Internal knowledge assistants
We build internal assistants that answer staff questions from your policies, procedures and reference documents, cite the exact sources they used and respect existing permissions, so people only get answers drawn from documents they are already allowed to open.
Answer
Part-time staff receive bereavement leave pro-rated to their scheduled hours, up to three days.
Drawn from two sources below. Check the policy for eligibility details.
HR Policy Handbook · §7.4
“Leave for part-time employees is calculated in proportion to scheduled hours…”
Updated May 2026 · Visible to all staff
Leave FAQ for Managers
“Bereavement leave does not require prior approval but must be recorded…”
Updated August 2026 · Visible to managers
Restricted source
1 matching document not used
Payroll exceptions log: you do not have access, so it was excluded.
Feedback
Was this answer correct?
Reports go to the HR content owner.
Different from a support bot
A knowledge assistant answers internal questions: HR policies, operating procedures, technical documentation, onboarding guides. It works across documents with different audiences, so access control is the core design problem.
A customer-facing assistant, by contrast, answers from public help content and escalates to your support team. That is covered by AI customer support.
The problem
Policies live in a drive, procedures in a wiki and the real answer in an email thread from last year.
Experienced staff become the search engine, interrupted many times a day.
An old copy of a procedure is easier to find than the current one.
Onboarding depends on who happens to be available to answer questions.
Trust, security and oversight
An assistant that ignores permissions is a data-leak risk. Permission-aware retrieval is a requirement, not an option.
Each document keeps its access rules, and retrieval filters by the person asking before anything reaches the model.
Staff can open the source and check it. Answers without a supporting source are not given.
Sources show their last update date, and stale documents can be flagged or excluded.
When the documents do not contain an answer, the assistant says so and suggests who to ask.
Wrong or missing answers are reported to the person responsible for the source.
Example scenarios
Fictional examples. They are not Alaysalee clients, deployments or measured results.
Situation
Managers at a 120-person company email HR with the same leave and expense questions.
What would be built
The assistant answers from the handbook and manager guides, citing sections, and respects the split between all-staff and manager-only documents.
A person stays involved
Questions about a specific employee’s situation are directed to HR.
Situation
Field technicians need the current procedure for equipment they service only occasionally.
What would be built
Technicians search procedures and checklists in plain language and get the steps with a link to the full document.
A person stays involved
Safety-critical procedures always link to the controlled document rather than paraphrasing it.
Situation
New hires at an agency ask the same setup and process questions in their first month.
What would be built
The assistant answers from onboarding guides and the internal wiki, and points to the right person when a document is missing.
A person stays involved
Unanswered questions are collected weekly so the wiki can be improved.
Scope and deliverables
Which repositories are in scope, who owns them and how their permissions work.
Connections only to the sources you authorize, using the platform’s own access controls where available.
Documents split, indexed and tagged with their access rules, refreshed on a schedule.
How answers are phrased, cited and limited, including the “no answer” behavior.
A set of real questions with known answers, used before launch and after changes.
How reports reach content owners and how corrected documents are re-indexed.
Capabilities and limits
Readiness
Fit
Engagement approach
Most projects start with one or two well-maintained sources and a pilot group.
We list the repositories in scope, their owners and their permission models.
Documents are indexed with their access rules attached.
Every answer links to its sources, and “no answer” is allowed.
Real questions with known answers are tested before launch and after changes.
Wrong or missing answers are routed to content owners.
Yes, that is the purpose, provided access is designed carefully. Documents are indexed with their permissions, and we agree which AI provider processes the text and under what terms before any real content is used.
The assistant shows when each source was last updated, and stale documents can be flagged or excluded. It cannot know that a document is wrong, which is why feedback goes to content owners.
Each indexed document keeps its access rules, and retrieval filters by the person asking before anything reaches the language model. We test this explicitly with restricted documents.
Language models can produce plausible but wrong text. Requiring a supporting source for every answer, showing citations and allowing “I don’t know” reduces that risk significantly, but staff should still check sources for important decisions.
It depends on the number of sources, their permission models and their quality. A pilot on one or two well-organized sources is the quickest way to learn what is needed.
Next step
Tell us where your internal knowledge lives and who needs it. We will look at sources, permissions and a sensible pilot.
We reply by email. An inquiry is not a contract, and there is no obligation.