MIT's 2025 NANDA initiative looked at 800 enterprise GenAI pilots. 95% failed to deliver ROI. The 5% that succeeded had one architectural property in common, and it isn't model size or budget.
In August 2025, MIT's NANDA initiative published a study titled The GenAI Divide: State of AI in Business 2025. The headline number, instantly viral, was that 95% of enterprise generative-AI pilots had failed to deliver measurable returns. The remaining 5% were not just slightly better. They were dramatically better, with rapid revenue acceleration that justified the investment many times over. The gap between the two groups did not correlate with company size, AI budget, or model choice. It correlated with something more fundamental: whether the pilot was deployed against a workflow narrow enough to verify the AI's output against ground truth.
The MIT research drew on 150 leader interviews, 350 employee surveys, and 300 public AI deployments. The pattern was strikingly consistent. Successful pilots tended to be:
The failures, conversely, almost all shared one trait: the AI was deployed on a task where there was no easy way to tell whether the output was right. A marketing assistant that drafts campaign copy. A "strategic analyst" that produces narrative reports. A general-purpose chatbot bolted onto enterprise data. In each case the output looks plausible, the user has no efficient way to verify it, and the company quietly accumulates a backlog of decisions made on unverified AI claims.
The report calls this "the trust gap": the space between what an LLM confidently outputs and what an organization can confidently act on.
The MIT finding lines up with what AI-safety researchers have been saying for years: large language models are good at producing plausible text, and the gap between plausible and correct is invisible from inside the model. In domains where correctness is easy to check (does the code compile? does the document parse?), the gap closes naturally because failures are immediately visible. In domains where correctness is hard to check (does this market analysis match reality? does this customer insight reflect what users actually do?), the gap stays open indefinitely, and the AI keeps shipping confident outputs that nobody fully audits.
Analytics sits firmly in the second category. An AI tool that says "Your bounce rate increased 14% on mobile last week" is producing output that's nominally checkable but in practice almost never checked. The user has to leave the tool, log into GA4, build the right report, filter it correctly, and compare. Almost nobody does. The AI's claim becomes the working truth, and decisions cascade from there.
This is why the MIT failure rate is so high. Most enterprise GenAI pilots are deployed without an architectural mechanism for the AI to be checked against ground truth. The model becomes the source of truth by default, and the only feedback signal is downstream business outcomes weeks or months later, by which time the original error is lost in the noise.
Askalytics is built around the architecture MIT's successful 5% had in common: verifiable evidence, narrow AI, traceable claims.
The 5% of pilots that succeed have a property that's easy to retrofit: every AI output is traceable back to a verifiable source, and the verification step is built into the pipeline rather than left to the user.
Askalytics' warehouse-grounded architecture implements this for analytics. Every property's GA4, Search Console, and Google Ads data is synced nightly into a structured data warehouse owned by Askalytics. The warehouse is the source of truth: stored evidence that what we report came from what Google's APIs returned. The AI runs its queries against that warehouse, not against live API calls, which means:
On top of that, every AI Insight is re-checked by a second, different AI model. This is the architectural shape the MIT report's successful 5% had in common: narrow, verifiable, with every output traceable to its source and re-checked before it ships. The unsuccessful 95% were trying to use one model to analyze and answer in one undifferentiated pass, with nothing to trace back to and no second opinion, and the trust gap broke them.
The MIT report didn't reveal anything new about model quality. The frontier models are stronger than they've ever been. What the report revealed is that model quality is no longer the bottleneck. What's failing enterprise pilots is architecture. The teams shipping AI products that actually work are spending less time tuning prompts and more time designing systems where the AI's outputs are verifiable by construction.
If you're evaluating AI analytics tools right now, the right diagnostic question is no longer "how smart is the AI?" but "can I trace any number it produces back to a row in my data?" If the answer is no, you have a 95-percenter on your hands.
For more on how we approach this, read about our traceable AI Insights and the Confidence Check that backstops every insight.