An internal knowledge assistant, something staff can ask a plain-English question and get an answer grounded in your business's own documents rather than the internet's general knowledge, is one of the highest-value first builds for an Australian business running Claude. It is also one of the easiest to get subtly wrong, because the failure mode isn't a crash, it is a confident-sounding answer that quietly isn't grounded in anything real.
Why grounding is the whole point
A general-purpose Claude conversation reasons from its training knowledge and whatever you paste into the chat. An internal knowledge assistant needs to reason from your actual policy documents, your product specs, your past client correspondence, and say clearly when the answer isn't in there rather than filling the gap with a plausible guess. That distinction, grounded versus merely plausible, is the entire value proposition, and it is also the part most first attempts skip.
The mechanism that makes this work is retrieval: before Claude answers, a search step pulls the most relevant chunks of your actual documents and hands them to the model alongside the question, so the answer is built from real source material rather than from memory. Get the retrieval step wrong, poor document chunking, a search index that misses the right file, and the grounding falls apart even if the model itself is perfectly capable.
What actually makes a knowledge assistant trustworthy
Source citations on every answer, so a staff member can click through to the actual policy document or email thread the answer was drawn from, not just trust it blind.
An explicit "I don't know" path, an assistant that says the information isn't in its indexed documents is far more useful than one that fills the gap with a confident guess.
Regular re-indexing, a knowledge base grounded in six-month-old documents quietly becomes wrong the day a policy changes, unless the index is kept current.
Scoped access, an assistant should only surface documents the asking staff member is actually permitted to see, not the entire company's file store indiscriminately.
Document chunking deserves a specific mention because it is the least glamorous part of the build and the most common source of quiet failure. Split a long policy document into chunks that break in the wrong place, halfway through a clause, separating a condition from its exception, and the retrieval step can pull back a technically-real quote that is misleading out of context. Sensible chunking, respecting section and clause boundaries rather than a fixed character count, is unglamorous work that determines whether the whole system is trustworthy.
Where Skills and MCP connectors fit
For a business already using Claude Cowork, a knowledge assistant often starts as a connector pointed at a specific folder of documents, contracts, policies, product manuals, paired with a Skill that defines how answers should be structured, always cite the source, always flag uncertainty, use the business's own terminology. That combination, connector for the data, Skill for the behaviour, is usually enough for a first working version without a bespoke retrieval pipeline.
A more sophisticated build, useful once document volume grows past a few hundred files or the business needs sub-second search across everything, adds a proper vector search index behind an MCP server, so retrieval scales without relying on Claude re-reading an entire folder for every question. That is a genuine engineering build, not a weekend project, but it is not necessary for most AU SMBs starting out.
A worked example of grounding failing quietly
A Brisbane professional services firm piloted an internal assistant against its HR policy documents and got glowing early feedback, until a staff member asked about parental leave entitlements and received a confident, well-formatted answer that turned out to blend the firm's actual policy with general knowledge about typical Australian leave entitlements the model had picked up during training. Nobody had told the assistant to cite its source, so nobody noticed the blend until someone checked the actual policy document and found a mismatch.
The fix was not a smarter model. It was a stricter Skill: always quote the specific policy clause, always name the document and section, and explicitly say when a question falls outside what the indexed documents cover rather than filling the gap. That one change, forcing citation discipline, turned a genuinely risky tool into a trustworthy one, without touching the underlying retrieval setup at all.
The Automata AI take
We build internal knowledge assistants for Sydney and Melbourne clients starting from exactly this connector-plus-Skill pattern, scaling to a proper search index only once the document volume genuinely needs it. A first working version, scoped to one department's documents with citations and access controls in place, typically runs A$5,000 to A$12,000 depending on document volume and how many source systems it needs to touch.
Book a brainstorm and we will scope what a grounded knowledge assistant would look like for your document library.
