Skip to content

Use case · Internal knowledge

Answers from your SOPs and policies, limited to what each person may see.

Staff ask in plain language and get a short answer with the exact section it came from. The assistant searches only the collections that person’s groups can open, shows which version it used, and says so plainly when no approved document covers the question.

Book a Strategy CallSee the step trace

At a glance

Connects to
  • Document sources
  • Your identity provider
  • Search index with permission metadata
  • Model provider or private model
  • Where staff ask
  • Admin view
Typical rollout
Week 1 → Weeks 5–6See the plan

The problem

Where the work gets stuck today

The answer usually exists, in an SOP, a policy PDF, a wiki page or a shared drive, but staff cannot find it quickly, so they ask the one colleague who knows. That person loses hours a week to the same questions, and new starters take far longer than they should to become confident.

Dropping every document into a chatbot creates a different problem. It will quote pay bands to someone who should never see them, mix an outdated procedure with the current one, and give no way to check where an answer came from.

  • The same procedural questions reach the same experts every day.
  • Staff follow an old version of a procedure because it was the first file they found.
  • Sensitive folders are either locked away from search entirely or exposed by accident.
  • Nobody knows which questions the documentation fails to answer.

Task catalog

What the system takes on

  • Task 01

    Answer with citations

    Short answers from approved SOPs, policies and guides, each linked to the exact section and document version used.

  • Task 02

    Respect existing permissions

    Only collections the asker’s groups can open are searched, using the groups your identity provider already manages.

  • Task 03

    Prefer the current version

    Superseded versions are removed from the index when a document changes, and the version date travels with every citation.

  • Task 04

    Say “I don’t know” and name the owner

    When no approved source answers the question, the assistant says so and points to the person or team who owns that topic.

  • Task 05

    Collect the gaps

    Unanswered and low-rated questions are grouped into a weekly report for each collection owner.

  • Task 06

    Meet staff where they work

    In your team chat via its API, an intranet widget or a small web app behind single sign-on.

Step trace

One run, step by step

Sample internal chat: a store manager gets a cited answer about no-receipt refunds, then asks about bonus bands and is told the HR collections are outside their access scope, with who to ask instead.
Sample data
ops-assistant · store managersSample data
  1. Visitor

    Can I approve a refund without a receipt, and up to what amount?

  2. Assistant

    Yes. Store managers can approve a no-receipt refund up to the limit set in the refunds policy, in the till with your manager code. Above that limit the area manager approves it.

    • Refunds policy · section 4 · current version
    • Till approvals SOP · step 3
  3. Visitor

    What bonus does an area manager get for hitting target?

  4. Assistant

    I can’t answer that from the documents your role can open. Pay and bonus bands are limited to the HR team. Your HR partner can help, or you can raise a request in the HR portal.

Access scope: Store managers (single sign-on group) · store operations collections searched · HR collections excluded

Read the trace as text (Sample data)
  1. 00:00Store managerStepAsks about approving a no-receipt refund, signed in through single sign-on.
  2. 00:00Access filterApprovalGroups from the identity provider: Store managers. Four store-operations collections in scope; HR collections excluded.
  3. 00:01RetrieverStepSearches only the collections in scope; best passages from the refunds policy and the till approvals SOP.
  4. 00:02AnswererDoneAnswers from those passages and cites both, with the version of each document.
  5. 00:03Store managerStepAsks about area-manager bonuses.
  6. 00:03RetrieverCheckNo passage in scope clears the relevance bar; the pay-band collection is not searchable for this group.
  7. 00:04AnswererDoneSays it cannot answer, names who can, and logs the question for the HR collection owner’s gap report.

Access scope

Same question, three roles, three answers

Permissions are applied before any text reaches the model. Switch the role to see which collections are searched and how the answer, its citations and its refusal change.
ops-assistant · access scopeSample data
Signed in as

Identity group: Store staff

Collections

  • Store SOPs: Searched
  • Refunds policy: Searched
  • Pricing playbook: Not visible to this role
  • Health and safety: Searched
  • HR policies: Not visible to this role
  • Pay and bonus bands: Not visible to this role

Question

Can I give a customer a discount to save a sale?

Answer

Associates can apply the promotions shown in the till, but not their own discounts. For anything else, call the shift manager, who can approve a goodwill discount.

Cited

  • Store SOPs · Till and promotions
  • Store SOPs · Escalating to a manager

Systems involved

What it connects to

Your existing tools stay the system of record; the AI reads and writes through their APIs.
Document sources
Shared drives, the wiki, the intranet and PDFs, connected through their APIs or scheduled exports. Only approved collections are indexed.
Your identity provider
Single sign-on and the groups you already manage. Access follows the account, so leavers lose it when their account is closed.
Search index with permission metadata
A vector and keyword index in your own cloud account, for example on Supabase or Cloudflare, with each passage tagged by the groups allowed to see it.
Model provider or private model
OpenAI, Anthropic Claude or Google Gemini on your account, or an open-weight model in your cloud for collections that must not leave it.
Where staff ask
Your team chat via its API, an intranet widget, or a small web app behind single sign-on.
Admin view
Sources, sync status, collection owners and the weekly gap report, for the people who keep the documents current.

Guardrails

  • Permissions are applied in the retrieval query from the asker’s groups, before any passage reaches the model. Nothing relies on asking the model to keep a secret.
  • Only approved collections are indexed. Drafts, personal folders and archives stay out unless an owner adds them.
  • Every answer cites its sources; when no passage clears the relevance bar, the reply is “I don’t know” plus the topic owner.
  • When a document is updated, the superseded version leaves the index, and sync runs both on a schedule and on change where the source supports it.
  • Topics you mark advisory-only, such as legal interpretations, medical questions or individual HR cases, are routed to their owner rather than answered.
  • Questions are logged for the gap report with access restricted to admins, and staff are told what is logged and for how long.

Human checkpoint

Every collection has a named owner. Owners get a weekly gap report: questions with no good source, answers rated unhelpful and the documents cited most often. They decide what to write, merge or retire. Sensitive topics are routed to the owner instead of being answered, and access changes come from your identity provider, so permissions are managed in one place rather than inside the assistant.

What we plan for

Typical targets

Typical targets are planning ranges we use when scoping, each with its basis. They are not results from a client project; your own baseline decides the real number.

  • Typical target

    None in the permission test set

    Answers that cite a source the asker cannot open

    Basis: Group filters are applied in the retrieval query and tested by asking the same questions from every role before launch, and again whenever a collection’s access changes.

  • Typical target

    30–50% fewer within the first quarter

    Repeat questions reaching your subject experts

    Basis: Planning range when the most frequent questions are covered by approved documents; the experts log incoming questions for a fortnight before launch and again after it.

  • Typical target

    Every answer, by design

    Answers with at least one citation to a current version

    Basis: A reply without a qualifying source is replaced by “I don’t know” and the owner’s name; checked on the evaluation question set before each release.

Rollout

Week by week

Typical ranges for a first release; the scoping call fixes the plan for your systems.

  1. Week 1

    Collections inventory, owners, permission groups and the advisory-only topic list; evaluation questions collected from staff.

  2. Week 2

    Connectors and sync for the first collections, permission tags on every passage, chunking and the citation format.

  3. Week 3

    The assistant in one channel, the evaluation set run, and cross-role permission tests.

  4. Week 4

    A pilot with one team, the first gap report and a review with the collection owners.

  5. Weeks 5–6

    More collections and teams, scheduled syncs, and the admin view handed over to your owners.

FAQ

Frequently asked questions

How does the assistant know which documents each employee may see?

It reads the groups from your identity provider when a person signs in, and every passage in the index carries the groups allowed to see it. The search only returns passages that match the person’s groups, so text they may not see never reaches the model in the first place.

Could staff get an answer from an outdated version of an SOP?

The index keeps only the current version of each document: when a file changes, the old passages are removed and the new ones added. Each citation shows the version date, so if a sync ever lags, staff can see it.

Can it answer from scanned PDFs, slide decks and spreadsheets?

Mostly, yes. Scans are run through text recognition, slides are read with their notes, and tables are kept as tables so rows and columns stay together. Poor scans and image-only slides are flagged during indexing so owners can replace them.

Do employees need a new app to use it?

Not necessarily. It can sit in your team chat via its API, inside the intranet, or in a small web app behind the single sign-on staff already use. Most teams start with the place where questions are already being asked.

Can managers see what individual staff are asking?

Owners see aggregated gap reports, not a feed of named questions. Individual logs are restricted to a small admin group for troubleshooting, kept for an agreed period, and described in a notice to staff before launch.

What if two documents contradict each other?

The assistant cites both and says they disagree instead of quietly choosing one. The conflict is flagged to both document owners in the gap report, which is often the first time anyone notices the duplication.

How is this different from a chatbot for our customers?

An internal assistant searches private documents and must respect who is asking, so permissions, single sign-on and owner reports are the core of the build. A customer-facing assistant answers from public help content and hands off to support; that is our customer support automation use case.

Next step

Have a workflow in mind? Let’s scope it together.

Book a free 30-minute strategy call. Bring one process you want to automate and we will talk through scope, integrations and a sensible first step.

Book a Strategy CallGet an AI Plan

Free 30-minute call · no commitment

Send an RFP or security questionnaire