AI knowledge assistant: answers from approved sources, with an owner
An AI knowledge assistant answers staff or customer questions by retrieving passages from approved documents, drafting a reply from those passages and showing where each part came from. It is worth building when every source has an owner, unanswerable questions reach a named expert, and answers are tested on real questions before anyone relies on them.
Submit a project See the related serviceReviewed by David Nguyen (CEO) · Updated 28 Sep 2026
OutsourcingVN is operated by Netbase JSC, which delivers this workflow through its AI Workflow Automation service, so read this page as a supplier's description. The wider decision of which workflow to automate first is covered in the AI workflow automation guide.
Who is in this workflow
-
The expert who keeps being asked
A senior engineer or policy lead answers the same questions every week; those interruptions are the cost the assistant exists to reduce.
-
The person asking
A new hire, support agent or partner who needs an answer now and cannot tell which document is current.
-
The content owner
Each collection of sources needs someone who decides what is approved, what is retired and when a document is due for review. Without that role the assistant quotes last year's policy with confidence.
-
The answer owner
Someone signs off on what the assistant may say about contract terms, HR cases or legal positions, and what it must refer elsewhere.
-
The platform engineer
They run the index, the permission sync and the logs.
Where do internal answers go wrong today?
-
Documents disagree
Versions of the same procedure sit in a wiki, a shared drive and an email thread, and nothing marks which one is current.
-
Search finds files, not answers
The reader still has to open twenty documents and reconcile them.
-
Permissions are handled badly
Either a chatbot exposes a restricted folder to everyone, or it is locked down so tightly that it cannot answer anything useful.
-
Unanswered questions leave no trace
Nobody sees what was asked and not found, so the gaps in the documentation are never fixed.
The target flow
-
The question arrives with an identity
The assistant knows who is asking, so retrieval is limited to documents that person could already open.
-
Retrieval comes before generation
The system finds the most relevant passages in approved collections only, each tagged with its document, version and owner.
-
The answer is grounded and cited
The model drafts a reply from the retrieved passages and cites them. A statement with no passage behind it is not written.
-
Weak evidence produces "I don't know"
When retrieval is thin or sources conflict, the assistant says so and routes the question, with the passages it found, to the named expert or queue.
-
Feedback reaches an owner
A reader can flag an answer as wrong or out of date, and the flag goes to the owner of the cited document rather than to a general inbox.
-
Gaps become a writing backlog
Unanswered and flagged questions are grouped weekly, so a content owner writes the missing page once instead of an expert answering it forty times.
Where does the system boundary sit?
A project of this shape owns ingestion, the index, the permission filter, cited answer generation, the hand-off route, the gap log and the evaluation set. It does not own the source systems, the identity directory or the decision about which documents are approved; that stays with the content owners. When questions arrive from customers on a messaging channel and must become routed cases, the workflow is customer support automation. When the job is extracting fields from incoming documents and approving them, it is document and approval workflows.
Sources, data and integrations
The objects to design are a source register (owner, review date, sensitivity), document versions, passages with metadata, the index, the permission mapping, the question log and the evaluation set. Typical connectors reach a wiki, file storage, ticketing and a CRM. Permissions should be synchronised from the source, never copied once and left to drift.
Retrieval choices, such as passage size, keyword and vector search combined, re-ranking and freshness rules, are explained in the guide to retrieval and context engineering.
A retrieved document can contain text that reads like an instruction; OWASP lists prompt injection as the first risk for model applications, so retrieved text is treated as data. Question logs hold personal data, so retention is decided before launch. Netbase's security practices include secure code review and version control, role-based access control, MFA for admin dashboards, contributors under NDA, and NDAs and DPAs on request.
What the model does, and what it must not do
The model rewrites the question for search, ranks passages, drafts the grounded answer and groups unanswered questions by topic. It must not invent a policy, answer from outside the approved collections, or act in other systems. An assistant that files tickets or changes records is a different design with its own controls, covered in AI agents versus workflow automation.
Netbase works with commercial and open-source AI models chosen per project (model-agnostic); no vendor partnership is implied. The model is chosen after the evaluation set exists and tested against it, as described in choosing an AI model for a business workflow. Netbase applies the ISO/IEC 42001 AI management system framework to its own AI delivery practice; that is how Netbase runs its own work, and it does not extend to the assistant a client operates. NIST's Generative AI Profile is a useful public risk list to review with your own team.
Delivery modules
- Source register and ingestion connectors, with versions and owner metadata.
- Index and retrieval, with permission filtering applied at question time.
- Answer service with citations, a refusal threshold and a hand-off route.
- Reader interface: a panel inside an existing tool, a web page or an API into a product.
- Feedback capture, gap log and a queue for each content owner.
- Evaluation set, regression runs and a monitoring view.
- Runbook for re-indexing, adding a collection, changing the model and pausing the service.
How is a knowledge assistant rolled out?
Start with one collection, one audience and the fifty questions they ask most, turned into an evaluation set with the expected source for each answer. Then run in shadow mode: the assistant drafts, an expert reviews, and nothing reaches the asker until the agreed threshold is met. How to assemble that set is covered in evaluating and testing AI workflows, and where the review point belongs in human-in-the-loop AI workflows.
Netbase's delivery lifecycle: discovery and strategic alignment; team assembly and architecture planning; agile execution with outcome-based milestones; modular components; training and rollout; ongoing support. Netbase delivers remote-first from Hanoi in Agile increments with weekly reviews, using AI-assisted engineering under human review. Milestone acceptance follows the project delivery page. For custom development the client owns the IP created for it; Netbase productized modules and products are licensed, not transferred.
How is success measured?
Take a baseline of expert questions per week and waiting time, then track the grounded-answer rate on the evaluation set, answers whose citation actually supports them, the hand-off rate, the time from a flagged gap to a corrected document, and expert interruptions per week. These are your numbers, not a benchmark; the methodology page explains how evidence is labelled here.
Where the pattern holds
-
Policies and handbook
- Who asks
- Employees
- What changes in the design
- Strict permissions by role and country; HR cases always hand off
-
Engineering runbooks
- Who asks
- Support and on-call engineers
- What changes in the design
- Freshness rules matter most; stale steps are worse than none
-
Product and pre-sales knowledge
- Who asks
- Sales engineers and partners
- What changes in the design
- Commercial answers need an answer owner's sign-off
-
Public help centre
- Who asks
- Customers
- What changes in the design
- A higher threshold, a visible route to a person, public sources only
It does not hold where answers were never written down, where the answer is a decision needing approval, or where it depends on today's transactional data, which is a report rather than retrieval.
What delivery record exists
Netbase has delivered anonymised client AI projects including retrieval-based knowledge assistants, document AI and MLOps pipelines. None of those knowledge-assistant projects has a published record on this site, so this page rests on capability and method rather than on a record a reader can open. Three published records are adjacent.
- Conversational AI with CRM sync. For a client that is not named, Netbase built a WhatsApp Business AI chatbot with intent and conversation-flow handling, an LLM API (GPT-4 in the stack) and CRM synchronisation, delivered in milestones from design and prototype to documentation, knowledge transfer and 30 days of support over 4-8 weeks. See the WhatsApp AI chatbot record.
- A review queue for model output. A classifieds platform used AI-powered content filtering that detects and flags offensive content into an admin moderation dashboard. See the online classifieds platform.
- A knowledge base as a product module. Phase 1 (2020-2023) of a multi-tenant cloud ERP SaaS for a US client covers CRM, real-time messaging, HR, knowledge base, custom fields and workflows. It is a knowledge base, not an AI assistant. See the cloud ERP SaaS record.
No volume, accuracy or business result is claimed for any of them.
This record covers phase-one scope for agency SMEs; the client and product are not named, and no usage or business result is claimed.
Keep Reading
Netbase delivered a bilingual classifieds marketplace, in English and Nepali, with paid ads, search, messaging, payments and an AI filter that flags offensive content into an administrator review queue. The client is not named, and no traffic, accuracy or commercial result is claimed.
Keep Reading
The client, sector, region and CRM product are not named, and no conversation volume or business result is claimed.
Keep ReadingCommon questions
Grounding in retrieved passages, citations and a refusal threshold make that much less likely, but not impossible. The evaluation set measures how often it happens on your questions, and the pilot does not go live until the agreed threshold is met.
Yes, if permissions are checked at question time against your identity directory rather than copied into the index once. Acceptance tests have accounts with different roles ask the same question.
No. Pick one collection, assign owners and retire obvious duplicates; the gap log shows what to write next.
Yes, where data rules require it. An open-source model hosted in your infrastructure is tested against the same evaluation set as any hosted option.
A support chatbot resolves and routes cases on a channel; a knowledge assistant answers from documents and cites them. The two can share one index.
Services behind this solution
AI workflow automation with evaluation and human control
AI Workflow Automation starts with one operating workflow, one accountable owner and one agreed way to judge the result. The goal is a workflow that handles the routine cases correctly on representative test cases, routes uncertain or high-impact cases to a person, and can be monitored and changed after handover. It is not a promise that every process can or should be automated.
Learn More
Scope a knowledge assistant
Bring the fifty questions your experts answer most often, the collections where the answers should live and who owns each, and the systems that hold permissions. If whether to build it at all is still open, an AI Workflow Blueprint settles that as a bounded piece of work, and the neighbouring workflows are listed under solutions. When the scope is clear, submit a project and name the collection you want answered first. OutsourcingVN is Netbase's own outsourcing-services platform.