Building knowledge management that people actually use

# Building knowledge management that people actually use

A mid-size accounting firm solved a thorny tax problem for a manufacturing client: how to treat a cross-border R&D credit under a specific state's rules. Eighteen months later, a different partner in a different office solved the same problem from scratch. Nine months after that, a third team did it again. Three times, three sets of billable research hours, one answer that already existed. It was sitting in a partner's sent folder.

This is the core failure of knowledge management (KM) in professional services: expertise gets created, then buried. The firm pays for the same intellectual work repeatedly, and when a senior person leaves, their knowledge walks out with them.

The good news: KM is a data problem, and data problems can be instrumented and measured. This lesson shows how.

Why most KM systems fail

Firms buy a document repository, mandate that people upload their work, and declare victory. Six months later the repository is a junk drawer: 4,000 documents, no tags, unsearchable, untrusted.

The failure is not the tool. It is that KM was treated as storage, not as a reuse system. Storage asks "where do we put things?" A reuse system asks "how does someone find and trust the right prior work in under two minutes, at the moment they need it?"

Those are completely different design goals.

The three things that must be true

For knowledge to compound instead of evaporate, three conditions have to hold:

1. Capture is cheap. If uploading and tagging takes more than a minute, people skip it under deadline pressure. They always skip it.

2. Retrieval is fast and trusted. A search that returns 200 results, half outdated, is worse than useless. People stop trusting it and revert to asking colleagues.

3. Reuse is measured. If you cannot see whether knowledge gets reused, you cannot improve the system or justify the investment.

Most firms nail zero of the three.

Instrumenting knowledge as data

Think of every reusable work product (a tax memo, a due diligence checklist, a client deliverable template) as a record with metadata attached. Metadata is just "data about the data": who made it, when, for what industry, under which rules, and how good it is.

Here is a minimal metadata schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.View full definition → for that accounting firm's tax memos:

memo_id:        TX-2026-0417
title:          State R&D credit treatment, cross-border
practice_area:  Tax
industry:       Manufacturing
jurisdiction:   [State-specific]
topic_tags:     [r&d-credit, cross-border, state-conformity]
author:         partner_id_882
created_date:   2026-03-11
reviewed_by:    partner_id_140
confidence:     High (partner-reviewed)
reuse_count:    0
last_reused:    null

Two fields do the heavy lifting.

`confidence` tells a searcher whether they can trust the memo without redoing the research. "Partner-reviewed" means reusable now. "Draft, unreviewed" means proceed with caution. This single field is the difference between a repository people trust and one they ignore.

`reuse_count` is your instrumentation. Every time someone opens, cites, or copies the memo into new work, the counter increments. Now reuse is a measurable event, not a hope.

Make capture automatic, not manual

The trap is asking humans to fill in all that metadata by hand. They will not do it consistently.

Pull what you can automatically:

• Author, date, practice area, office: from the document system and the person's profile.
• Industry and client type: from the client-matter record in your practice management system.
• Topic tags: increasingly auto-suggested by AI. A large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → (a type of AI trained on text) can read a memo and propose tags like r&d-credit and cross-border, which the author confirms in one click.

The human should only confirm and set confidence. That gets capture under the one-minute threshold.

Retrieval that beats asking a colleague

The benchmark for KM search is not "better than nothing." It is "faster and more trusted than walking over to ask a senior colleague." That is a high bar, because asking a person is fast and the answer comes pre-vetted.

To beat it:

Rank by confidence and recency, not just keyword match. A partner-reviewed memo from 2025 should outrank a raw draft from 2021, even if the draft mentions the search term more often.

Show the trust signals in the result. Reviewer name, date, and reuse count should appear in the search result itself, so the searcher can judge before clicking.

Use semantic search where you can. Traditional keyword search misses a memo titled "cross-border research incentives" when someone searches "R&D tax credit foreign." Semantic search matches on meaning, not exact words. Many KM and document platforms in 2026 offer this out of the box, often called "AI search" or "retrieval-augmented" features.

A caution on AI: retrieval-augmented generation (RAG), where an AI drafts an answer from your firm's documents, is powerful but only as trustworthy as the underlying documents and their confidence tags. Garbage in, confident-sounding garbage out. Instrument confidence first, then layer AI on top.

For a solid, vendor-neutral primer on organizing knowledge work, the Atlassian guide to knowledge management is a free and practical starting point.

Measuring whether it actually works

Here is where most firms stop paying attention, and where the value actually lives.

The one metric that matters most: reuse rate

Reuse rate = (number of work products reused at least once) / (total work products captured).

If you captured 500 memos and 40 were ever reused, your reuse rate is 8 percent. That tells you 92 percent of your captured knowledge is dead weight. It also tells you your tagging or search is probably broken, because good work is not being found.

Track it monthly. A healthy, well-instrumented system should show reuse climbing as your tag quality and search improve.

Supporting metrics

• Time-to-find: how long from search to opening a useful result. If it creeps above two minutes, people abandon the system.
• Top reused items: your "greatest hits." These reveal what expertise the firm leans on most, and what deserves a polished, template-ized version.
• Orphan knowledge: high-confidence work from people who have since left the firm. This is knowledge you already captured before it walked out the door. Its reuse count proves the system paid for itself.

A simple ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → story

Say a tax memo takes an estimated 6 hours to research from scratch. If a well-tagged, reviewed memo gets reused 5 times, you avoided roughly 30 hours of duplicate work. Multiply across a practice and the KM system stops being an IT cost and becomes a margin lever. Note these are illustrative figures; measure your own.

Rolling it out without a revolt

Technology is the easy part. Adoption is where KM initiatives die.

Start with one practice group and one document type. Do not try to instrument the whole firm at once. Pick, say, the tax group's memos. Prove the reuse rate goes up. Then expand.

Recruit the busiest senior person as a champion. If the partner everyone respects uses the system and cites it in a meeting ("I pulled this from our KM base, saved me a day"), adoption follows. Mandates from IT do not move behavior. Peer signals do.

Reward reuse, not just contribution. Many firms reward people for uploading documents, which produces the junk drawer. Instead, spotlight the memos with the highest reuse counts and credit their authors. This rewards writing something genuinely reusable, and it rewards the reviewer who tagged it "high confidence."

Close the loop on retrieval failures. When someone searches and finds nothing useful, capture that. A search that returns zero good results is a signal: either the knowledge does not exist yet (create it) or it exists but is mistagged (fix the metadata).

Watch the confidentiality line

Professional services work is full of client-confidential material and, for lawyers, privileged information (legally protected communications between attorney and client). Your KM schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.View full definition → needs a confidentiality field and access controls so a reusable *template* is shareable firm-wide while *client-specific* content stays locked to the engagement team. Instrument this from day one. Retrofitting it after a leak is painful.

Key Takeaways

• KM is a reuse system, not storage. Design for fast, trusted retrieval at the moment of need, not for filing documents away.
• Instrument two fields above all: `confidence` and `reuse_count`. Confidence makes knowledge trustworthy enough to reuse; reuse count makes value measurable.
• Automate capture so it takes under a minute. Pull metadata from your systems and let AI suggest tags. Humans should only confirm and rate confidence.
• Track reuse rate monthly. A low reuse rate means your knowledge is dead weight or your search is broken. Both are fixable, and both cost you billable hours.
• Win adoption through peer champions and by rewarding reuse, not uploads. Mandates create junk drawers; visible reuse creates a system people trust.