Explore Our Network.

Artificial intelligence is no longer something happening somewhere else. It is already embedded in how decisions are made, information is surfaced, resources are allocated and risks are assessed. In many cases, AI doesn’t just support decisions — it quietly makes them.

That shift has been fast, uneven, and largely unquestioned.

For the last decade, the dominant narrative around AI has been efficiency: faster processing, fewer errors, lower costs, more scale. All of that is broadly true. But efficiency alone is not intelligence, and automation alone is not wisdom. The more responsibility we hand over to systems, the more important it becomes to ask who is still accountable when those systems are wrong, biased, incomplete, or simply operating without context.

Human judgement is not a weakness in AI systems. It is the missing layer.

Much of today’s AI operates by pattern recognition at scale. Large language models predict likely sequences of words. Computer vision systems classify pixels. Optimisation engines maximise for predefined outcomes. These tools are extraordinarily powerful, but they are also fundamentally indifferent. They do not understand consequence. They do not weigh competing values. They do not pause when a decision feels wrong.

That indifference matters in the real world.

In food and beverage, automated recommendations increasingly determine which venues are discovered, which suppliers are surfaced, and which businesses quietly disappear from view. In recycling and materials recovery, automated classification systems influence whether an item is reused, downgraded, or destroyed. In both cases, the system is only as good as the assumptions baked into it — and those assumptions are often invisible once the system is running.

The danger is not that AI makes mistakes. Humans do that too. The danger is that automated systems make mistakes at speed, at scale, and without friction.

This is where human arbitration becomes essential.

Human judgement introduces pause. It introduces responsibility. It introduces the ability to override optimisation when optimisation is not the right goal. In practical terms, that means designing AI systems that expect human intervention, not systems that quietly exclude it. It means acknowledging that some decisions are not binary, not clean, and not reducible to probability scores.

Good applied AI does not aim to remove people from the loop. It aims to put them in the right place in the loop.

There is a growing tendency to treat human involvement as technical debt — something to be minimised as models improve. We think that framing is backwards. In complex, high-impact environments, human judgement is not a stopgap. It is a control surface. It is the difference between a system that merely functions and a system that can be trusted.

This matters especially in sectors that intersect with the public, the environment, or physical infrastructure. When AI systems help determine how food is produced, how waste is processed, or how resources are valued, the outcomes ripple far beyond the immediate decision. A misclassification is not just a data error; it can become an environmental cost, a financial loss, or a reputational failure.

An AI-first world does not remove the need for ethics, context, or experience. It amplifies the need for them.

The most resilient systems we see emerging today are not the most autonomous ones, but the most considered ones. They are designed with escalation paths, uncertainty thresholds, and clear moments where human judgement is required — not as an afterthought, but as a feature.

That design philosophy demands a different kind of development culture. One that values context over novelty. One that treats data as something to be interpreted, not worshipped. One that recognises that intelligence is not just about prediction, but about responsibility.

As AI continues to shape how decisions are made on our behalf, the real question is no longer whether machines can think. It is whether we are willing to remain accountable for what they do.

Human judgement still matters - not despite AI, but because of it.

Ben Long CTO, EPC

For most of the internet’s life, search worked in a fairly simple way. You asked a question, received a list of links, and decided for yourself what to trust. The search engine’s role was to retrieve information; judgement was left to the user.

That model is now largely gone.

Today, search increasingly delivers answers, not options. Large language models summarise, interpret and recommend — often without requiring a click, and frequently without showing their workings. For users, this feels frictionless and helpful. For businesses and institutions, it represents a profound shift in how visibility, authority and trust are assigned.

What changed is not just the interface. It’s the power dynamic.

Traditional search rewarded structure, relevance and authority over time. If you invested in clear information, consistency across platforms and credible signals, you had a reasonable chance of being found. AI-driven search, by contrast, collapses that ecosystem into a single response. The system decides what matters, what is “good enough”, and what is excluded entirely.

This has consequences.

In food and beverage, AI-driven discovery increasingly determines where people eat, what they try and which venues are considered worth mentioning at all. In sustainability and circular systems, it shapes how information about recycling, recovery and environmental impact is interpreted by the public and by decision-makers. In both cases, being absent from AI answers is often worse than ranking poorly ever was.

The uncomfortable truth is that many organisations don’t yet understand how or why they appear — or don’t appear — in these systems.

Large language models do not search the web in the way people imagine. They synthesise from training data, live retrieval systems, structured information, behavioural signals and contextual cues. They favour clarity over creativity, consistency over novelty, and signals of trust over marketing language. They are also deeply sensitive to ambiguity, contradiction and noise.

This means that simply “having a website” is no longer enough. Nor is publishing content for its own sake. What matters now is whether your information is understandable to machines as well as humans.

The shift from search to answers also removes an important safeguard: user scrutiny. When a list of links is replaced by a single response, the opportunity for comparison disappears. Errors become harder to spot. Biases become harder to challenge. And subtle omissions can quietly reshape perception without anyone noticing.

That’s not inherently malicious. But it is consequential.

From a systems perspective, this is a move from discovery to delegation. We are delegating interpretation, prioritisation and judgement to machines designed to optimise for fluency and likelihood, not truth or consequence. In low-stakes contexts, that may be acceptable. In high-impact environments, it demands caution.

This is why applied AI work today cannot be separated from questions of governance and responsibility. If systems are answering on our behalf, we need to understand how those answers are formed, when they should be challenged, and where human judgement must re-enter the loop.

The organisations adapting best to this shift are not chasing algorithms. They are focusing on coherence. They invest in clear data structures, consistent narratives, and signals that establish credibility across platforms. They think about how their information will be interpreted, not just indexed.

They also accept a harder truth: visibility in an AI-first world is not guaranteed. It must be earned through clarity, reliability and restraint.

Search has not disappeared. But it has been absorbed into something more opinionated, more opaque and more powerful. The move from links to answers feels small on the surface. In practice, it changes who is heard, who is trusted, and who quietly fades from view.

Understanding that shift is no longer optional.

There is no shortage of experimentation in AI right now. New models, new architectures, new benchmarks and new demos appear almost weekly. Much of this work is impressive, and some of it will eventually reshape entire industries. But experimentation alone is not value — and confusing the two is becoming increasingly common.

Applied AI is not about what a system can do in theory. It is about what it can do reliably, repeatedly and responsibly in the real world.

The distinction matters because most organisations are not operating in laboratories. They are operating in environments shaped by regulation, legacy systems, human behaviour, imperfect data and real consequences. In those settings, novelty is rarely the primary constraint. Trust is.

Experimental AI explores what is possible. Applied AI focuses on what is usable.

This difference shows up most clearly once systems leave the demo phase. Experimental systems often perform impressively under ideal conditions: clean inputs, cooperative users, narrow objectives. Applied systems must survive ambiguity, edge cases and human unpredictability. They must fail gracefully. They must explain themselves. They must integrate with workflows that already exist.

In food and beverage, this can mean the difference between a recommendation engine that looks clever in isolation and one that actually helps venues be discovered without distorting choice or visibility. In recycling and materials recovery, it can be the difference between a classification model that performs well in controlled testing and one that copes with contamination, degradation and inconsistent inputs at scale.

Applied AI is constrained by reality — and that is its strength.

One of the quiet risks in the current AI cycle is the tendency to ship systems before their social or operational impact is fully understood. The pressure to deploy quickly can lead to brittle solutions that work until they don’t, often in ways that are hard to diagnose after the fact. When systems are embedded deeply enough, reversing them becomes costly, both technically and reputationally.

This is why applied AI development is as much about design discipline as it is about modelling.

Good applied systems are opinionated about where automation should stop. They define thresholds for uncertainty. They create explicit handoff points for human review. They are designed to surface doubt rather than hide it. These are not compromises; they are safeguards.

By contrast, experimental systems are often judged on peak performance rather than consistency. They optimise for accuracy on benchmark tasks, not for interpretability or operational resilience. That is appropriate in research contexts. It is risky in production ones.

The challenge is that the language of experimentation has leaked into commercial settings. Terms like “autonomous,” “self-improving” and “fully automated” are often used aspirationally, without sufficient discussion of the costs when those aspirations fall short. In practice, most valuable AI systems today are neither fully autonomous nor fully automatic. They are collaborative.

Applied AI assumes humans will remain part of the system — not because machines are inadequate, but because responsibility cannot be automated away.

This is particularly important in sectors where outcomes affect livelihoods, environments or public trust. When AI systems influence which businesses succeed, which materials are reused, or which information is surfaced, the consequences extend well beyond technical performance. In those contexts, reliability and accountability matter more than raw capability.

The most successful applied AI projects we see are also the least visible. They do not announce themselves loudly. They work quietly in the background, supporting decisions rather than replacing them outright. They are boring in the best possible way.

That is not a failure of imagination. It is a mark of maturity.

As AI tooling continues to improve, the gap between what is possible and what is appropriate will only widen. Navigating that gap is not a technical problem alone. It is an organisational and cultural one.

Real value shows up not where AI is most impressive, but where it is most dependable.

We are living through a period where data is often treated as a proxy for understanding. The assumption is simple: if you collect enough data, intelligence will emerge. If you store it, process it and model it, insight will follow.

In practice, this is rarely true.

Data is inert. It does nothing on its own. It has no intent, no hierarchy and no sense of consequence. Intelligence only appears when data is interpreted within context — and context is something machines still struggle to establish without help.

The confusion between data and intelligence is understandable. Modern systems can process volumes of information that no human ever could. They can identify patterns, correlations and anomalies at extraordinary speed. But pattern recognition is not the same thing as understanding. A system can tell you what is happening without having any sense of why, or whether it matters at all.

This distinction becomes critical in real-world environments.

In food and beverage, large datasets might show purchasing trends, foot traffic patterns or menu preferences. But without context — seasonality, local culture, supply constraints, human behaviour — those signals can be misleading. Optimising purely on what appears to perform well can quietly flatten diversity, reinforce incumbents and crowd out nuance.

In recycling and materials recovery, the risks are even clearer. Sensor data, imaging data and classification outputs can indicate material type or condition, but they do not inherently understand contamination, downstream usability or environmental trade-offs. Treating raw data outputs as “intelligent decisions” can result in systems that are technically accurate yet operationally wrong.

More data does not fix this problem. It often amplifies it.

As datasets grow, so does noise. Contradictions multiply. Biases become harder to detect, not easier. Without a framework for interpretation, large datasets can give the illusion of certainty while obscuring uncertainty entirely. The system appears confident because it is precise, not because it is correct.

This is where context-driven development matters.

Context is not metadata. It is not a label you attach after the fact. Context is an understanding of environment, constraints, incentives and consequences. It is knowing which signals matter, which ones are incidental, and which ones should be ignored altogether. Humans do this intuitively. Machines do not, unless they are explicitly designed to.

Applied AI systems that work well are almost always conservative about what they claim to know. They surface confidence levels. They highlight ambiguity. They provide space for human interpretation rather than trying to eliminate it. In other words, they treat intelligence as a shared process, not a final output.

There is also a cultural issue at play. Organisations are often more comfortable investing in data infrastructure than in judgement frameworks. Data feels objective. Intelligence feels subjective. One can be measured easily; the other requires accountability. But avoiding judgement does not remove it — it simply pushes it upstream, into the assumptions embedded in the system.

Those assumptions are rarely neutral.

Training data reflects historical behaviour, not ideal behaviour. It encodes past decisions, including past mistakes. Without careful oversight, systems trained on that data can reproduce inequities, inefficiencies or environmental harm at scale — all while appearing rational and evidence-based.

This is why intelligence cannot be reduced to computation alone.

True intelligence involves prioritisation, trade-offs and values. It requires deciding what not to optimise for. It involves recognising when a technically correct answer produces an undesirable outcome. These are not edge cases; they are central to how real systems behave in the wild.

The most effective AI implementations we see today are those that treat data as a starting point, not an authority. They invest as much effort in governance, interpretation and review as they do in collection and modelling. They acknowledge that intelligence emerges from the interaction between systems and people, not from datasets alone.

Data will continue to grow. Models will continue to improve. But intelligence — the kind that can be trusted — will always require context, judgement and responsibility.

Data is powerful. It just isn’t wise.

Automation is often framed as progress by default. If a task can be automated, the assumption is that it should be. Faster processing, fewer manual steps, lower costs. In isolation, these goals are reasonable. The problem arises when automation is mistaken for judgement.

Automation accelerates decisions. Arbitration determines whether those decisions are right.

Much of today’s AI-driven automation operates on the premise that once a system performs “well enough,” human involvement becomes unnecessary. Thresholds are lowered, oversight is reduced, and intervention is treated as friction. Over time, the system becomes invisible — not because it is perfect, but because it is fast.

Speed, however, does not discriminate between good decisions and bad ones.

When automated systems make mistakes, they tend to do so consistently and at scale. A human error might affect a handful of cases. An automated error can propagate across thousands of decisions before it is noticed — if it is noticed at all. The absence of arbitration removes the natural checkpoints where assumptions are questioned and edge cases are surfaced.

This is not a theoretical concern. It is already visible across multiple sectors.

In food and beverage, automated discovery and recommendation systems increasingly influence which venues are surfaced, which suppliers are preferred and which products are deemed “relevant.” When those systems misinterpret context — local nuance, seasonality, cultural factors — the result is not just a technical error. It is a market distortion that quietly favours uniformity over diversity.

In recycling and materials recovery, the stakes are even higher. Automated classification systems determine whether materials are reused, downgraded or destroyed. Without arbitration, borderline cases are forced into binary outcomes. The system optimises for throughput, not for environmental value. What looks efficient on a dashboard can translate into unnecessary waste downstream.

Arbitration is the mechanism that slows systems down at the right moments.

It introduces the ability to pause, escalate and override. It allows uncertainty to be acknowledged rather than smoothed over. Importantly, it assigns responsibility — not to the machine, but to the people who designed, deployed and rely on it.

Well-designed applied AI systems treat arbitration as a feature, not a failure. They define explicit decision boundaries. They surface confidence levels rather than hiding them. They make it clear when a system is unsure, and what should happen next. These design choices are not about distrusting technology; they are about understanding its limits.

One of the risks in the current AI cycle is the conflation of autonomy with maturity. Systems are often described as “fully automated” as though that represents the end state. In practice, fully automated systems are rarely the most resilient ones. They are brittle, because they lack a mechanism to respond when reality diverges from expectation.

Arbitration provides that mechanism.

It also forces a different conversation within organisations. Instead of asking “Can this be automated?”, the better question becomes “Where should automation stop?” That question has no universal answer. It depends on context, consequence and tolerance for error. In low-impact environments, automation can be aggressive. In high-impact ones, restraint is a form of responsibility.

There is also a cultural dimension to this shift. Automation without arbitration often emerges from a desire to avoid difficult decisions. Delegating judgement to a system can feel safer than owning it. When outcomes are poor, blame can be deflected onto the technology. But accountability does not disappear simply because a decision was automated. It becomes harder to trace.

Arbitration brings accountability back into focus.

It makes explicit who is responsible for outcomes and under what conditions. It requires organisations to think about governance, not just performance. And it encourages a more honest assessment of what systems are actually doing, rather than what they are assumed to do.

The goal of applied AI should not be to remove humans from decision-making altogether. It should be to place them where they add the most value — at points of ambiguity, risk and consequence. Automation handles repetition. Arbitration handles responsibility.

Without arbitration, automation simply accelerates whatever assumptions are already embedded in the system. If those assumptions are flawed, the result is not progress. It is momentum in the wrong direction.

Faster mistakes are still mistakes.

Across food systems, circularity and applied AI, a common pattern is emerging. The technology is moving faster than the structures designed to govern it, and faster than the industries trying to adopt it can comfortably absorb.

That tension is not inherently a problem. Periods of rapid change are often productive. But they do require clarity about what problems are actually being solved — and for whom.

In food and beverage, AI is now deeply embedded in discovery, logistics, pricing and promotion. Recommendations shape dining choices. Search summaries influence where people go and what they trust. Optimisation tools quietly prioritise certain businesses, cuisines and suppliers over others. None of this is new in isolation, but the concentration of influence is.

What we’re seeing is a shift from competitive visibility to delegated visibility. Businesses are no longer just competing with one another; they are competing for inclusion in opaque systems that summarise, recommend and decide on the user’s behalf. For smaller or more distinctive operators, that shift can be uncomfortable. Differentiation is harder to express when systems reward clarity and consistency over character.

At the same time, the upside is real. When designed well, AI systems can reduce friction, surface genuinely relevant options and help good operators reach the right audiences. The difference lies in how those systems are trained, governed and constrained — and whether human judgement remains part of the process.

In recycling and circular systems, the picture is similar but the stakes are higher.

AI-driven identification, classification and optimisation tools are increasingly used to determine material pathways. These systems promise efficiency, scale and consistency — all of which are badly needed. But they also introduce new risks when decisions are reduced to thresholds without context. A marginal call in a dataset can become a permanent loss in the physical world.

What stands out is how often circularity is framed as a purely technical problem. It isn’t. It is a systems problem that sits at the intersection of economics, regulation, behaviour and infrastructure. AI can help navigate that complexity, but only if it is designed to recognise uncertainty rather than erase it.

Across both sectors, we are seeing growing fatigue with hype.

Organisations are becoming more sceptical of grand claims about autonomy and disruption. There is a quiet shift toward pragmatism — toward tools that integrate, explain themselves and fail visibly rather than invisibly. This is a healthy correction. It reflects a growing understanding that AI is not a silver bullet, but a force multiplier for whatever assumptions already exist.

Another theme we see repeatedly is the importance of context-driven development.

Generic solutions struggle in environments shaped by local nuance, regulatory constraints and human behaviour. Systems that perform well in one market or material stream often fail in another. The most effective approaches acknowledge this from the outset, designing for adaptation rather than uniformity.

This has implications for how AI products are built and deployed. One-size-fits-all architectures are giving way to modular systems that can be tuned, audited and adjusted. Human oversight is being reintroduced not as a concession, but as a requirement for trust.

Perhaps the most important shift we’re seeing is cultural.

There is a growing recognition that delegating decisions to machines does not remove responsibility. It redistributes it. Boards, founders and operators are beginning to ask harder questions about how systems behave at the edges, not just how they perform on average. They are asking who intervenes when outputs feel wrong, and how those interventions are logged, learned from and improved.

This is particularly visible in organisations operating at the intersection of commerce and public impact. In these environments, the cost of getting things wrong is not just financial. It can be reputational, environmental or social. AI systems that ignore those dimensions are unlikely to be tolerated for long.

What we’re seeing, in short, is a move away from spectacle and toward stewardship.

Applied AI is becoming less about proving what is possible and more about deciding what is appropriate. The organisations navigating this shift best are not those moving fastest, but those thinking most carefully about where automation belongs — and where it doesn’t.

The next phase of AI adoption will not be defined by bigger models alone. It will be defined by better judgement, clearer accountability and systems that are designed to work with people, not around them.