The Promise vs. The Reality
Last year was meant to be the year of the agent. The point at which AI would move beyond personal productivity and start to reshape enterprise productivity, supercharging business processes, and finally unlocking the long-promised AI ROI. The uncomfortable truth is that, so far, the hype hasn’t quite met reality. Which raises the obvious question: are agents a busted flush?
The answer is yes…and no.
The Core Technical Constraint
There are real challenges in designing and deploying agents at scale inside operational business processes, which is where they need to be if you want to see ROI and transformative change. At the heart of almost every agent architecture today sits a large language model. These models are generative by design, and as a result, much of today’s agentic engineering effort is focused on constraining, corralling and shaping that generative behaviour to achieve the consistency, predictability and compliance that business processes demand.
We intuitively understand the limitations of LLMs, but we often forget them when the results appear convincing. LLMs do not understand content; they generate language based on mathematical and statistical patterns, and they are exceptionally good at it. But understanding, intention and judgement are not native capabilities. This matters because many of the failure modes we see today: hallucinations, brittle reasoning, inconsistent outcomes, are direct consequences of this limitation.
Recognising it is not an argument against agents; it is a guide to where agentic approaches can be safely and productively applied today, and where they should not yet be trusted, particularly in areas of customer proximity, regulatory exposure or irreversible financial impact.
Architectural Immaturity
Context windows remain limited and transient. Long-lived memory and durable context are still immature. Approaches such as RAG were a major step forward, and enterprise context graphs and newer orchestration protocols improve things further. But these remain workarounds, or at least works in progress, rather than a complete solution to enterprise-scale context and assurance.
The Intention Gap
The deeper issue is not the availability of data. Enterprises are drowning in data. The issue is intention.
Most enterprise documentation exists to demonstrate rigour and defensibility, often for regulatory or audit purposes; it is a poor reflection of how the enterprise operates day to day. Organisations are very good at documenting what they do and how they do it. They are far less good at documenting why. Purpose, goal hierarchies, acceptable trade-offs, escalation thresholds and risk appetite are rarely explicit. They live in human judgment, experience and organisational norms.
Humans infer intention naturally. Agents cannot. This “intention gap” is one of the least discussed but most significant constraints on scaling agents into real business operations.
Theoretical vs. Practical Automation
Agents are often positioned as the breakthrough that finally unlocks the judgment-heavywork that traditional automation could not reach. The contradiction is resolved by drawing a distinction that is frequently blurred in enterprise AI conversations: the difference between what work is theoretically automatable and what is practically automatable in today’s enterprise context. Agentic technologies dramatically expand the theoretical frontier. They do not automatically expand the practical one.
There is also a scale problem that is easy to underestimate. Even a single regulated process,such as customer onboarding, decomposes into multiple capabilities, identity verification, eligibility checks, authentication, and compliance screening, each of which breaks down into multiple processes and execution steps. It is at this lowest level, where work is still manual, that agents can create value. It is also where context and intention matter most.
Agentifying already-automated steps does not seem sensible and likely delivers little to no return. The opportunity sits squarely in the manual work, and it is here that both legacy automation and agents struggle without explicit context and intention.
The Governance Paradox
In practice, even small to medium enterprises may end up with hundreds of agents operating across execution processes, all of which require monitoring, compliance and orchestration. This will need more agents existing purely to keep the broader agentic system safe and compliant. These agents are a necessary operational overhead, but not a source of direct ROI.
Here we see a paradox: there is low appetite currently for full agentic autonomy, but without it, at least in these governing and orchestration agents, the whole agentic architecture cannot effectively scale. Human-in-the-loop is baked in, either by gating scale or diluting the economic case.
Why Waiting Is the Wrong Response
Faced with these realities, it would be understandable to adopt a wait-and-see posture, to slow down investment and wait for the gap between AI agent promise and proof to close. That would be a mistake. The technical limitations of agents, predictability, memory, and reasoning are being addressed at speed. Week by week, new techniques emerge: better orchestration, improved tool use, more structured agent frameworks.
It is reasonable to bet that the technology will mature faster than most enterprises can re-shape themselves to absorb it. What is far less mature is enterprise readiness.
A New Category of Actor
At their core, operating models have historically had two building blocks: people and systems. People do things. Systems do things. People and systems govern each other. Agents change that, they introduce a third category: non-human actors with delegated, bounded autonomy.
Agents are not people, and they are not traditional systems. They can make decisions and take actions, but the ethical, legal and organisational accountability for those actions still sits with humans. There are emerging engineering patterns for building and controlling agents. There are far fewer patterns, arguably none that are widely accepted, for governing agents as operational elements of an enterprise.
Engineering governance and operational governance are not the same thing. Being good at the former does not materially change the risk profile of the latter. This creates unanswered questions for decision-making authority, auditability, controls, explainability and compliance. These questions will not be resolved by better prompts or better models alone; they require new thinking about operating models, accountability and governance.
The Work That Matters Now
This year may not be the year of large-scale agent deployment into run-the-business operations. But it should be the year of preparation. There is meaningful work that can be done now, largely adjacent to the technology itself.
Building data and context platforms that assume large agent ecosystems. Expanding governance models to accommodate non-human actors. Identifying where agents can safely augment work today and where they should not. Beginning the work of making enterprise intention explicit and portable, rather than implicit and human-bound.
Enterprises have time, but the work is non-trivial. Those who start now will be ready to move deliberately when the technology matures. Those that do not may find themselves technically capable, but operationally flat-footed, watching the opportunity arrive before they are ready to take it.
Ready or Not
The agents may not be here at scale yet. But make no mistake, they are coming. All of which assumes an enterprise culture willing to delegate elements of judgement to non-human actors,a shift that could prove harder than any technical constraint.
