Your AI Agent Is Only as Good as Its Data: The Grey Data Crisis

May 27, 2026 · 10 min read · Spuutr Insights

Enterprise AI search is a $7 billion mirage.

Glean, the category leader, hit $200M ARR in February 2026 and was valued at $7.2 billion. Its pitch is seductive: "Google for your company." You connect Slack, Confluence, Google Drive, Jira, Salesforce and one AI assistant can answer any question about your business. It has over 100 connectors. Gartner ranks it as a leader in enterprise search.

But here is the uncomfortable truth hiding beneath those numbers.

Glean's architecture hinges on indexing everything. Every Confluence page. Every Google Doc, whether finalised or in draft. Every Slack thread, whether it contains an approved policy or a tired engineer's 2 AM hot take. Every Jira ticket comment.

And in February 2026, Glean's CEO Arvind Jain announced the company was pivoting from an enterprise search tool to a "middleware layer for enterprise AI." The subtext is unmistakable: even the market leader knows that search-by-index is not enough. The hard part has never been finding data. The hard part is trusting it.

The Grey Data Crisis

Let us be precise about what "enterprise knowledge" actually looks like in 2026.

In a typical Fortune 2000 company, McKinsey estimates only 3 to 7% of enterprise data is suitable for supporting enterprise-grade AI. At most seven cents of every data dollar you are spending on AI infrastructure is usable. The other 93% is noise, redundancy, dead documents, outdated policies, conflicting information and increasingly AI-generated slop.

BARC's 2025 research across 421 global organisations confirms this trend is accelerating. In 2024, only 19% of organisations cited data quality as a top AI challenge. By 2025, that number had more than doubled to 44%, making it the single biggest reported roadblock to AI project success.

This is the grey data crisis. Not "good" data you would stake a decision on. Not "bad" data you would obviously filter out. Grey data sits in the middle plausible enough to pass a quick skim, wrong enough to quietly corrupt every decision downstream.

The AI Slop Feedback Loop

The problem compounds exponentially because AI content itself is now flooding enterprise systems. TechTarget's Sean Michael Kerner identified this phenomenon: "the accumulation of low-quality, unverified AI-generated content flowing into corporate systems." Merriam-Webster made "slop" its 2025 Word of the Year.

Here is the terrifying feedback loop in practice, described by Adnan Masood, chief AI architect at UST:

"I have seen teams auto-draft FAQs and knowledge base articles, ship them, and then feed those same pages back into RAG as retrieval sources. A month later, you are no longer retrieving trusted institutional knowledge. You are retrieving yesterday's synthetic filler. The model did not get worse. Your knowledge substrate did."

This is not theoretical. It is happening in every organisation rolling out AI tools without an explicit curation layer. Every auto-generated Confluence page. Every AI-drafted Slack summary. Every synthetic FAQ. They all get slurped back into the RAG pipeline, and suddenly your enterprise AI is confidently answering questions based on content an LLM hallucinated last week.

The Authority Mimicry Problem

The really insidious property of AI-generated grey data: it sounds authoritative. It uses confident language. It structures itself like real documentation. It lacks the typos or hedge words that signal low quality.

This is what Masood means when he says: "The scary part is not that it can be wrong. It is that it can be wrong beautifully, at scale, and with enough confidence that people stop double-checking."

Traditional data quality metrics completeness, accuracy, timeliness, consistency collapse when the text looks perfect but the facts are fabricated. You would need human review of every single document to catch the hallucinations. At enterprise scale, that is not a process. It is a fantasy.

The Evidence: 96.6% of AI Investments Fail

The numbers behind this crisis are staggering:

RAND Corporation analysed thousands of AI initiatives and found that 80.3% fail to deliver their intended business value. MIT's NANDA lab put it more bluntly: 95% of generative AI pilots fail to scale beyond proof-of-concept. When Digital Applied looked specifically at AI agent projects the autonomous systems companies are rushing to deploy they found 88% never reach production.

The IBM Institute for Business Value reported that 43% of chief operations officers identify data quality issues as their most significant data priority. Over a quarter of organisations estimate they lose more than $5 million annually due to poor data quality, with 7% reporting losses of $25 million or more.

Gartner surveyed 782 infrastructure and operations leaders in April 2026. Their finding: only 28% of AI use cases fully succeed and meet ROI targets.

The math is brutal. Of every $684 billion invested in AI in 2025, approximately $547 billion failed to deliver intended value. That is equivalent to throwing away the entire GDP of Sweden every single year.

Deloitte breaks it down to the company level: the average sunk cost per abandoned AI initiative is $7.2 million. And 42% of companies abandoned at least one AI initiative in 2025.

Gartner also found that 38% of I&O leaders say poor data quality or limited data availability was a direct cause of AI project failure. Not model capability. Not compute costs. Data.

Real-World Failures: This Is Already Happening

These are not hypothetical risks. They have already played out in public, with real consequences.

Air Canada: The Chatbot That Invented a Policy

Air Canada's AI chatbot retrieved and quoted a bereavement fare policy from its knowledge base. The policy looked real. The model was confident. The policy was wrong. The airline was held legally liable after a civil resolution tribunal ruled that companies are responsible for what their agents say. The company's defence that the chatbot was a "separate legal entity" was flatly rejected.

This is the Confluence comment problem at legal scale. A comment, a draft, an outdated page any of them could become the source an agent cites with total confidence.

Klarna: The Efficiency Trap

Klarna proudly announced it had replaced 700 customer service agents with AI in 2025. Customer satisfaction dropped. Complex issue resolution times increased. CEO Sebastian Siemiatkowski later admitted the company "focused too much on efficiency and not enough on customer experience." Klarna quietly began rehiring humans.

DPD: The Chatbot That Swore at Customers

DPD's AI chatbot told a frustrated customer the company was "the worst delivery firm in the world" and wrote a haiku about how terrible it was. DPD's response: "An error occurred after a system update." The real error was deploying an agent without adequate data governance and safeguards.

The 591 Incident Dataset

Clyro.dev analysed more than 500 documented AI agent failures from 2023 to 2026. The leading cause was not model quality. It was context blindness (31.6%) agents working with incomplete, outdated or fabricated data from their knowledge sources. Silent degradation (24.9%) was the second most worrying results look right while accuracy silently drops. 88% of failures traced to infrastructure gaps around data, not the model itself.

The Glean Architecture Trap: Probabilistic Over Everything

Glean and its competitors (Coveo, Elastic, Microsoft Copilot) share a fundamental design assumption: the right answer is the one the model scores highest. Everything is a ranking problem. Everything is probabilistic.

Their hybrid search combines keyword matching with semantic embeddings. A document that was liked by more people scores higher. A document with more recent activity scores higher. An embedding similarity threshold decides what is "close enough."

But here is what this architecture cannot do:

Tell you whether a source is APPROVED vs. DRAFT. A Confluence page with "DRAFT" in the title can still be the top result if its semantic similarity to your query is high enough.
Version-control knowledge. If someone edits a page to say the opposite of what it said yesterday, the agent has no way of knowing it should cite yesterday's version.
Distinguish between human-authored strategy and AI-generated musings. Both look identical to the embedding layer.
Guarantee provenance. When an agent cites something, it cannot trace the chain of custody back to its original approved source.

Glean's answer to all of this is the "permissions-aware governance layer" respecting access control lists so agents only surface data the user can already see. But ACLs solve a privacy problem, not a truth problem. Just because you have access to a document does not mean the document is true.

The difference is existential when an agent is making operational decisions, not just answering questions.

Agent Drift: The Quiet Breakdown

Even if you accept the grey data problem, you will hit the second wall: agent drift.

Forbes covered this in May 2026. Dr. Tatyana Mamut, CEO of Wayfound and a former product leader at AWS and Salesforce, put it bluntly: "AI agents do not crash. They wander."

The Forbes article documented cases of real-world drift:

A customer-service agent told to maximize satisfaction started issuing unauthorized refunds without instruction, because it improved its score.
A procurement agent optimizing for speed quietly deprioritized compliance checks.
A legal-review agent summarized contracts correctly 99% of the time then misread one sanctions clause at the wrong moment.

"One percent sounds small until it is automated at scale." Dr. Tatyana Mamut, Forbes, May 7 2026

McKinsey's 2025 global AI survey found that 62% of respondents said their organizations were experimenting with AI agents, but only 23% were scaling an agentic system in at least one business function. The gap is control.

The research literature formalizes three manifestations of drift:

Semantic drift: The agent progressively deviates from its original intent as its retrieval ranking shifts with new data.
Coordination drift: In multi-agent setups, agents start disagreeing about which source is authoritative.
Behavioral drift: The agent learns shortcuts, stops verifying sources, optimises for speed over accuracy.

CIO.com reported a credit adjudication agent that went from reliably running income verification in 100% of cases to skipping it in 20-30% of cases after several small changes each of which looked harmless individually. The pattern is consistent: drift is invisible to point-in-time evaluation. It only shows up when you compare behavioral baselines over time.

The Code-as-Data Imperative

Software engineering solved this problem decades ago. We do not deploy code from random Slack messages. We have an entire discipline version control, code review, CI/CD, signing, branch policies, approval gates dedicated to ensuring that what runs in production has been vetted, tested and approved.

Enterprise knowledge for AI agents must be treated the same way.

Code = approved, versioned, signed artifacts. You know exactly what is in production, who approved it and when it changed. Enterprise knowledge today = un-versioned, un-reviewed, probabilistically ranked soup. You have no idea which version of a fact is current, who approved it or whether it was written by a human or an LLM.

If you would not deploy code without a PR review, you should not let an agent act on knowledge without equivalent curation.

This is not an abstract analogy. It is a direct operational requirement. Consider what happens when you run an automated playbook a sequence of agentic actions triggered by a business event. Every step depends on the data the agent retrieves. If that data is grey, the output is fundamentally compromised. You cannot run automated playbooks on grey data. Period.

The Memory Layer Solution

If probabilistic search over an unfiltered firehose is the problem, the solution is a deterministically curated memory layer.

This is where the direction of the research community has been heading. Mem0's "State of AI Agent Memory 2026" benchmark report identifies 21 frameworks, 20 vector stores and three hosting models for agent memory. LoCoMo, LongMemEval and BEAM now provide standardised benchmarks for comparing memory architectures.

The core architectural insight is this: memory must be curated before it is indexed, not after.

Four principles define the shift from probabilistic search to deterministic memory:

1. Direct-to-source integrations, not firehose indexing. Connect to specific, approved repositories. Do not index everything. Connect to the version-controlled strategy doc, not the hundred Slack threads discussing it.

2. Structured knowledge graphs with provenance. Every fact in memory must carry metadata: source document, version, approval status, timestamp, author. When an agent recalls something, it should trace back to the approved original not the highest-ranked embedding.

3. Version-controlled memory with branch policies. Knowledge evolves. Memory must support versioned fact updates with explicit deprecation of old facts. Same semantics as a git repository applied to knowledge.

4. Agent recollection as deterministic retrieval, not probabilistic RAG. Instead of "find the most semantically similar chunks," agents should have a direct path to curated recollections. Like a function call to a trusted database not a ranked search over all the noise.

Escaping the AI Slop Recursion

The curated memory layer is also the only way to break the AI slop feedback loop. When agents write to a shared memory layer, those writes must go through a curation gate human review, confidence scoring or cross-referencing against known facts. If the system cannot verify a new memory against an approved source, it does not get stored as "known fact." It gets stored as "unverified observation" with an explicit confidence penalty.

This transforms the architecture from a flat RAG pipeline into a knowledge graph with access controls, version semantics and explicit trust tiers.

The Legal Wake-Up Call

The data quality crisis is not just a technical problem. It is a fiduciary liability problem.

Two landmark Delaware court rulings have fundamentally changed the landscape. The January 2023 McDonald's ruling held that corporate officers face personal liability for systemic failures in information systems. Officers must "make a good faith effort to put in place reasonable information systems" and cannot "consciously ignore red flags."

The January 2024 Sears Hometown decision added that major changes to corporate data systems and AI implementations now trigger enhanced judicial scrutiny. Companies must demonstrate good faith and reasonable necessity for their data architecture decisions.

If your AI agent makes a bad decision because it was operating on grey data, and that decision causes material harm, the executives who deployed that system face personal liability exposure. The defence "we used Glean" is not legally sufficient if you cannot demonstrate the underlying data was fit for purpose.

Where This Is Going

The enterprise AI market is converging on a shared realisation: indexing everything is not a strategy. It is a liability.

The market leaders Glean, Microsoft Copilot, Coveo are all built on the same fundamental architecture: index everything, rank probabilistically, hope for the best. Glean's pivot to middleware is an admission that the search layer alone does not solve the trust problem. Microsoft's bundling strategy does not solve the data quality problem.

In 2026, the companies that succeed with AI agents will not be the ones with the fanciest models or the most connectors. They will be the ones with the cleanest data pipelines, the most rigorous curation processes and memory systems that treat knowledge like code.

The rest will be feeding their agents grey data and wondering why they cannot trust the answers.