AI Pilot Purgatory Has a Knowledge Problem

One number that explains the enterprise AI problem

Only 25% of enterprise organizations have moved 40% or more of their AI pilots into production, according to Deloitte's 2026 State of AI in the Enterprise report. Another 37% are still using AI "at a surface level with little or no change to core processes."

That's not a slow start. That's three quarters of enterprises stuck somewhere between experiment and deployment, burning budget while their most capable AI systems sit in demo purgatory.

McKinsey's numbers tell the same story from a different angle. 88% of organizations report using AI in at least one business function. But only about one-third say they've begun scaling AI programs. For agents specifically — the systems enterprises are betting on for real process automation — just 23% report scaling an agentic system anywhere in the enterprise, while 39% are still experimenting.

This is the AI pilot gap. Wide experimentation, thin production. And the conventional explanations — governance deficits, change management failures, ROI measurement problems — are real but incomplete.

Why the standard explanations miss the hardest part

The Databricks numbers are striking enough on their own. Companies using AI governance tools get over 12 times more AI projects into production. Organizations using evaluation tools move nearly 6 times more AI systems to production. The case for governance and evals is not subtle.

So why aren't more enterprises adopting them? And why do governance-mature organizations still report production failures at meaningful rates?

Banking shows where the real friction lives. According to a March 2026 analysis drawing on Boston Consulting Group data, globally one in four banks is actively using AI to gain competitive advantage — the rest remain in fragmented, inconclusive pilots. McKinsey puts the potential annual value addition to global banking at $340 billion. Yet the majority of the world's financial institutions cannot demonstrate in measurable terms what their AI investments are actually delivering.

India's banking and financial services sector is instructive. 94.1% of Indian BFSI firms claim to use AI to improve operational efficiency. Only 19.1% track whether it is contributing to revenue (WION). That 75-point gap between "we have AI" and "we know what it's doing" is not a measurement problem. It's a trust problem. And trust problems usually trace back to reliability problems.

When agents give wrong answers, enterprises stop trusting them. When they stop trusting them, they stop scaling them. The question is: why do agents give wrong answers in production when they worked well in pilots?

The fault line is production capacity, not model access

InfoWorld's recent analysis framed the enterprise reality as a split between fast-learning and slow-learning teams, rather than a smooth adoption curve (InfoWorld). A hedge fund engineering head with fleets of agents in full production. A data engineer at a large retail bank with no agents and sparse LLM use. Same era, same model access, radically different outcomes.

OpenAI's 2025 enterprise report is direct on the cause: the primary constraints are now organizational readiness and implementation, not model performance. The models are good enough. The infrastructure around them often is not.

Databricks' data on multi-agent deployments makes the stakes clearer. Multi-agent workflows grew 327% over the past year. As organizations shift from single-agent experiments to chained, interdependent systems, the margin for error shrinks. One unreliable agent in a workflow contaminates everything downstream.

The enterprises winning on production rates share a few traits: evaluation discipline, governance tooling, and — less discussed — a coherent knowledge foundation that their agents can actually read.

The under-reported blocker: what agents are reading

There's a trust stack in enterprise AI that most discussions cover incompletely. Governance gets attention. Evaluations get attention. Observability and rollback are starting to get attention. What gets almost no attention is the quality of the source documents agents retrieve from.

Consider what a production AI agent is actually doing when it handles a customer query, processes an onboarding document, or responds to a support ticket. It retrieves relevant content from an internal knowledge base — policies, SOPs, product specifications, compliance documentation — and generates a response grounded in that content. The model is not the variable. The source material is.

In most enterprises, that source material is in rough shape. SOPs contradict each other because they were updated at different times by different teams. Policies have version drift — the document someone uploaded two years ago is still in the retrieval index, alongside the current version. Internal knowledge is scattered across folders, formats, and systems, and only some of it is accessible to the agent at all.

This matters differently in production than in pilots. Pilots run on curated datasets. Production runs on the actual enterprise document estate — and that estate has rarely been audited, cleaned, or maintained with AI retrieval in mind. We've written about this problem from the readiness angle, but the production failure angle makes it more concrete: you can have world-class governance and evaluation tooling and still watch agents fail systematically because what they read is wrong.

When agents act on documents, the quality of those documents becomes execution risk — not just an answer quality problem. An agent that reads a stale returns policy doesn't give a bad customer experience; it potentially makes a bad business commitment. At scale, that's a different category of problem.

What production-ready AI actually requires

Governance and evaluations are necessary but not complete. The full production-readiness stack has a knowledge layer that sits beneath the better-known trust architecture.

Source attribution in retrieval. Every agent response should be traceable to the specific document it came from, and to the version of that document. This isn't just for auditability — it's the mechanism that lets you identify when retrieval is pulling from outdated sources. Without it, you're flying blind on why the agent got it wrong.

Contradiction detection across documents. Enterprise document estates grow by accretion. New policy goes in, old policy stays. New SOP supersedes old SOP, but old SOP never gets removed. The agent doesn't know which one to trust, so it picks one — often the wrong one. Scanning for contradictions before and after ingestion is not optional for production-grade systems.

Maintenance that keeps pace with operations. A knowledge base that was accurate at pilot time degrades over the production lifecycle. Organizational changes, regulatory updates, product revisions — all of it creates drift between what's in the system and what's actually true. In the ROI analysis we've seen across enterprise AI deployments, knowledge freshness consistently separates the systems that hold up over time from the ones that quietly degrade.

Permission-aware access. Agents operating in production need to know what they're allowed to show to whom. A knowledge base that doesn't enforce document-level permissions isn't production-ready, regardless of how accurate the retrieval is.

Support for the actual document estate. Most enterprises have a mix of clean digital documents and messy legacy content — scanned PDFs from five years ago, slides repurposed as policy documents, spreadsheets someone turned into a knowledge base because nothing else existed. Production-grade retrieval has to handle all of it, not just the clean subset that performed well in the pilot.

This is the layer Mojar (/) is built around. Not just retrieval, but governed retrieval — source attribution on every answer, contradiction detection across the full document set, document-level maintenance that keeps the knowledge base current, and a hybrid parsing approach that handles the kinds of messy enterprise documents most RAG systems quietly reject.

What to watch

The next 90 days will clarify which enterprises are actually closing the pilot gap versus which are running more pilots. Watch for production metrics, not adoption announcements.

Organizations tracking production rates rather than deployment counts will get there first. Those who treat governance as a project management layer while leaving the knowledge foundation unaddressed will keep watching their agents degrade between quarterly reviews.

The firms that move AI from demo to scale won't be the ones with the flashiest pilots. They'll be the ones who built a knowledge layer sturdy enough for agents to act on when nobody's watching.