6 May 2026
The data your organisation most needs to act on is the data your AI systems trust the least. That is not a technology gap. It is a structural risk.
There is a number that should concern every senior insight leader: 80.
Eighty percent of enterprise data is unstructured. It does not sit in rows and columns. It does not slot neatly into dashboards. It lives in survey responses, call transcripts, chat logs, review platforms, and open-text feedback fields that accumulate, largely untouched, in the background of every customer interaction your organisation has.
For VoC and CX professionals, this is not new information. You have known for years that the richest signal in your data estate is the one that is hardest to process. A customer who scores you a 6 in an NPS survey tells you a number. The same customer explaining why they gave you a 6 tells you something worth acting on.
The question has always been: how do you act on it at scale?
For a while, generative AI looked like the answer. LLMs can read a thousand survey responses and return a coherent summary in seconds. That is genuinely impressive. But impressive and trustworthy are not the same thing. And as agentic systems move from surfacing insights to triggering actions, the difference between the two has significant consequences.
The Architecture of the Agentic Problem
Agentic AI represents a meaningful step change in how organisations use data. Rather than producing outputs for humans to evaluate, agentic systems are designed to act: routing cases, triggering alerts, initiating workflows, escalating risks. The reduction in operational friction is real, and the efficiency case is compelling.
But there is a question that sits underneath every agentic deployment, and most organisations have not answered it with sufficient rigour: what happens when the system acts on something it got wrong?
For structured data, this is a manageable risk. If your agentic system misreads a transaction value or misclassifies a ticket type, the error is visible, traceable, and correctable. The data model has clear boundaries.
Unstructured text data is different. Language is contextual, nuanced, and critically, interpreted. When an LLM processes a customer comment, it does not measure meaning. It generates a plausible version of meaning. That is not a flaw in how these models are built. It is precisely what they are designed to do.
The technical term for this is non-determinism. Ask an LLM the same question twice and you may get two different answers. Run the same dataset through the same model on two separate occasions and you may get divergent outputs. For exploratory tasks, this variability is acceptable, even useful. For systems designed to make automated decisions, it is a foundational problem.
The Trust Deficit at the Heart of Unstructured Analytics
Consider what a typical CX agentic workflow might look like. Customer feedback arrives across channels: surveys, support tickets, social, chat. The agentic system analyses it, identifies elevated churn risk signals, and triggers an intervention: a retention offer, a case escalation, a flag to the account team.
For this to work reliably, the system needs to consistently identify what the language in those interactions actually means. Not approximately. Consistently.
A global financial services firm running this at scale across hundreds of thousands of monthly touchpoints cannot afford a system that sometimes identifies fee frustration as a churn signal and sometimes categorises it as a routine enquiry. The commercial and regulatory consequences of that inconsistency are significant.
This is why the 80% figure matters as much as it does. The majority of your customer signal is sitting in unstructured text. Your agentic infrastructure is being built to act on data. But your current AI layer cannot reliably tell you, with statistical confidence, what that text is actually communicating, and it certainly cannot do so with the consistency required to trigger automated action responsibly.
The result is a gap. A large volume of highly valuable signal remains outside the decision boundary of your most capable systems.
What Deterministic Analysis Actually Means
This is where Relative Insight enhances insight propagation. First, the platform uses dThe solution to this gap is not better prompts or larger models. It is a different architectural approach.
Deterministic analytics operates on a different principle to generative AI. Where an LLM generates a plausible interpretation, a deterministic system measures statistically significant difference. The same input produces the same output. Every time. The methodology is traceable. The outputs are auditable. The differences it surfaces are quantified, not described.
Applied to the text data that makes up the bulk of your customer and employee signal, this changes what is possible at the agentic layer.
When your system identifies that customers who mention waiting time in a specific syntactic context are 14 times more likely to feature in churn data over the following 90 days, that is not a summary. That is a measurable signal. It can be validated. It can be defended. It can be acted on by an automated system with the same confidence you would apply to a structured metric.
This is the enabling condition for agentic systems to work reliably with unstructured data: not more generative capability, but a deterministic foundation that provides the statistical rigour those systems need to act with confidence.
The Compounding Value of Unlocking the 80%
The commercial case here extends beyond operational efficiency. Consider what becomes possible when your agentic infrastructure can consistently interpret and act on the full breadth of your customer signal.
- The business case becomes more defensible. When a VP of Insight presents a recommendation to the leadership team that originated from an automated text analysis system, the question they will face is: how do we know this is right? A deterministic answer to that question carries a fundamentally different weight than a generative one.
- Churn models become richer, because the language signals that precede churn are often more predictive than behavioural metrics alone. Early-stage dissatisfaction frequently surfaces in text before it shows up in usage data or CSAT scores.
- Personalisation becomes more precise, because the words customers use to describe what they want reveal preference patterns that structured data flattens or misses entirely.
- Root cause analysis becomes faster, because rather than manually reviewing samples of open-text feedback to understand why a metric has moved, your system can surface the specific language differences that explain the shift, consistently and at scale.
The Hybrid Architecture That Makes This Real
None of this is an argument against generative AI. The fluency of LLMs is genuinely valuable: for surfacing themes, drafting summaries, making insights accessible to non-technical audiences, and reducing the friction of analytical workflows. These are real advantages.
The point is that fluency and rigour serve different purposes, and a well-designed analytics architecture recognises that distinction.
The organisations that will extract the most value from agentic AI are those that deploy generative and deterministic systems in complementary roles. Generative layers make insight accessible. Deterministic engines make it trustworthy. The agent acts as the interface. The deterministic system operates as the underlying authority, quantifying, validating, and anchoring outputs in evidence that can withstand scrutiny.
This is not a speculative architecture. It is the practical requirement for any organisation that wants to responsibly extend agentic decision-making into the domain where the majority of its most valuable signal lives.
The Question Worth Asking
If you are currently building or evaluating agentic analytics infrastructure, one question cuts through the complexity:
For every automated action your system is designed to trigger based on customer or employee language data, can you demonstrate the statistical basis on which that action was initiated?
If the answer is no, or not consistently, you have an 80% problem. The data is there. The value is there. The missing piece is the layer that makes it trustworthy enough to act on.
That layer exists. The organisations that build it into their agentic stack now will find themselves in a structurally different position to those that realise they need it later.
Relative Insight’s deterministic analytics platform is purpose-built to provide the statistical rigour that agentic systems require to act confidently on unstructured text data. To understand how it works in the context of your VoC or CX infrastructure, get in touch.
Want to learn how you can replicate this set up, with tailored, automated insight delivery going to your stakeholders? Book a call with our team to find out more.
