Most large companies, from financial services to retail, have made AI personal. Tens of thousands of assistant licences have helped employees save hours a day delivering real and measurable uplifts in individual productivity. Yet almost none of this has translated into strategic value:
- Cost-to-income has not moved.
- The combined ratio has not moved.
- The numbers the board actually report on are unchanged.
The productivity gains are genuine, but they simply do not aggregate into any benefits the organisation can point to.
This is the question every data and AI leader is now being asked, in one form or another. We have adopted AI. So where is the value? The honest answer is that personal productivity and strategic value are not the same achievement, and the path between them does not start with more adoption or better prompts. It starts with something far less exciting: saying precisely what you mean by AI.
AI is three things, and only one of them is new
What sits under the label of AI in most businesses is three different technologies, with different value mechanics, and they do not perform the same job.
- Deterministic automation.
- Classical machine learning (ML) and optimisation.
- Generative AI.
Deterministic automation is rules execution with the same input, same output, every time. It saves hours on high-volume, low-variance work, and most organisations banked those gains years ago. It does not predict or learn anything. It is not really AI, which is worth clarifying, because a good deal of the AI return in board packs is automation savings wearing a new label.
Classical machine learning and optimisation is where strategic value has actually lived for the last two decades. Primarily seen in pricing, underwriting, fraud, retention, allocation, claims triage, the output is a number or a decision, clear enough to be measured against reality, so that the question “did it work?” has a numerical answer. These are profit and loss lines, not productivity stories, and the governance around them is mature. Optimisation and machine learning are often filed as separate disciplines, but in production they rarely are: forecasts feed solvers, constraints are learned from data, and one specialist technical team (normally Data Science in mature businesses) owns the decision end to end.
Generative AI is the new arrival, and the source of the confusion. It is genuinely powerful and broadly useful because it writes code, drafts and reviews text, answers questions over documents, and synthesises across material that was previously too large or too unstructured to use. Developers ship faster with it; analysts get through more. The utility is real, and it is large, but its value mechanic is different from the other two. It does not make a better decision per unit. It compresses the cost of a knowledge-work step. And that kind of value, however large, is scattered as it is captured by individuals, spread across thousands of small tasks, and hard to see on the measures the board tracks. The technology is not weak as productivity value and strategic value are simply different things, and one does not automatically become the other.
There is one place where generative AI earns strategic value close to standing alone, and it is worth being precise about it. In document-heavy work, evidence review, due diligence, suitability files, regulatory horizon-scanning, and retrieval over a large internal document store, a well-engineered system can change what the work costs to do entirely. That is more than productivity. But this is a narrower set of cases than the market assumes, and it is not the same as the assistant rollout most organisations are measuring.
What goes wrong when you treat them as one
Once the three technologies are seen clearly, the failure patterns are easy to read.
- Automation savings reported as AI return. The numbers are real but the provenance is wrong. The AI investment line absorbs work that predated it.
- Generative AI asked to deliver decision-grade uplift. It cannot, because that is not what it does. The disappointment that follows gets read as “AI does not work,” which is the wrong conclusion drawn from the wrong experiment.
- Adoption used as a proxy for value. Weekly active users is an input metric. High adoption with no movement in the underlying business measures is a sign that value is not flowing.
- The decision layer quietly starved of investment. The most reliable source of strategic value in the business loses budget to generative initiatives at an earlier stage of maturity, and the opportunity cost is real and rarely quantified.
Turning productivity gains into strategic value
This is the transition most businesses have not made. Going from personal productivity to strategic value is not going from less AI to more AI. It is going from generative AI standing alone to generative AI wrapped around a decision core.
Put plainly, the assistant stops being the product and becomes the interface. The work that has to be correct and defendable, the price, the risk score, the optimisation, the forecast, sits in a deterministic or machine-learning core that can be tested. The generative layer pulls structure out of messy inputs, retrieving the right context, orchestrating the steps, and explaining the result in language a person can act on. The value is generated in the core and the generative layer makes the core usable, faster, and accessible to people who could never query it before.
This also changes the sizing question. A frontier model is rarely the right tool for the bounded jobs inside one of these pipelines: extracting fields, classifying a document, drafting a section against a fixed template. Smaller models, often run inside the organisation’s own environment, are good enough for those steps, cheaper at volume, and easier to govern where the data cannot leave the perimeter. The heavy reasoning is reserved for where it earns its cost. Most corporate demand is text-based knowledge work of exactly this shape, and it needs the right model in the right place, not the largest model available.
The pattern is visible once you look for it. For example, an underwriter who interrogates a pricing model in plain language and understands what is driving the number, instead of reading a static dashboard; or a reserving process where the movement is decomposed deterministically and the commentary is drafted from that decomposition, with every figure traceable back to the engine; or a claims workflow where the triage decision is a model and the handler talks to it. In each case the language model never touches the number. It makes the number legible and actionable. That is the difference between a tool that saves a person time and a capability that changes how the work is done.
Organisations miss this because the standalone assistant is easy to see. The hybrid system’s value is in the part you cannot see, the core that does the real work, and that is harder to sell and harder to build. It needs people who can build the decision layer, not just configure the assistant. So the market buys the visible thing, gets productivity, and wonders where the value went.
What businesses need to do
Three things follow for any business that wants to successfully make the transition:
Separate the spend and the metrics. Productivity tooling belongs with productivity measures, time saved, adoption, user satisfaction. It does not belong on the same dashboard as a credit or pricing model, where putting them together flatters one and hides the other. Measured apart, the value question finally has a transparent and honest answer.
Stop restricting the decision layer. Strategic value lives in the decision systems, the classical machine learning and optimisation layer, and in many organisations it is quietly losing headcount and budget to generative initiatives at an earlier stage of maturity. Mostly this is not a decision anyone takes deliberately as the team loses two people and a project slot. If the business genuinely means to shift investment from decision systems to generative AI, that should be a choice the executive owns, not something that happens by drift. For most businesses, the core decision systems are still the higher-value bet for the next few years.
Build for the hybrid, not the demo. The standalone assistant is the right destination only if productivity is the ambition. If the ambition is value, the generative layer must be engineered to sit on a decision or data core, with the evaluation, faithfulness checks, and monitoring that let it be defended to internal audit, to the regulator, and to the organisation itself. Without that, it is not a system in production, it is a demo with users.
Putting the assistant in its place
The next phase of AI will not be won by the organisations with the most AI licences. It will be won by those that can say precisely what they mean by the word and how it is being deployed in core parts of the business; it will be won by those that use the generative layer as the interface to something that does real work rather than asking it to be the engine (which it is inherently bad at due to context length limitations).
Personal productivity made AI familiar, and that was useful. But familiarity is not understanding, and understanding is where value starts. “We rolled out an assistant” is not a strategy, it is a procurement event.
The road to value runs through something far less exciting than the vibe-coded one-shot demos suggest: foundational engineering and data science principles. Data and AI leaders need to define the three things properly and build use cases around the core business functions that can move revenue, margin or profit, and fund the work that moves those numbers. Keep the generative layer where it belongs, as the interface, not the engine. None of it starts until you stop calling three different things by one name. Stop calling it AI, and the rest gets easier to see.
