Your search tools work. Your knowledge management still costs a day a week.

Knowledge management is the discipline of making an organization's information findable, current, governed, and reusable, so that the right answer reaches the right person without a hunt. Enterprise search is one component of that. It retrieves documents. Knowledge management decides which documents exist, which are authoritative, who owns them, and how they stay accurate over time. The distinction matters because the market has already split. The AI-specific knowledge management segment grew 47.2% to $7.71 billion in 2025, while the broader knowledge management software category grew at roughly 13.8% a year toward a projected $74.22 billion by 2034 ^[11][2]. That divergence signals a platform transition, not simple category expansion. AI-powered approaches are the growth engine, and the organizations still treating knowledge as a static file store are the ones falling behind.

APQC repositioned knowledge management in its 2026 research as a strategic capability rather than a support function ^[3]. The reframe reflects what executives are discovering the hard way. You cannot deploy a capable model onto a fragmented knowledge base and expect a capable result. The model inherits whatever it is fed.

Tricky Wombat made with Google Gemini 3.1 Flash Image, Sep 2026

Source

What does the data show about time lost to search?

The foundational measurement has held for over a decade. Employees spend close to a fifth of the working week searching for and gathering information ^[1]. What has changed is that more tools have not fixed it. In Coveo's 2025 Employee Experience Relevance Report, based on a survey of 4,000 workers at companies with 5,000 or more employees, 42% of the information employees encountered was rated irrelevant to their role, workers searched an average of four different systems, and they spent close to three hours a day doing it ^[5]. Frustration with unhelpful tools reached 47% in 2025, up 12 percentage points from the prior year ^[5]. The trend line runs the wrong way despite rising technology budgets.

Translate the McKinsey figure into money. A 1,000-person organization at a $60,000 average salary loses roughly $12 million a year in productive capacity to information search alone ^[4]. That is the equivalent of hiring five people and assigning one of them to do nothing but look for files every day.

Tricky Wombat made with Google Gemini 3.1 Flash Image, Sep 2026

Source

What are practitioners reporting?

The people inside these systems have already adapted. In the same Coveo research, 49% of employees said they had encountered AI hallucinations in their workplace tools, and one in four did not know where to start when they needed to find information ^[5]. Workers are fact-checking machines by default now, because they have learned the answer might be wrong.

The executive view shows the same tension from a different angle. Deloitte's 2026 enterprise AI research found that two-thirds of organizations report productivity gains as the top benefit of AI, yet only 34% said they had genuinely reimagined a business process. The majority were optimizing what already existed, and data-related issues remained the number-one challenge, cited by 48% of respondents ^[12]. High satisfaction and a data problem sitting underneath it. That combination is the tell. Most organizations are running AI as a faster version of the status quo, and the status quo has a broken knowledge layer.

Statistics describe the shape of the problem. Case studies show the mechanism. The organizations below span automotive, consumer goods, and medical technology, on three continents, at wildly different sizes. What connects them is not the tool each chose. It is what they did, or failed to do, with the knowledge underneath.

Reckitt: replacing 200,000 scattered files with one entry point

Reckitt, the global consumer goods company behind Dettol, Lysol, and Nurofen, ran its insights function across more than 5,500 SharePoint folders and assorted internal systems, with no unified way to search across them ^[13]. Every day, teams commissioned consumer research without a reliable way to check whether the same study already existed somewhere in the archive, so they duplicated work. Then the company consolidated its insights knowledge base onto a single platform, migrating roughly 200,000 files into one entry point that used AI to surface not just documents but specific paragraphs, sentences, and slides. Because of that, research duplication fell across global teams, and the base has since grown past 400,000 files with about 2,000 added weekly ^[13]. The takeaway is an economics that rarely appears in ROI models: the cost of not finding what already exists. 200,000 files is not a knowledge base. It is an archive. The difference between the two is findability.

Siemens Healthineers: building an AI agent from ticket history

Siemens Healthineers, the medical technology division of Siemens, ran a global HR service function on more than 850 knowledge articles across eight languages, resolved by first-level agents doing manual lookups ^[14]. The normal pattern was a person reading a knowledge article to answer a routine request, one ticket at a time. In a single day-long internal hackathon in summer 2025, business analysts and solution architects built a prototype agent trained on actual ticket history and those knowledge articles, capable of reclassifying, routing, and resolving inquiries autonomously when the answer existed in the base ^[14]. The instructive detail is the ratio. The prototype took one day. The 850 documented articles, structured case histories, and multilingual content it depended on took years. The AI was the easy part.

An international goods manufacturer: when knowledge lives in inboxes

An international goods manufacturer operating in more than 12 countries across the Americas kept its technical process documentation on local hard drives, shared by email, with competing versions and no authoritative source ^[15]. Every day, an engineer needing a procedure risked pulling an outdated file, and the stakes were operational, not clerical. Manufacturing downtime in the sector runs between $36,000 an hour in fast-moving consumer goods and $2.3 million an hour in automotive ^[15]. Because the institutional knowledge needed to respond quickly was eroding as experienced workers retired, the company's average recovery time from disruptions rose 65% over five years, from 49 to 81 minutes per incident ^[15]. The consulting team implemented a content model, designated repositories by knowledge type, metadata enrichment, and a structured offboarding procedure to capture departing experts' knowledge before it walked out the door. The lesson lands without a chart: recovery time got worse not because the equipment aged, but because the knowledge did.

What patterns emerge across these cases?

Three different industries, one shared diagnosis. In each, the technology was available and the knowledge was not usable. Reckitt had the files and could not find them. Siemens Healthineers had the articles and could not apply them at scale. The goods manufacturer had the expertise and could not keep it. The systemic data backs the pattern. IBM research found that 68% of enterprise data remains unanalyzed and inaccessible to AI systems, Eptura's 2025 Workplace Index found only 4% of organizations run fully integrated systems, and 82% of enterprises report workflow disruptions caused by siloed data ^[6]. Fragmentation is not the exception. It is the default state of enterprise information, and it is the exact condition in which AI produces its worst results.

Tricky Wombat made with Google Gemini 3.1 Flash Image, Sep 2026

Source

Zoom out from single cases and a spread appears. The same technology, deployed by different organizations, produces results that differ by an order of magnitude. That spread is the whole argument, because it means the technology is not the variable.

What do users and practitioners consistently report?

Across survey after survey, the reported themes converge. Employees encounter irrelevant information, do not trust what they find, and burn hours verifying AI output ^[5][5]. Executives report data quality as their leading obstacle ^[12]. Knowledge management practitioners, per APQC, are under pressure to move from counting activities to proving business outcomes, and they name AI integration as the force reshaping their priorities ^[3]. Different populations, same underlying complaint: the knowledge is there, and it is not usable.

What drives the gap between strong and weak outcomes?

Consider the arithmetic of hidden cost. A 1% hallucination rate sounds tolerable until you scale it. Across 1,000 employees each asking an AI tool 10 questions a day, that is 100 fabricated answers every day, 500 a week, and a workforce learning not to trust the system. That distrust is expensive. The average AI user now spends 4.3 hours a week verifying output, which annualizes to about $14,200 per employee, roughly $7.1 million a year for a 500-person firm spent checking answers rather than acting on them ^[8]. Layer that on top of IDC's estimate of $5,700 per worker per year in baseline productivity loss from poor knowledge management ^[7], and the compounding is obvious. Organizations that fix the knowledge layer cut both numbers. Organizations that do not pay both, and pay them every year, while the AI tools they bought quietly make the trust problem worse.

Here is the pattern every previous section has been circling. The technology is rarely the variable. What separates the organizations getting returns from the ones abandoning pilots is a decision made before any model is selected: whether the knowledge was made findable, current, and governed first.

The evidence is direct. MIT Project NANDA analyzed more than 300 enterprise generative AI initiatives and found that roughly 95% showed no measurable profit-and-loss impact, against Gartner's projection of $644 billion in generative AI spending for 2025 ^[16][6]. The diagnosis was not model failure. It was unstructured, ungoverned knowledge feeding capable models. The same causal story appears at the retrieval level. Syntheses of production deployments find that hallucination in production is overwhelmingly a retrieval problem, not a generation problem.

AI does not create knowledge quality. It inherits it, then amplifies it in both directions.

That is why Gartner predicts 60% of AI projects will be abandoned through 2026 for lack of AI-ready data, not for lack of model capability ^[16]. The bottleneck moved. It is no longer the model. It is what you feed it.

Tricky Wombat made with Google Gemini 3.1 Flash Image, Sep 2026

Source

What does a successful implementation look like?

Givaudan, the world's largest maker of flavors and fragrances with more than 16,900 employees, gives the clearest proof ^[17]. Its European insights team held research fragmented across external platforms, internal systems, personal folders, and colleagues' memories, and when a project arrived, teams searched disconnected repositories and often resorted to inference when the evidence was not accessible in time. Givaudan adopted an AI-powered insights platform built on a design its team nicknamed "The Genie," where every AI-generated answer is grounded in Givaudan's own validated research and linked back to specific source documents, so users can audit the inputs and challenge the conclusion ^[17]. Because the answers were auditable, the team trusted them, and because they trusted them, they used them. Decision-making time fell 3x on many projects and up to 10x in specific project categories, work that took weeks now finishes in days or hours, and the team supports more client briefs with the same headcount ^[17].

Set that against Reckitt, which deployed a platform from the same vendor and saw a real but far more modest lift in research efficiency ^[13]. The technology was comparable. The distinguishing factor was the grounded, source-linked design that built enough trust to drive full adoption. That contrast is the thesis in miniature, and it holds without a single word of comparison.

The reframe also scales down. A mid-sized Class II medical device maker, published under the name New Horizon Biotech, built an AI knowledge infrastructure with mandatory source citation, groundedness monitoring, and human approval gates ^[18]. In a six-month pilot the organization reported evidence retrieval time falling 88%, from 12.5 hours to 1.5 hours, an AI groundedness score of 0.94, and a 6-month ROI of 82.4% ^[18]. Present those figures as reported by the publication rather than as audited results, and the point stands: the governed knowledge layer, not the model, produced the return, and it did so for a company nowhere near enterprise scale.

Both sides of the ledger are calculable, and both are larger than most executives assume. On the cost side, AI hallucinations were estimated to cost businesses $67.4 billion globally in 2024, split across $18.2 billion in direct losses, $21.5 billion in operational cleanup, and $27.7 billion in reputational damage, with the total projected to reach $112 billion in 2025 ^[8]. Treat that as a single-source industry estimate rather than an audited figure, and it still frames the scale. Add the quieter costs: $5,700 per worker per year in lost productivity ^[7], and 60% to 80% of AI project time consumed by data preparation before a model is even chosen ^[9]. Most of your AI budget is spent on knowledge work whether you plan for it or not. The only question is whether that spend is deliberate or accidental.

The return side is equally concrete. A Forrester Total Economic Impact study of a composite enterprise deploying an AI-powered market intelligence platform found a 411% three-year ROI, total benefits of $8.93 million against $1.75 million in costs, and a net present value of $7.18 million ^[19]. The benefits decomposed into speed-to-answer efficiency, a 27% reduction in duplicate research studies, legacy tool consolidation, and 3% incremental revenue growth in targeted markets ^[19]. Both sides of the ledger trace to the same variable. The cost of poor knowledge management and the return on good knowledge management are the same number measured from opposite directions, and that number is set by the infrastructure, not the tool.

Tricky Wombat made with Google Gemini 3.1 Flash Image, Sep 2026

Source

The pattern across every case is the same. The organizations getting returns built a knowledge layer that is findable, current, and grounded before they scaled AI on top of it. That is what Tricky Wombat builds. We frame the work around the pipeline that feeds the model, because the pipeline is what determines the output. A capable model on a broken pipeline produces confident wrong answers. The same model on a sound pipeline produces auditable right ones. Three things decide which one you get.

1. Structure and govern the knowledge before it reaches the model

Most systems point a model at a document store and call it retrieval. The store contains competing versions, orphaned files, and content no one has owned in years, so the model retrieves whatever is closest in vector space, current or not. Tricky Wombat builds a governed knowledge layer first: a taxonomy and metadata model, assigned content ownership, and version control that marks which document is authoritative. Data preparation is 60% to 80% of the work in any real AI project ^[9], so we treat it as the project rather than as setup.

2. Retrieve at the right granularity, not the whole document

Most systems retrieve entire documents and hand the model a wall of text, where the relevant sentence is buried among ten pages of context and the model has to guess. Tricky Wombat retrieves at the level of the passage, the paragraph, and the specific claim, so the model receives the precise evidence rather than the whole file. Retrieval quality is where production hallucination is won or lost, and granularity is a large part of retrieval quality.

3. Ground every answer in a citation the user can open

Most systems return an answer and ask the user to trust it, which is exactly how the fact-checking tax appears and why 49% of employees report encountering hallucinations ^[5]. Tricky Wombat links every generated answer back to the validated source it came from, so a user can open the citation and verify it in one click. This is the design that separated Givaudan's 10x result from a modest lift: auditable answers earn trust, and trust drives the adoption that produces the return ^[17].

The pipeline is not a one-time build. Tricky Wombat monitors the knowledge base continuously, re-processes content as sources change, and verifies that citations still resolve to live, correct material. Because the system watches for stale and orphaned content and flags it for owners, the knowledge layer gets more accurate over time rather than decaying the way an email-and-hard-drive archive does. The organizations that win are not the ones that deployed the newest model. They are the ones whose knowledge layer improves every quarter.

Two companies bought comparable AI platforms and got results an order of magnitude apart. A manufacturer's recovery time got worse while its equipment stayed the same. Roughly 95% of AI pilots produced no measurable return against hundreds of billions in spending. Every one of these outcomes traces to the same variable, and it is not the model. It is whether the knowledge feeding the model was made findable, current, and governed before anyone pressed deploy.

This is a sequencing decision, and the sequence is not optional. APQC's 2026 research found that the organizations distinguishing themselves at AI treated knowledge architecture as prerequisite infrastructure, built before deployment, not patched afterward ^[10]. The ones that skip that step are not saving time. They are moving the cost downstream, where it shows up as abandoned pilots, a fact-checking tax, and a workforce that has quietly stopped trusting the tools.

So the question was never whether your company needs knowledge management. Your people already prove it, one lost day a week at a time. The question is whether you build the knowledge layer before your next AI project stalls, or after.