AI at the Frontier of Risk Intelligence

There is a reflex in our industry right now to answer every problem with the largest available model. A supplier needs screening? Ask a frontier model. A thousand news articles need reading? Ask a frontier model. It feels like progress. In practice it produces a system that is slow, costly, and hard to audit, and one that is wrong more often than anyone admits.

We have taken a different route. Turning messy real-world signals into risk intelligence is not one problem; it is dozens of smaller ones, each with a technique that suits it best. The frontier of this work lies in fitting the right method to each task, and that discipline is what lets us analyse more deeply and at greater scale than the systems we compete against.

The bigger-is-better reflex is miscalibrated

Large reasoning models are remarkable, and they are also the most expensive and least predictable component you can drop into a pipeline. NVIDIA's research team put numbers on the gap in their position paper Small Language Models are the Future of Agentic AI: serving a 7-billion-parameter model is roughly ten to thirty times cheaper, in latency, energy, and compute, than serving a 70-to-175-billion-parameter one. Their conclusion is blunt. Insisting on a large general model for narrow, repetitive tasks is a misallocation of computational resources, and most of what production systems actually do is narrow and repetitive.

The waste shows up even within a single model. Reasoning models can generate seven to ten times as many tokens as non-reasoning models to reach comparable accuracy on simple work, deliberating at length over a question as trivial as the capital of France. Researchers call this the overthinking problem. Paying frontier prices for it, across millions of documents, is how a risk platform prices itself out of doing the job thoroughly.

Most of the work is sorting, not reasoning

A large share of what a risk pipeline does is sorting. Is this article about forced labour or a routine labour dispute? Which company does this sentence name? Is the sentiment toward this supplier negative, and how strongly? These are classification and extraction problems, and they were largely solved before the current wave of generative AI arrived.

BERT and the encoder models that followed it handle them very well. By reading text in both directions at once, an encoder builds a rich representation of a passage, and a small task-specific layer turns that into a sentiment score, an entity tag, or a category label. The model is fast, costs a fraction of a frontier call, and returns the same answer for the same input every time. A well-chosen specialist can reach parity with a far larger model on a focused task with around a hundred labelled examples. For sentiment, named-entity recognition, and document classification, this is the correct tool, and reaching past it buys nothing but latency and a larger bill.

Context lives in vector space

Judging whether two pieces of text mean the same thing is a different problem, and it has its own well-suited tool. Sentence embeddings map a passage to a point in high-dimensional space, positioned so that text with similar meaning sits close together. Comparing two passages then becomes a matter of measuring the distance between two vectors.

This is what lets us connect a report written in local language about conditions at a specific site to a commodity, a supplier, and a sub-national region, even when none of the obvious keywords match. A search for forced-labour indicators can surface a document that never uses those words but describes the reality of them. Because the heavy lifting happens through fast nearest-neighbour search over precomputed vectors, this holds up across very large corpora while staying cheap to run. It is the backbone of how we connect signals across our commodity risk and supply chain mapping work.

The same property reaches across languages. Risk signals rarely arrive in English, and the sharpest ones often surface first in local-language reporting, court filings, or community posts. Multilingual encoders such as XLM-RoBERTa, trained across a hundred languages, and the multilingual sentence transformers built on them place text from different languages into one shared vector space. A query framed in English can then surface a document written in Portuguese or Bahasa Indonesia, because the two sit close together by meaning rather than by spelling. For a supply chain that crosses dozens of jurisdictions, that cross-lingual reach is what stops a serious signal from being missed because it happened to be written somewhere else.

Reasoning is the exception, not the default

None of this means generative reasoning has no place. It means its place is specific, and a growing body of research supports treating it as the exception rather than the default. Studies have found that for many tasks a reasoning model performs just as well, and considerably faster, when the extended thinking step is skipped entirely. The pragmatic architecture that follows is a heterogeneous one: lightweight specialists handle the bulk of the work, and a large model is called in only for the problems that genuinely need it.

The same logic as a quick reference:

This triage can be made systematic. In a model cascade, each input goes first to the cheapest model that might handle it and is escalated to a larger one only when a confidence check says the cheap answer cannot be trusted. The economics are stark: published work shows a learned cascade matching the accuracy of the strongest single model at up to ninety-eight percent lower cost, helped by the fact that prices between the smallest and largest models differ by two orders of magnitude. A surprising share of inputs are handled correctly by a small model, which leaves the expensive one for the genuinely hard remainder. The principle we hold to is the same throughout: the default is the lightest competent tool, and capability is added only where the work demands it.

The clearest measure of what those models are uniquely good at comes from METR, which tracks the length of task a model can complete reliably. The time horizon for what frontier agents can finish at fifty percent reliability has been doubling roughly every seven months, driven by better handling of long, multi-step work where mistakes have to be noticed and corrected along the way. That is the shape of problem a frontier model earns its keep on: synthesising a coherent narrative from twenty fragmentary sources, weighing conflicting evidence across the dimensions of a CAHRA assessment, reasoning through a tangled chain of ownership. For everything shorter, something cheaper and steadier will do.

The quiet case for restraint: errors compound

There is a deeper reason to keep generative steps to a minimum, and it is mathematical rather than financial. In any pipeline that runs many stages in sequence, error compounds. A step that is right ninety-nine percent of the time sounds excellent, yet chained across a hundred stages it completes the whole task correctly only about thirty-seven percent of the time. The arithmetic is unforgiving, and no amount of model size repeals it.

What makes generative steps especially dangerous in this setting is that their failures are silent. A traditional software error throws an exception you can catch. A model that quietly extracts the wrong entity or misreads a sentence produces output that looks perfectly plausible, and every downstream stage treats it as fact. The same property that makes these models fluent is why they generate confident falsehoods even on straightforward questions, and some researchers argue this is a structural feature of how they work rather than a flaw to be patched out.

Deterministic components behave differently. An encoder that returns a label and an embedding model that returns a vector give stable, reproducible outputs that can be logged, checked, and explained. Every generative step you replace one of those with trades that stability for variability and reintroduces the risk of an invisible fabrication. Keeping generation to the minimum the task requires therefore produces a system that is more predictable and more accurate, with lower cost as a side effect.

When a step does call for generation, the same vector search that finds context can also fence it in. Retrieval-augmented generation conditions the model on specific passages pulled from the source material, so its answer is tied to evidence on the record rather than to whatever it absorbed in training. This lifts factual accuracy and, just as usefully, leaves a trail: each generated claim can point back to the document it rests on. It does not make hallucination impossible, since a model can still stray from the passages it was handed, which is one more reason to keep the generative surface small. Grounding makes the rare generative step safer, and using it sparingly is what makes the system dependable.

Scale is the whole game in due diligence

This restraint is not an academic preference. It is what makes the regulatory job possible at all. The EU's Corporate Sustainability Due Diligence Directive requires in-scope companies to identify, prevent, and account for adverse human rights and environmental impacts across their value chains, which means processing a great deal of data about a great many suppliers on a recurring basis.

The depth of the problem is easy to underestimate. Network analysis of EU supply chains suggests that while a small share of firms carry forced or child labour risk in their first tier, more than ninety-nine percent are likely to have such risk by the third tier. Risk does not live at the border or the obvious supplier; it hides several steps down, scattered across thousands of relationships.

No team screens that by hand, and no budget routes it through a frontier model document by document. Efficient machine learning is the only way to assess an entire portfolio rather than a handful of priority commodities, and that coverage is the difference between checking your most-scrutinised supplier and checking all of them. It is also where a well-engineered platform pulls clear of competitors still paying premium rates for every step.

Efficiency buys freshness as well as breadth. Due diligence is an ongoing obligation rather than a one-off audit, and risk does not wait for the next quarterly review: a strike, a court ruling, or a field report can change a supplier's profile overnight. Because lightweight models are cheap and fast enough to run continuously, the same corpus can be re-screened as new information lands, so a signal is caught when it appears rather than months later. A system that could only afford to look once a quarter would be blind in exactly the window when a problem is newest and most open to being addressed.

Auditable by design

The final argument for a lighter touch is that someone will ask how a conclusion was reached. From August 2026, most provisions of the EU AI Act take effect, and its requirements for higher-risk systems centre on traceability, record-keeping, and explainability. The principle underneath those requirements is that accountability cannot be handed to the machine.

A pipeline built mostly from deterministic, inspectable components answers that demand naturally. When a risk score can be traced back through the specific signals and stages that produced it, the assessment becomes defensible in front of a regulator, an investor, or a downstream buyer in exactly the way tightening expectations now require. A system whose every judgement rests on free-form generation cannot make the same promise. Human expertise stays at the centre of our work, supported by a platform whose reasoning can be followed rather than taken on trust.

The frontier is about fit, not size

The frontier of risk intelligence is not the largest model on the market. It is the judgement to match each part of a hard problem to the method built for it: encoder models for sorting, embeddings for context, and frontier reasoning held back for the long, complex synthesis that nothing else can do. That is how we have built a platform that goes deeper and runs at greater scale than the alternatives, while staying accurate and auditable enough to stake a decision on.

If you would like to see what that looks like applied to your own supply chains, get in touch.