why we built it

In insurance, banking, legal, healthcare and manufacturing, the important questions cross documents: which customer orders depend on a lot that failed inspection, which contracts have uncapped liability, which borrowers are close to breaching a covenant. Basic chatbots answer from one passage at a time and can't show their work.

Stack

ReactPythonPostgres + pgvectorKnowledge graphGoogle CloudFrontier and open-weight LLMs

What we built

Agent mode that plans multi-step searches across a project's documents and the web
Graph RAG: entities and relationships extracted so answers can follow links between documents
Page-level citations on every statement, with answer-quality evaluation built in
Vector store management, enrichment and memory across conversations
Workflows that run scheduled, repeatable document reviews and deliver results to a team
MCP server and API so other AI tools and agents can query Akurion directly
Enterprise controls: SSO, project-level access, roles, API keys, audit and usage dashboards

Feature tour

Akurion.ai features in motion

Agent mode: a planned search with every claim citedA question about vendor spend is asked inside one file, the agent's search steps appear, and the answer streams in with a citation on every line.

Chat commands and the answer APITyping @fileName in the chat composer scopes a question to one file, then the same question is sent to the POST /api/v1/queries/answer endpoint, which returns a cited JSON answer.

Chat history: find and continue any threadThe threads panel, grouped by day, is searched for "spend", an older thread reopens with its earlier answer, and a follow-up question gets a new cited answer.

Credit usage and answer feedback dashboardShort silent loop of the credit usage page: totals, the weekly credit trend, credits by operation, weekly thumbs-up and thumbs-down feedback, and remaining credits.

Rating answers and reviewing feedbackA cited answer gets a thumbs-up and a short note, then the Insights page shows the feedback count and positive rate update and the note at the top of recent feedback.

Graph RAG: linking names across documentsThe same supplier and person are highlighted across a contract, a purchase order and an invoice, linked into a relationship map, and a question is answered by following those links with citations.

Connected tools over MCP, with approval for actionsTools are connected in the Connected tools table, the agent searches a logistics tool and cites two late shipments, and a follow-up task is created only after the user approves it.

Memory: standing instructions the assistant keepsA user asks for answers in tables, a later answer arrives as a table, and the instruction appears in the admin Memory panel with Edit and Delete.

Project-level roles and partner accessTwo projects show their members and roles, an outside partner is added to one project as a Member, and the other project stays locked to them.

Self-serve single sign-on (OIDC)OIDC single sign-on is switched on in account settings and a company email domain is allowed, then a teammate signs in with SSO through their identity provider.

Vector search by meaningDocuments dissolve into points that cluster by meaning, and a question in different words lights up the nearest points, returned as three cited passages.

Videos

Akurion.ai walkthroughs and explainers

18 questions Akurion answers across teamsAnimated, narrated montage of 18 example questions answered from documents for claims, legal, credit, healthcare, manufacturing, HR and IT teams, each with a short cited answer. Uses illustrative fictional data.

Akurion Workflows: from one prompt to a reliable processAnimated, narrated explainer of Akurion workflows: a process described in one sentence becomes a plan that uses connected MCP tools, runs now or on a schedule, branches, pauses for human approval and writes tasks and briefs to other tools.

Akurion AI Quality: observe, evaluate and guard answersNarrated explainer mixing animation with product screens: answer traces with steps, sources and timing, automatic faithfulness and relevance scores, test sets generated from your documents, run comparisons and guardrails.

Insurance claims: which open claims hit a flood exclusionAnimated, narrated story: Maya, a claims lead, connects policies and claim files to Akurion and asks which open claims hit a flood exclusion or sublimit. The answer names two claims, each cited to a policy page, using fictional sample data.

Legal: finding uncapped liability across contractsAnimated, narrated story: Marcus connects master agreements, amendments and side letters to Akurion and asks where liability is uncapped. The answer lists two agreements, each cited to the exact page, using fictional sample data.

Banking: spotting borrowers close to a covenant breachAnimated, narrated story: James, a credit analyst, connects compliance certificates and loan files to Akurion and asks which borrowers are close to a covenant breach. The answer flags one breach and one tight borrower, each cited, using fictional sample data.

Healthcare: matching a device recall to equipment recordsAnimated, narrated story: Omar connects recall notices, equipment inventories and work orders to Akurion and asks which infusion pumps are in the IP-300 recall. The answer gives the serial range and departments, each cited, using fictional sample data.

Manufacturing: tracing a bad batch to customer ordersAnimated, narrated story: Tom, a quality engineer, connects production logs, batch records and shipping notes to Akurion and asks which customer orders hold batch B-0915. The answer separates shipped and held orders, each cited, using fictional sample data.

Supply chain: which purchase orders will miss their dateAnimated, narrated story: Mei connects supplier notices, trackers and quality reports to Akurion and asks which open purchase orders will miss their due date. The answer lists two late orders with reasons, each cited, using fictional sample data.

HR: who must sign a new policyAnimated, narrated story: Hannah, an HR business partner, connects policies, rosters and agreements to Akurion and asks who must sign the new on-call policy. The answer names the employees, deadline and managers, each cited, using fictional sample data.

IT support: finding the fix for a VPN resetAnimated, narrated story: Sam connects IT guides, wiki pages and solved tickets to Akurion and asks how to reset VPN access on a new laptop. The answer gives the steps and the approval rule, each cited to a guide page, using fictional sample data.

Want something similar?

Tell us about the process you want to automate.

A 30-minute call is enough to tell whether AI is a fit, what a first project would look like and what it would take to run it in production.

Start a conversation dassjaylabs@gmail.com

Akurion: ask your documents, get answers you can check

Why we built it

Stack

What we built

Akurion.ai features in motion

Akurion.ai walkthroughs and explainers

Services behind this product

RAG & Document AI

AI Agent Development

Business Process Automation

Private & On-Prem LLMs

Tell us about the process you want to automate.