VIENJA
Get a quoteFree 30-min call
Back to case studies
AI ProjectsAn AI product built by iPlus

Marginalia — Policy Answers with the Clause Cited

Employees ask about leave, per diem or security rules in Slack or Teams and get an answer with the exact clause. When the rules are silent, it goes to HR, never a guess.

Book a 30-minute callSee inside the product

On sample data · measured again before and after go-live

questions reaching HR each week
33−65%
Before: 94
of HR time back each week
20 h
citations on the right clause (HR audit of 50)
94%
How it works

How Marginalia works

The problem

  1. “How many extra leave days after five years?” “What is the per diem for a trip abroad?” Every question goes to the same person in HR.
  2. Nearly 100 repeat questions a week, and HR digs through the handbook, the travel policy and each amendment to answer.
  3. Answers from memory or from an old version go wrong, and a wrong answer turns into a complaint.

What it does

  1. 01Staff ask in Slack or Teams, and the assistant answers with the document name and clause number.
  2. 02Clicking the citation opens the original, with the passage highlighted so anyone can check it.
  3. 03It uses only the version in force. When the rules are silent it says so and sends the question to HR. It never guesses.
  4. 04Each week HR gets a list of questions the rules do not cover and places where two documents disagree.
Real product, sample data

Inside Marginalia

01/10

An employee asks about per diem: the answer is worked out step by step, with the exact clause beside it.

Rollout

Putting it to work

Rollout with iPlus

  • Weeks 1–2 · AssessmentChoose the documents to load and write a test set of questions with HR.
  • Weeks 3–8 · BuildLoad documents from SharePoint or Google Drive, connect Slack or Teams and pilot in one department.
  • Monthly · AI AdvisoryOptional, at least 3 months, to watch quality and close gaps and conflicts in the rules.
  • You provideThe policies in force, one HR contact and read access to the document folders.

Running cost

  • SetupOne-time, fixed price, quoted after a 30-minute call.
  • AI APIUnder US$ 60 a month for about 1,500 questions, billed to your own account with the provider.
  • HostingYour own cloud or a small server.

Where the AI gets its information

  • SourcesYour documents on SharePoint, Google Drive or uploaded files, plus reference law where you want it.
  • Right versionEach document has a version, an effective date and an owner; answers come only from the one in force.
  • By roleEach person gets answers only from the documents their role may see.
  • No trainingYour data is not used to train models.
Built withClaudeRAGVector databaseOCRSharePointGoogle DriveSlackMicrosoft TeamsNode.jsPostgreSQL

Frequently asked questions

Short answers to what companies usually ask before they start.

Ask anything else in 30 minutes
What happens when the assistant is wrong?
Every answer shows its source clause, so errors are easy to spot. When unsure it routes to HR instead of guessing, and HR audits a sample each month.
Do employees need training?
No. They ask in Slack or Teams as they would message a colleague.
Does it work with our software?
Yes. Documents come from SharePoint or Google Drive and answers appear in Slack or Teams. An HR system can be added later for leave balances.
Is our data safe?
Documents stay in your own systems; iPlus keeps none. Each person sees only permitted documents, and your data is not used to train models.
What does it cost each month?
Mainly the API, paid directly to the provider, estimated under US$ 60 at about 1,500 questions, plus hosting. AI Advisory is optional.
How long until it runs?
A 2-week assessment, then 4–6 weeks to build and pilot.
What if we stop using it?
Your documents and data stay yours and can be exported, and you can stop at any time under the contract.
Can we start small?
Yes. Begin with a few documents and one department, then widen once you trust the results.

Related projects

A faxed order is read and matched 3 of 3; the price is 12% over the old quote, so Orderly holds it for review.
AI Projects

Orderly — AI Agent for Purchase Order Entry

Orderly reads purchase orders from fax, email and PDF, finds the right part numbers and creates the order in kintone. Unusual orders wait for a person.

The office floor: seven AI agents work beside staff, and every decision that matters goes to the approval desk
AI Projects

Atrium — An AI Agent Office That Works Beside Your Staff

Atrium is an AI agent office: seven agents sort email and WhatsApp, draft quotes, enter orders and chase receivables, while people approve every step that matters.

Luma answers a braces price from the 2026 price list and books Saturday in the chat, with no front-desk work.
AI Projects

Luma Care — Patient Care Chatbot for Dental Clinics

A chatbot that answers patients 24/7 on WhatsApp, LINE and the website, books on the real dentist calendar and hands clinical cases to a dentist. Every answer shows its source.