Most AI initiatives fail for a reason that has nothing to do with the AI itself. Eighty percent of AI projects fail to deliver their intended outcomes, and only about 30 percent of AI pilots progress beyond the pilot stage. Given how capable current AI tools genuinely are, that failure rate points to something upstream of the technology: organizations are implementing AI before they’ve done the groundwork AI actually depends on.
That groundwork is knowledge management. AI doesn’t create knowledge, it retrieves, summarizes, and reasons over what already exists in your organization. If what exists is outdated, contradictory, or missing entirely, AI doesn’t fix that. It just delivers the mess faster and with more apparent confidence. Here is what needs to happen before implementation, not after.
1. Audit what your AI will actually be trained on
Start with a real inventory, not an assumption. Most organizations believe their documentation is “mostly fine” until someone actually counts. You’re looking for four things specifically:
- How much content is outdated (policies, procedures, or product information that no longer reflects reality)
- How much is duplicated or contradictory (two documents giving different answers to the same question)
- How much critical knowledge was never written down at all, and exists only in the heads of a few experienced people
- How much is scattered across systems that don’t talk to each other (a wiki here, a shared drive there, Slack threads nobody can search)
This audit alone often takes weeks, and it should. AI systems don’t distinguish between your best content and your worst. They treat everything in scope as equally authoritative unless you’ve explicitly told the system otherwise, which is impossible to do if you don’t know what’s actually in there.
2. Fix content quality before you upgrade the tool stack
There’s a strong pull toward solving quality problems with better technology: a smarter retrieval system, a larger model, more sophisticated prompt engineering. This is the wrong order of operations. Poor data quality weakens AI performance regardless of how advanced the system built on top of it is. A retrieval-augmented generation pipeline pulling from a disorganized knowledge base doesn’t produce better answers than a simple search would; it just produces confidently wrong answers faster. Content quality work, deduplication, updates, clear ownership of what’s current, has to happen first, or the AI layer inherits every existing problem and hides it behind fluent language.
3. Name a single accountable owner
AI readiness efforts fail quietly and predictably when responsibility is spread across a committee or absorbed into someone’s existing job as an add-on. What actually works is an internal owner with dedicated time and a clear mandate, someone whose job it is to drive this, with the authority to make calls and the time to follow through. Committees are good at producing documents. They are much worse at making the dozens of small decisions readiness actually requires: what gets prioritized, what gets cut, what “good enough” means for this rollout.
4. Involve the people who hold knowledge no document contains
Readiness assessments are often run as a leadership or IT exercise, disconnected from the people actually doing the work. This is a mistake, because the people who currently do the work hold knowledge no document contains, and excluding them loses that knowledge while gaining resistance instead. The employee who knows which supplier consistently sends the wrong paperwork, or which client relationship has an unwritten exception to standard policy, is sitting on exactly the kind of tacit knowledge that AI systems can’t infer and documentation rarely captures. Bring them in before rollout, not after complaints start.
5. Decide governance before you decide tools
Governance is not the policy document you write after go-live to cover yourself. It’s the set of decisions that determines whether “ready” means anything at all. Specifically, before implementation you need clear answers to:
- Who verifies AI-generated output before it reaches a real decision
- Who is accountable when that output is wrong
- Who has access to query what, and why that access is structured the way it is
- Whose knowledge is well-represented in the system, and whose is likely to be underrepresented or missing
These are organizational decisions. They don’t live inside a platform’s settings menu, and no vendor can make them for you.
6. Score your readiness honestly, not optimistically
Every gap you round up or explain away during a readiness assessment becomes a cost you pay later, usually at a worse moment and in front of more stakeholders. The discipline here is straightforward but uncomfortable: score your data and process condition as if a skeptic were auditing it, because eventually one will be. Vendor-led assessments in particular tend to conclude that you’re ready for that vendor’s product; keep the readiness judgment yours, and use vendor input only for questions of technical feasibility.
Why this matters more than tool selection
Most enterprise AI capability is now broadly comparable across vendors. What actually separates the roughly 30 percent of pilots that survive from the 70 percent that don’t is rarely the sophistication of the model. It’s whether the organization did the unglamorous work first: a clean, current knowledge base, one named owner instead of a diffuse committee, frontline expertise included rather than assumed, governance decided in advance, and an honest, sometimes uncomfortable, assessment of where things actually stand.
None of that work is fast. But it’s the difference between AI that genuinely improves how your organization thinks and decides, and AI that just makes existing problems move faster.
Stephanie Barnes explores this readiness gap in depth in Knowledge Management in the Age of AI: Keeping Humans at the Centre, a workshop on 16 September 2026 grounded in Entelechy’s Radical KM framework.