How to build AI search over past consulting deliverables

What should AI search over deliverables actually do?#

AI search over consulting deliverables should let a consultant ask a question in plain language, such as whether the firm has redesigned a distribution network for a cold-chain business, and get a short answer that cites the exact slides and reports it drew on. It should also let them filter by sector, service line, date and engagement type.

The point is reuse of the firm's thinking: prior approaches, frameworks, analyses and lessons. Consultants searching a shared drive by file name or keyword often miss the right deck because it uses different words, and the knowledge sits idle while teams rebuild material from scratch.

The build follows six steps, with the confidentiality step placed early on purpose.

• 1Inventory where deliverables live and what formats they take.
• 2Exclude or restrict material that client contracts limit.
• 3Mirror source permissions so search never widens access.
• 4Ingest files with their structure and metadata.
• 5Answer with citations to the source page or slide.
• 6Collect feedback and keep the index current.

Step 1: inventory where deliverables live#

The deliverable inventory starts with every location a final report or deck might sit: SharePoint and Teams sites, Box or Google Drive, an old file server, PSA document links, proposal libraries and, too often, email attachments. Recorded client readouts and their transcripts belong on the list as well. Microsoft 365 firms should note a split that catches many teams out: Microsoft says files shared in a Teams chat are stored in the sharer's OneDrive, while files posted to a channel sit in the team's SharePoint site, so a deliverable sent in a chat may never reach the project site.

For each location, note the formats, the date range, who owns it and whether finals are separated from drafts. Drafts multiply quickly in consulting, and an index full of near-duplicate versions gives worse answers than a smaller set of finals.

Step 2: exclude what client contracts restrict#

The client-confidential exclusion step sorts content into tiers before anything is indexed. Many consulting MSAs limit use of client information to the engagement, some make deliverables client property and some require return or destruction at the end. Others allow internal reuse of know-how but not of client facts. Legal or the engagement partners need to tag clients by what their contracts allow.

Ethical walls matter too. If the firm serves competing clients in the same sector, deliverables for one should not surface for a team serving the other.

Sanitized case summaries are the bridge between tiers. A one-page write-up of the problem, approach and outcome, with the client described only by sector and size, can usually be searched firm-wide even when the underlying deck cannot. Writing them at engagement close, while the team still remembers the work, is far cheaper than reconstructing them later from restricted files.

Step 3: mirror permissions before indexing#

Permissions in AI search should mirror the source system exactly, so the tool never shows a passage the user could not open directly. The common failure is indexing everything through a broad service account and filtering later, which lets summaries leak content from files a user has no right to see.

Prefer tools that check the user's access at query time against the source system, or that copy access lists at index time and refresh them often. Test with real accounts from different practices before launch, including someone deliberately outside a walled engagement.

Step 4: ingest decks, PDFs and models with their structure#

Ingestion for consulting material has to respect structure. A slide's meaning sits in its title, body text, speaker notes, chart labels and table cells, so treat each slide as its own chunk and keep those parts together. PDFs need page-level chunks and text extraction that copes with multi-column layouts and scanned pages.

Spreadsheet models are a special case: index the model's description, tab names and key assumptions rather than thousands of cells. Attach metadata to every chunk, including sector, service line, engagement date, tier and file location, so filters and citations work.

Steps 5 and 6: answer with citations, then learn from feedback#

Answers should come from retrieved passages and cite the file and slide or page for each claim, which is the core of retrieval-augmented generation. When the index holds nothing relevant, the tool should say so rather than compose a plausible answer from general knowledge.

Feedback keeps the system honest. Let users flag answers as wrong, stale or too sensitive, route flags to a named content owner and review search logs for questions that found nothing. Those gaps show where the firm's knowledge is missing or badly labeled.

Illustrative: an operations consulting firm builds search across two archives#

Illustrative: a fictional operations consulting firm keeps recent deliverables on SharePoint and older work on a file server that partners still dig through by memory. Consultants keep rebuilding warehouse layout analyses the firm has already done.

The COO and the knowledge manager inventory both archives, retire drafts and tag clients by contract type with help from counsel. Deliverables for clients whose agreements bar reuse are excluded, a walled sector practice gets its own restricted tier and methods and templates go firm-wide. The firm chooses an enterprise search tool that checks SharePoint permissions at query time, chunks decks by slide and cites slide numbers. Practice leads review flagged answers on a regular cadence.

Where internal search ends and data licensing begins#

Internal AI search is the firm using its own records for its own people; licensing records to an outside AI developer is a separate decision with a separate rights review. A deliverable that is fine to search internally may still be client property that cannot be licensed.

SourceX does not build search tools. When a firm considers licensing, SourceX starts with Supply in the SourceX five-step transaction, and the same inventory and tiers help: firm-owned methods, templates and SOPs are more often in scope than client deliverables, which are usually carved out at Rights. Any package that proceeds is documented in a SourceX Evidence Packet.