Internal knowledge agent
We design, build and deploy an AI agent that answers your team's questions from your own SOPs, policies, HR handbook and product docs — with the exact source cited, and an honest “that's not in our documents” when it isn't.
At a glance
Frees your experts, HR and IT from re-answering the same questions all day, and gives every employee an instant, consistent, sourced answer — value spread across the whole company.
Retrieval over your documents, keeping the index fresh, grounding every answer in a citation, a calibrated “I don't know”, and role-based access so people only see what they're allowed to.
What changes when it works
Off-the-shelf first: you may not need a custom build
Be honest before you build. If your documents already live in tools Claude connects to — Google Drive, SharePoint, Slack, GitHub and more — and you're OK with per-seat pricing, Claude for Work with connectors can stand up an internal knowledge agent with almost no engineering: point it at your sources and go. As a Claude partner we'll say it plainly — if that's all you need, start with a plan. A custom build (the rest of this page) earns its keep only when you need what the seat product can't do on its own: role-based access your compliance team can prove, retrieval scoped per department, PDPA-bound or on-prem data that can't sit in a seat product, your own chat surface like WhatsApp or Teams, or auditable citations and an eval gate.
Time per application
Finding the answer today
~10–20 min
With the agent
~seconds
Time saved
≈ 10–15 minutes per question — plus the hidden interrupt cost on your experts. At 500 questions a month that's ~80–125 hours of staff time back.
Indicative; actual savings depend on how often staff currently hunt for answers or interrupt colleagues.
Our agentic development process
Every use case follows the same seven stages — from framing the problem to production.
Our process, built on Anthropic's agent guidance and the Agent GPA eval framework: Building Effective Agents · Agent GPA
Stage 1 · Frame
The business problem
In most Malaysian companies the answer to “what's our medical claim limit?”, “how do I apply for leave?”, “which form do I use for a supplier?” already exists — in an HR handbook, an SOP, a policy PDF, a product manual. But nobody can find it. So staff message HR, IT or a senior colleague, wait for a reply, and the same questions get answered dozens of times a week. The knowledge is there; the access isn't.
Stage 2 · Map
The manual workflow today
- 01Employee has a question
- 02Searches the shared drive / intranet, gives up
- 03Messages HR / IT / a senior colleague
- 04Waits for a reply (minutes to hours)
- 05The expert stops their own work to answer — again
Where it breaks: The bottleneck isn't missing knowledge — it's retrieval. The answer exists in a document nobody can find, so it gets re-asked and re-answered.
Stage 3 · Design
The agentic workflow
We turn your documents into an agent staff can just ask. It works out which knowledge source the question belongs to, retrieves the relevant passages, drafts a grounded answer, and verifies every claim traces to a citation before replying — and abstains honestly when the answer isn't in the docs.
If a claim isn't backed by a retrieved source, it re-retrieves or abstains rather than guess.
Watch a case flow through the agent — it does the reading and matching; a human still makes the decision.
See it in action
Staff message
Sourced answer
- Panel clinic: RM 500 per visit, outpatient (HR Handbook §4.2)
- Non-panel: not covered unless pre-approved by your manager (§4.3)
- Claim via the HR portal within 30 days, keep the receipt
Cites the exact document & section · says “not in our docs” rather than guess · ~RM 0.01–0.05 per question.
Step by step
- 1
Understand the question
It interprets the question — including mixed Bahasa Malaysia / English / Chinese and vague phrasing — and works out which knowledge domain it belongs to: HR, IT, finance policy, or a product manual.
- 2
Retrieve the right passages
It searches the indexed document set (semantic plus keyword) and pulls the specific passages that answer the question — not the whole document — so the answer is grounded in exact text.
- 3
Draft a grounded answer
It writes a short, direct answer using only the retrieved passages, in the language the person asked in, and keeps the source reference attached to each point.
- 4
Verify the citations (the check)
Every claim in the answer must map to a retrieved source; if a sentence isn't backed, it re-retrieves or drops it — so it cannot confidently state something that isn't in your documents.
- 5
Answer or abstain honestly
It returns the answer with the source, or says “I couldn't find that in our documents” and offers to route the question on. No confident guesses.
- 6
Route the exceptions to a human
Sensitive, out-of-scope or genuinely unanswerable questions go to the right owner — HR or IT — with the question attached, never a dead end.
Stage 4 · Build
How we build it with Claude
It answers only from your own documents, never the open internet — so it can't invent a policy you don't have. Every answer carries its source, restricted documents stay restricted to the right roles, and re-indexing on change means an updated SOP is live the moment you publish it, not weeks later.
Integrations
- Your document store — SharePoint, Google Drive, Confluence, Notion or PDFs
- The chat surface staff already use — Teams, Slack, WhatsApp or web
- Optional: your HRIS / ITSM to route the questions it can't answer
Under the hood
- Pattern: a retrieve → ground → cite-check chain on the Claude Agent SDK, with a light routing front-end that picks the right document set per question — each capability exposed as an MCP tool.
- Ingestion & retrieval (RAG): documents are chunked by their own structure — sections, clauses, tables — so each passage is self-contained and keeps its citation, not blind fixed-size slices that cut a policy mid-clause. Retrieval is hybrid: a vector (embeddings) index for meaning plus a keyword index for terms that must match exactly — policy codes, form numbers, “§4.2”, SKUs; we over-retrieve then re-rank the top passages before the model answers. Pure semantic search misses exact identifiers; pure keyword misses paraphrase — hybrid catches both.
- Context & freshness: only the re-ranked passages enter the window (prompt-cached for the stable system context), and the index is rebuilt incrementally on document change — so an updated SOP is live the moment it's published, no stale answers.
- Tools: the search/retrieve tools and a topic router/classifier — all read-only; no tool can write to your source systems.
- Grounding & citations: answers are constrained to the retrieved passages and every claim carries a source; unsupported sentences are dropped — the anti-hallucination guarantee, enforced, not hoped for.
- Access control: retrieval is scoped to the asker's role and enforced at the tool layer — restricted HR or finance documents are only searchable by people entitled to them, so the agent can't surface what someone isn't cleared to see.
- Untrusted input: both the question and the document text are treated as untrusted (prompt-injection resistant) — a booby-trapped document can't make the agent leak or misbehave.
- Observability: every answer logs the question, the passages used and the sources cited — so any answer can be audited, and a wrong or outdated document traced and fixed.
- Eval harness: a golden set of real staff questions with known correct answers plus known “not in the docs” cases; scored on both answer correctness and on abstaining when it should — regression-tested on every change and every document update.
Stage 5 · Architect
Single model or multi-agent?
Answering from documents is one tightly-scoped task per question, so a single grounded model with a cheap classifier in front is the right call — reliable, cheap, and easy to audit.
Multi-agent designs (≈15× the tokens) buy nothing here — this isn't a parallel, breadth-first task.
Which model does what
Classify & route the question
Cheap, fast intent and topic classification at high volume — most questions never need more.
Draft the grounded answer
Strong reasoning over retrieved passages and messy mixed-language questions — the workhorse.
Nuanced / high-stakes policy interpretation
Reserved for the small share of ambiguous, consequential questions where wording really matters.
What it costs — an estimate
AI usage — per question
short retrieval + grounded answer, mostly at Claude Haiku/Sonnet token rates
~RM 0.01–0.05
AI usage — at 500 questions/month
most questions resolve on Haiku; only nuanced ones reach Sonnet or Opus
~RM 25–120 / month
Build (one-off)
indexing your documents + wiring the chat surface is a few weeks; we quote fixed after a short discovery
scoped per document set
Ongoing
small next to the expert time it frees
monitoring + re-indexing
Indicative only, shown in Ringgit at roughly RM 4.70 to the USD; per-token rates follow Anthropic's published pricing (confirm current figures). Actual AI usage depends on question volume and how much text each answer needs.
Stage 6 · Evaluate
How we measure success
- Grounded-answer checks — every claim in the answer must trace to a retrieved source
- Abstention accuracy — it says “not in the docs” when the answer genuinely isn't there
- Business KPIs — questions self-served, expert interruptions removed, time-to-answer
We score the agent with the Agent GPA framework — Goal, Plan, Action, a current standard for agent reliability — on top of the human baseline. (reference)
Setting the baseline
First we sample how staff get answers today — how long it takes, how often an expert is interrupted, and how consistent the answers are. That human baseline is what every agent metric is measured against, so the lift is provable rather than assumed.
What we test — Goal · Plan · Action
Answer correctness
On held-out real questions, the answer matches the correct documented answer
Right source retrieved
It pulls the passage that actually answers the question
Abstains when it should
On questions the documents don't cover, it says so instead of guessing
Every claim cited
No sentence in the answer lacks a source
No hallucinated policy
It never states a rule that isn't in the documents
Access respected
It never surfaces a document the asker isn't entitled to see
The go-live gate
The agent doesn't go live to staff until it beats the human baseline and clears these GPA targets on a golden set of real questions — including ones deliberately not in the docs. Every change and every document update is regression-tested against the same set before it ships.
Stage 7 · Deliver
How we'd deliver it
We start with a proof-of-concept on one document set — your HR handbook or your IT SOPs — and prove it answers correctly and abstains honestly before widening to more sources. A focused first rollout is a matter of weeks.
Free consultation
Is your team asking HR and IT the same questions all day?
Tell us which documents hold the answers and how your staff ask today — we'll tell you honestly whether an internal knowledge agent is worth building, and what it would take.
Related use cases
Frequently asked questions
- Where does the agent get its answers?
- Only from the documents you give it — your SOPs, policies, handbooks and manuals — never the open internet. If an answer isn't in your documents, it says so rather than guess, and every answer cites the exact source.
- How does it actually find the right answer in our documents?
- Two steps. When we set it up, we break your documents into meaningful pieces — by section and clause — so it can quote the exact paragraph rather than point vaguely at a 40-page PDF. Then, for each question, it searches two ways at once: by meaning (so “how much can I claim?” finds the “RM 500 limit” clause even in different words) and by exact term (so “form BE” or “clause 4.2” matches literally). It takes the best-matching passages, answers only from those, and shows the source. Searching by meaning alone would miss exact codes; by keyword alone would miss the paraphrased questions people actually ask — doing both is why it lands on the right passage.
- Can staff see documents they're not allowed to?
- No. Retrieval is scoped to each person's role and enforced at the tool layer — restricted HR or finance documents are only searchable by people entitled to them. The agent can't surface what someone isn't cleared to see.
- What happens when our documents change?
- We re-index on change, so an updated policy or SOP is used the moment it's published — no stale answers. Keeping the source documents correct is the one thing that stays with your team.
- How is it different from a chatbot like ChatGPT?
- A generic chatbot answers from the public internet and can make things up. This agent answers only from your documents, grounds every claim in a citation, respects who can see what, and abstains when it doesn't know — built for internal accuracy, not chit-chat. It's the internal counterpart to our customer-facing support triage agent.
Examples are based on real, anonymised engagements; details are generalised. Anchor Sprint is a member of the Anthropic Claude Partner Network — a deployment and rollout partner, not a reseller. This is general information, not legal or compliance advice.