Your AI copilot is confidently wrong. Faster.
By Connie Yang, Co-founder and CEO, Nth AI. Published 2026-07-06.
Every enterprise now has a chat-with-your-data tool, and most executives trust it less than the ticket queue it was supposed to replace. The problem was never the model. It is that AI got bolted on top of data nobody governed, so it inherits every disagreement in the stack and states it with total confidence. The fix is not a smarter copilot. It is a trusted definitions layer underneath it.
Ask two executives at the same company for last quarter's gross margin and you will get two numbers. Both came from dashboards. Both dashboards came from the data team. Neither executive can tell you why they disagree, and by the time someone reconciles them, the decision that needed the number has already been made on gut feel.
That was the state of enterprise analytics before AI. Then chat-with-your-data arrived and promised to fix it. Point a copilot at the warehouse, ask a question in plain English, get an answer in seconds.
Here is what actually happened in most enterprises: the answers got faster and the trust got worse.
The receipts are in. MIT's NANDA initiative found that 95% of enterprise GenAI pilots delivered zero measurable return to the P&L, despite $30 to $40 billion of enterprise investment (The GenAI Divide, 2025). Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (June 2025). And this is the second time we have watched this movie: self-serve BI, the last wave that promised to free executives from the queue, has been stuck below 20% adoption for more than seven years (Gartner / BARC). IBM reports that only 29% of employees use the BI tools their employer already pays for.
Two waves of tooling, a decade apart, failing the same way. That pattern is not a coincidence. It is a diagnosis.
The bottleneck has two layers, and AI only touched one
The visible layer is the queue. Every dashboard, report, or analysis goes through the data team and the BI team. Each request is a ticket. Each ticket is weeks. Executives learn to stop asking, and decisions get made on stale numbers or no numbers at all.
The invisible layer is the reason the queue exists in the first place. The data is scattered across dozens of systems that do not agree with each other, and the definitions that make a number correct, what counts as revenue, what counts as an active customer, how gross margin is actually calculated, do not live in any system at all. They live in people's heads. Finance has one version. Sales has another. The BI analyst who built the dashboard three years ago had a third, and she left.
Chat-with-your-data tools attacked the first layer and ignored the second. They put a conversational interface on top of whatever data already existed, ungoverned, contradictory, and undefined. The result is exactly what you would predict. AI bolted onto data nobody trusts does not fix the bottleneck. It industrializes it. The wrong answer that used to take three weeks now arrives in four seconds, beautifully formatted, with no way to tell it apart from a right one.
Confidently wrong, faster.
The governance numbers explain why this keeps happening. IBM finds 87% of organizations claim to have an AI governance framework, but fewer than a quarter have fully implemented one. Economist Impact puts comprehensive AI governance at 8% of organizations. The models got dramatically better. The foundations they sit on did not move.
Everyone is automating a slice. The slices don't add up.
Step back and the market attacking this problem splits into three camps.
The copilots and text-to-SQL tools automate the interface. They are only as good as the foundation underneath them, and they do nothing to build it.
The pipeline and ETL tools automate data movement. They move bytes between systems, but they do not capture what the numbers mean, and their output is still raw material waiting for a BI team to shape it.
The semantic layer and governance vendors define truth on paper. But they do not build the pipelines below or the experience above, so an enterprise still pays for the labor on both sides of them.
Each camp automates one slice and hands the rest back to you as a project. The project is the actual product. That is why the standard fix today is still labor: internal data engineers, BI analysts, and data scientists, plus outsourced implementation teams at seven figures per engagement and six to nine months to first value.
And here is the industry's tell. In June 2026, AWS announced a $1 billion investment in a new forward-deployed engineering unit, embedding engineers directly with customers to make AI actually work, following the same playbook OpenAI and Anthropic built first. When the largest platforms on earth staff the gap with humans, they are conceding, publicly and at billion-dollar scale, that the gap exists and that their software alone does not close it.
Throwing people at it is the old answer. Labor scales linearly with every new request. Software should not.
The unlock is not a smarter model. It is a trusted definitions layer.
Every company runs on definitions. What counts as revenue. What counts as an active customer. How margin is calculated, by product, by region, by quarter. Today those definitions are tribal knowledge, and every department carries its own version.
Now imagine they were captured once, reviewed and approved by the humans accountable for them, and then enforced everywhere: every dashboard, every report, every AI answer resolving against the same governed set of definitions. Not a policy document. A living layer that sits between the raw data and everything that consumes it.
Two things change immediately.
First, answers become trustworthy, because there is exactly one way to compute each number and it is the approved way. An AI operating above that layer is not guessing what you mean by churn. It is bound to the definition your CFO signed off on, with lineage back to the source.
Second, the queue disappears, because self-serve stops being dangerous. When the definitions are enforced by the platform rather than by a reviewer's vigilance, an executive can generate their own dashboard, in minutes, scoped to their own department, and the data team does not have to fear what comes out. The data team's role shifts from producing every artifact by hand to reviewing and governing what the system produces. They move up a level. Nobody gets displaced by a ticket queue vanishing. They get promoted out of it.
Personalization and governance stop being a trade-off. Every executive sees a different surface, their data, their vocabulary, their scope. Every number underneath resolves to the same truth.
Why this matters right now
Three observations for anyone deciding where this market goes next.
The budget already exists. Enterprises are not being asked to fund a new category. The spend is already in the P&L as data and BI headcount and outsourced implementation engagements. The question is only whether that line item keeps buying hours or starts buying software.
The timing window is real. Agentic systems just crossed the threshold where the repeatable middle of implementation work, connecting sources, drafting definitions, generating pipelines and reports, can be produced and tested by software with humans approving the output. Two years ago that sentence was fiction.
And the economics compound in one direction. A labor model costs roughly the same on engagement fifty as on engagement one. That is as true for a $1 billion forward-deployed engineering unit as it is for a traditional consultancy. A software model gets faster and cheaper with each deployment, because the scaffolding transfers. Any approach that scales with headcount is competing against a curve it cannot match.
This is the thesis we are building against at Nth AI. Nexus turns fragmented enterprise data, metrics, policies, and business logic into governed context that AI can run on. It automates the implementation work required to build semantic models, ontologies, mappings, and controls as reviewable artifacts the customer owns, so agents can understand the business and act reliably instead of guessing from raw systems.
If your executives have learned to stop asking, or your copilot answers fast and gets trusted never, that is not an AI problem. It is a foundation problem, and it is now solvable with software.