Australian consultancies that have built deep expertise across years of client work usually have a knowledge problem the founders themselves do not see. The expertise sits in three places: the senior partners' heads, the working files of past matters, and a Confluence or SharePoint site nobody opens. An internal AI knowledge base done well consolidates all three. Done badly, it simply adds a fourth location nobody opens.
The cost of doing nothing is easy to underestimate because it never shows up as a line item. For a 60-person Australian consultancy with a fully loaded average hourly cost of $215, where senior consultants spend eight hours a week answering some version of "have we done something like this before", more than $1.6M of annual capacity is tied up in lookups. A working internal AI knowledge base returns 50 to 70 percent of that, and it does so without hiring anyone.
What "structure that works" actually means
The single biggest failure mode for consultancy knowledge bases is treating the KB as a dumping ground. Every deliverable, email thread and half-finished draft is dropped in. Nothing is curated. Nobody can find anything, so everybody goes back to asking the partner who was on the job. The structure that works in practice has three layers:
A curated front layer holding the firm's frameworks, methodologies and "how we think" documents
An indexed middle layer with case studies, past engagements and reusable artefacts
A searchable back layer of raw working files, surfaced through retrieval but never browsed directly
The proportions matter. The front layer is small, perhaps 200 documents. The middle layer runs to a few thousand. The back layer is unbounded. The retrieval layer composes across all three, which is what lets a junior consultant in the Sydney office find the methodology, the precedent engagement and the raw model in a single query instead of three Slack messages and a week of waiting.
Curation is the unlock
A knowledge base without curation produces mediocre answers, because retrieval surfaces too much noise for the model to synthesise anything sharp. Curation is not optional, and it is not a one-off project. It is an operating discipline, and it is cheap compared to the capacity it recovers.
A senior consultant owns each domain folder and reviews it quarterly
Every artefact carries a "still current" flag with an expiry date
Past engagements get a structured summary template covering problem, approach, outcome and lessons
The firm's frameworks are tagged so retrieval can prioritise them when answering
Without this discipline, retrieval surfaces 12 outdated proposals before the partner finds the current methodology, and the firm quietly concludes that the knowledge base does not work. The technology gets the blame for a curation problem.
Privacy and confidentiality
Consultancy knowledge bases hold client-confidential material, which makes the Privacy Act and individual client confidentiality deeds live constraints rather than afterthoughts. The retrieval layer must enforce client-level walls so a consultant staffed on Client A cannot retrieve from Client B's matter files. Firms serving APRA-regulated clients should expect these controls to be tested in vendor due diligence questionnaires, so it pays to build them in from day one:
Folder-level access controls tied to the staffing matrix
Retrieval queries scoped to the user's authorised projects
An audit log of every retrieval that surfaces client-confidential content
Automatic redaction of client names in any output that flows outside the firm
The "Claude on top" decision
Most Australian consultancies should put Claude on top of the knowledge base rather than relying on the underlying tool's native search. The difference in answer quality is large, because Claude synthesises across multiple retrieved chunks and cites its sources, where native search returns a ranked list of documents and leaves the synthesis to a busy human. As a deployment, it looks like this:
A Claude Skill scoped to the firm's knowledge base
Every claim in an answer cited against the source document
A refusal to answer when retrieval finds no supporting material
Access controls enforced above the model, so Claude only ever sees what the querying user is entitled to see
The citation and refusal behaviours are what turn the KB from an interesting demo into something partners trust. A wrong answer delivered confidently to a client costs more than the entire build, so the system has to show its working every single time.
What it costs and where to start
A working internal AI knowledge base for a 60-person Australian consultancy costs $180,000 to $400,000 AUD to build, and $50,000 to $120,000 a year to operate, with payback typically inside nine months against the recovered senior capacity. Start small: pick one practice area, build the curated front layer first, and prove retrieval quality on real questions from real consultants before widening the scope. The firms that fail are the ones that try to ingest everything on day one.
If your firm is sizing a knowledge base engagement, or wants an honest read on whether the files you already have can support one, book a brainstorm call with Automata AI and we will walk through the three-layer structure against your own material.
