Use caseProcurement
Lucy needs a signed contract and its notice period by Thursday. It was filed years ago under a supplier name that has since changed.
A question in plain words finds the documentEvery answer names the documents it usedAccess rights apply before the search runs
Your payroll provider is Paylink Inc., called Arbor Payroll until 2023. The signed data processing agreement is dated March 4, 2021.
source · supplier contracts, 4 documentsThe document answers · it opens at the cited page
Arbor Payroll became Paylink Inc. in 2023, and nobody renamed the folder. That is how documents go missing in a repository: filed under a supplier's old name, a project that ended, a folder created by someone who left. Staff search by file name, then ask around. IT gets asked to open broad access to speed things up, and compliance refuses. The search ends with a colleague's memory.
Lucy types her question the way she would ask a colleague. The assistant turns it into a search on content and metadata, with her rights, and returns the documents it used. Each person finds everything their access allows, and nothing more.
A search box and a settings page look like nothing new. The care is in the order: rights are set where the documents live, the question is asked in plain words, the answer names its documents, and every exchange is kept. The steps below start with Oliver's access settings and end with his November usage counts.
Before anyone searches, Oliver settles two things: who may read what, and where the model runs.
Oliver Mercer is IT director at a company of 2,400 people. In the repository, access is set by group, down to the document. Procurement reads supplier contracts. Legal owns a litigation space that only its own group opens. The search adds no rights of its own.
Contracts hold prices, names and personal data. Oliver sets the search task on a model hosted at the company's site, and the data residency routing rule keeps every call there. Usage is tracked by department, so he can tell who uses the search.
Lucy meets the payroll provider on Thursday about the renewal, and needs the signed agreement and the notice period.
Lucy Gardner does not remember the provider's former name or where the agreement was filed. She types: the signed data processing agreement with our payroll provider. The question becomes a search, and the answer comes back as a table of four documents, each naming its document class.
Lucy opens the services agreement in the viewer and asks how much notice the company must give to leave. The answer quotes clause 14.2 and its page. She reads the clause herself before writing the date in her notes.
The table listed an amendment from 2025. Lucy asks what it changed compared with the agreement. The answer names the clause it touched: exhibit 3, the list of sub-processors, with two names added. Every other clause stays as signed in 2021.
Lucy asks whether there has ever been a dispute with Paylink. The answer comes only from documents she may open, and none mentions one. She emails Helen Faber in legal, who asks the same question with her own rights. Helen finds a 2025 settlement and tells Lucy what she can use for Thursday.
Lucy has what she needs. The results go into her renewal file, and IT reads how the search is used.
Lucy exports the table of documents to Excel for her renewal file. Each row keeps the document name, its class and its status, so her manager finds the same documents with his own access.
Every Monday Oliver opens the exchanges in the console. Each one keeps the user, the question, the answer, the cited documents, the model and the date. He sees that facilities staff often search for the same lease documents, and asks them to agree on a document class.
The console counts requests and tokens by department. Procurement and facilities use the search most. Oliver brings the counts to the steering committee instead of a guess about adoption.
Lucy's questions about Paylink hit all three cases in one morning. Each case gives the person asking a different next move.
The table names the document and its class, and the answer quotes the page.
Next: the reader opens it and checks the clause herself.The table lists both, with their status, so nobody quotes the draft by mistake.
Next: the reader picks the signed copy, or compares the two in the viewer.The answer says so, and comes only from documents the person may open.
Next: the question goes to the team that owns the subject, which searches with its own rights.It grants no access. It shows nothing the person asking could not open. Across the nine steps on this page, access is decided by the people who own the content, where it is stored.
The assistant turns a sentence into a search and quotes its sources. Permissions apply before the call, and every exchange is kept with its model and date.
Counted on Lucy's search before the Paylink renewal. A repository with tighter document classes would return shorter tables.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
The people who search, the people who run the repository and the people who guard access all have a stake, and each applies its own test.
The document exists, and only the person who filed it knows where.
Each search tool you have tried wanted its own copy of the documents and its own rights.
Someone always wants broad access opened so people find things faster.
Open the plain-language search on one document class first, such as the supplier contracts your teams ask about most. A few weeks of real questions will show what people actually look for.
Your repository, your document classes and your access rights stay as they are.
Next step
Bring a few questions your teams ask every week and a test set of anonymized documents. We will run the search with two users who have different rights, and compare.