How a consultancy uses professional services AI automation to stop rewriting proposals

The knowledge a consultancy keeps losing

A consultancy sells expertise, then struggles to find it again. Every engagement leaves behind a proposal, a methodology, a set of slides and a review of what worked. That body of work is the firm’s real asset, and most of it is sitting somewhere in Notion, in an old deck, or in the memory of whoever ran the project. The next proposal lands, and a consultant opens a blank page anyway.

The cost is quiet but constant. A junior consultant spends most of a day looking for the pricing model the firm used on a similar scope last year, then rebuilds it from a half-remembered version. A partner gets interrupted for the third time that week to point at a document they know exists. A proposal goes out reusing an approach the firm has already refined twice, except this version misses the refinements because nobody could find them. The work was done. It just was not reachable.

Search does not fix this on its own. Notion search finds documents with the right words in them, not the right answer to a question asked in a consultant’s own language. It cannot tell a current methodology from a superseded one, and it happily returns a stale proposal next to the good one with no way to know which is which. For a firm whose product is knowing the answer, not being able to find your own answer is an expensive problem. It also carries a real risk, because the material being searched includes confidential client engagement detail that has to stay controlled.

A knowledge base grounded in the firm’s own documents

The version that works is a knowledge base built on retrieval-augmented generation (RAG), sitting over the firm’s own content in Notion. A consultant asks a question in plain language. The system retrieves the most relevant passages from the firm’s proposals, methodologies and past engagements, and an OpenAI GPT model turns them into a direct answer that cites the documents it came from. The consultant gets the answer and a link to the source, so they can verify it before it goes near a client.

RAG is the right core here, and the reason is specific to this problem. The obvious alternative is to fine-tune a model on the firm’s documents, but fine-tuning bakes the knowledge in at a point in time. The day a methodology is updated or a proposal is retired, a fine-tuned model is quietly wrong and has no way to tell you. Retrieval keeps the knowledge in the documents, where the firm already manages it. Update the Notion page and the next answer reflects it. Just as important, retrieval lets every answer cite its source, which is what makes the output trustworthy enough to reuse. A model that asserts an approach with no provenance is no safer than guessing.

The supporting pieces sit around that core and stay deliberately plain. Notion remains the source of truth, so consultants keep working where they already work. The GPT model is held on a tight rein, instructed to answer only from retrieved passages and to say when the material is missing rather than fill the gap. Access follows the firm’s existing permissions, and client-identifying detail is tagged and controlled so the knowledge base respects confidentiality and any NDAs, with personal information handled under the Privacy Act 1988.

Building it, and where it got hard

The model was rarely the hard part. The friction lived in the source material, and one problem stood in for most of the rest.

The documents in Notion were messier than anyone wanted to admit. There were three versions of the same methodology with no marker for which was current, proposals from engagements the firm no longer ran that way, and review notes that contradicted the polished method they were meant to support. Pointed at that, the knowledge base retrieved confidently and answered wrongly. It would surface a two-year-old pricing approach with the same certainty as the current one, which is worse than no answer, because a consultant might reuse it.

The fix was curation, not a cleverer prompt. We worked with the firm to mark a single source of truth for each methodology, tag superseded material so retrieval could down-weight it, and weight recent documents ahead of old ones where two answers competed. Then we leaned on the one feature that makes the whole thing safe to trust. Every answer cites its source, so when the system is unsure or the material is thin, it says so and shows the consultant where to look rather than inventing a tidy reply. A human can always open the cited page and check. That combination, curated sources plus visible provenance, turned a knowledge base people did not trust into one they reached for first.

Confidentiality shaped the rest. Proposal reuse had to draw on the firm’s structure, pricing logic and method, not on a former client’s confidential specifics, so the system reuses how the firm works rather than copying what a particular engagement contained. Client-identifying detail stayed controlled behind the firm’s permissions, which kept the build aligned with the confidentiality and privacy obligations the firm already carries.

What changed

In a representative build, the time to a usable proposal first draft fell by roughly half, because consultants started from the firm’s most relevant past work instead of a blank page. Questions that used to mean interrupting a partner or digging through old folders were answered from the knowledge base in seconds, each with its source document attached. New starters reached useful output sooner, asking the system what the firm’s method was rather than waiting for a senior consultant to free up an hour.

These figures are illustrative. They describe the pattern we see rather than a published result for a named firm. The shape is the point. The knowledge the firm had already paid to create starts reaching the next proposal while it still matters, the senior people stop being a lookup service for their own documents, and a junior consultant can find the firm’s best answer instead of rebuilding a worse one.

Where this fits

A grounded knowledge base is one application of our Process Optimisation service, built on a retrieval-augmented core, for the realities of an Australian professional services firm. It is a contained, high-return place to start, because the knowledge already exists and the value comes from making it reachable and trustworthy rather than from anything exotic. If your team keeps rewriting work it has already done, the place to start is to map where your proposals and methods live and decide which questions the firm should never have to answer twice.