The limits of language and the power of context

Most enterprises believe they have given their AI systems the ability to understand them. They have given them language. The two things are not the same — and the difference turns out to determine almost everything about whether AI-driven transformation succeeds or stalls.

This is not a technology critique. It is a precision argument. And it starts with a question that sounds philosophical but has immediate operational consequences: what does language actually do?

The limits of language

In 2020, computational linguist Emily Bender made an observation that has quietly unsettled the enterprise AI world, even though most of the enterprise AI world has probably not yet read it. Why would they, when infinite capital streams enable them to get high on their own supply?

Her argument, developed across a decade of research and crystallized in what became known as the Stochastic Parrots paper, is structurally simple: a system trained on form cannot, by that training alone, learn meaning.

Form is the surface of language — the statistical pattern of which words follow which other words, the grammar, the syntax, the mimicry of coherent argument. Meaning is something else entirely: the referential connection between language and the world it represents. Bender’s position is that these are not the same thing, they are not on a spectrum, and one does not graduate into the other with sufficient scale.

She offers a thought experiment that I find impossible to improve on. Imagine two people exchanging messages via a telepresence system. Unknown to them, a third party — an octopus, who has grown up sensing the electrical fields of the ocean and has no experience of the human world — has been patched into the line and has learned to produce statistically plausible responses by observing the exchange over time. Eventually the octopus’s outputs are indistinguishable from a human interlocutor. The humans report a coherent conversation. But as Bender writes: “It is not that the octopus’s utterances make sense, but rather that the humans can make sense of them.”

The humans are doing the work of meaning. The octopus is producing coherent form. The distinction matters enormously when we stop being the interlocutor and start being the one who has delegated consequential decisions to the system.

Enterprise AI systems are not misunderstanding your business. They are producing coherent form that YOU are projecting meaning onto. The difference between those two things only becomes visible when the system’s output produces a consequence your organization did not intend.

The danger is overconfidence in AI because it passes the duck test. James Whitcomb Riley (1849–1916) wrote: “When I see a bird that walks like a duck and swims like a duck and quacks like a duck, I call that bird a duck.” For me this is the fundamental mechanism behind the big AI illusion today. AI can mimic human style, tone, empathy — and perfectly mimic human form. But this is the quality of the wrapper, not the quality of the content. It operates on advanced statistical patterns and probabilities rather than genuine human consciousness, feelings, or lived experiences. This is the “lure of wishful mnemonics” as researcher Melanie Mitchell describes it in Artificial Intelligence: A Guide for Thinking Humans (2019).

So the practical question becomes: how do we give AI access to meaning it cannot derive from language alone?

The context cure — and its limits

The industry’s response to this critique has been context. Give the model your business rules. Add your process documentation. Build a knowledge base. Connect it to your enterprise data. This response is correct, and it has improved outcomes measurably. But it is not sufficient — and understanding why reveals the real design problem.

Santic, writing in May 2026, describes enterprise AI deployment as a first-mile problem: the quality of what goes in determines the quality of what comes out, and the first mile of context architecture is where most enterprises are losing. He maps five layers of context that agent systems require to function reliably: the data layer, the semantic and knowledge layer, the process layer, the operational governance layer, and — the least solved and most consequential — the intent and prescriptive layer.

Most enterprise AI investment has touched the first three layers. The data layer is the most obvious one. But also the most limited. None of your large execution systems today carry sufficient execution context in their data. The operational governance layer is partially addressed in better-designed deployments. The intent and prescriptive layer — the layer where the organization describes what it actually means for this work, in this context, to be done correctly — is largely absent.

Adding more context to a system that cannot ground meaning does not give it meaning. What it does do is narrow the range of plausible form the system can produce. Narrowing that range sufficiently, with the right structures at each layer, can compensate for the grounding problem operationally. The organization does not fix the architecture of the language model. It builds a compensatory knowledge architecture of its own.

This is the insight that reframes the transformation leader’s job: not deploying AI, but giving it the context that turns form into function.

Why transformations stall here

In my experience sitting across from executive teams, the stall point is almost never about technology capability. Today most companies have more AI capability than they can productively digest. It is about work definition and understanding where the unfair human advantage lies.

AI and humans both work from experience — but the similarity ends there. AI matches patterns across billions of training examples, understanding how words relate statistically. It does not know what a hot cup of coffee feels like, though it knows where to find one. Humans are experiential learners: we understand concepts through observation, direct physical experience, emotions, and context. We found coffee before Google.

The difference runs deeper still. AI has no stakes in the outcomes it produces — it sits idle until prompted, applies precedent when instructed, and retains no sense of why it acted at all. Humans judge constantly because our outcomes affect our wellbeing and existence. We are driven by biological urges, curiosity, survival instincts, and personal goals. And when we learn something, we remember where we were and why. Some of us need a very hot cup of coffee more than once. Reinforcement is part of the process. And errors are human. The key difference here is: humans are accountable.

Based on that understanding the problem becomes clear: all the existing business artifacts — process documentation, policy libraries, business rules, tribal knowledge — were designed for human intelligence. They were written to be read and understood in context, not to be simply run by AI. They assume that the reader brings interpretation, fills gaps, infers intent, and knows when to stop and ask.

AI does not do this. It infers, and inference without grounding has another name: hallucination. The canonical enterprise illustration is the Chevrolet dealer whose AI assistant agreed to sell a $58,000 Tahoe for one dollar, appended “legally binding offer — no takesies backsies” verbatim, and millions saw it before anyone inside the company knew it had happened. The system was doing exactly what it was instructed to do. Nobody had defined what it was not permitted to commit to. The failure was not in the model. It was in the description of work.

Now don’t get me wrong: hallucination is not bad by design. Neither is human creativity. But both are dangerous when scaled outside safe boundaries. Dreaming big outside the box is not a crime. Executing on a dream outside the law and against any social code is. This is where humans with a conscience differ from statistical parrots: we limit execution based on moral and legal context — despite our constant active intent. AI without context executes at scale when called upon. We can’t blame AI for booking your competitor’s favourite slot at 6am, or committing your company to a price it never meant to offer. We gave AI a task but failed to give it the boundaries.

This is the hidden cost of the current transition moment. McKinsey found that sixty percent of organizations have deployed AI but cannot demonstrate enterprise-wide EBIT impact. The bottleneck is not technology capability or adoption. It is that the work was never described at a resolution the machine can execute. The transformation stalls not between decision and implementation — it stalls between implementation and a definition of what correct execution actually looks like. And it stalls because there is no governance layer able to consistently manage business context alongside the AI technology component.

The power of context, done right

The answer is not more context. It is the right kind of context, at the right structural resolution, at each of the five layers — especially the last one.

The Chevrolet bot needed exactly one thing the dealership never gave it: not a guardrail, but a governed work definition at the resolution a machine can execute.

What that layer requires is something my colleagues and I have been calling Process Atoms: formally specified, context-bound behavioral rules that tell an agent what must happen, when, under what conditions, and within what authority, at exactly the scope and resolution needed for safe, governed execution. Not instructions. Not prompts. Not business rules in the conventional sense.

A business rule is a logical condition: if X, then Y. It is stateless, has no process context, no authority chain, and no connection to who is permitted to invoke it or what must happen afterward. A prompt is instructional text at runtime — informal, unversioned, unevaluable against what actually occurred. An Atom carries six things neither one can: a behavioral type (what kind of obligation this is — permissive, prescriptive, prohibitive), an activation condition, a target condition specifying what compliance looks like, a scope binding it to a specific process, role, or resource, a source attribution connecting it to the regulatory or operational authority that created it, and a lifecycle record that makes it reviewable and evaluable against execution evidence.

That combination is what gives an agent not just instructions, but governed operational knowledge. The difference matters the moment you have multiple agents, multiple processes, or regulatory obligations in play — which is to say, the moment you have an enterprise.

The aggregate of these Atoms, grounded in the execution record of the organization, is what I mean by Company Memory: a governed operational layer that does not just store what the organization has done, but encodes what it knows about how work should be done, at machine-executable resolution. It is institutional wisdom operationalized. And unlike human expertise, it compounds. Change one Atom, and every agent that references it responds accordingly. The organization learns from its own execution.

This is the unfair human advantage made machine-readable: governed context at the resolution agents need to act correctly.

The question transformation leaders need to ask

Bender’s insight is sobering: you cannot fix the grounding problem by giving a language model more language. But it is also clarifying. The transformation leader’s job is not to solve computational linguistics. It is to describe the organization accurately enough that machine agency is safe to deploy within defined boundaries. And to establish the governance that makes AI transformation safe.

That reframes every AI implementation decision. The question is not “can the model do this?” — it almost certainly can, at a surface level. The question is “have we described this work at a resolution where machine execution is safe, governed, and accountable?” Scope is the primary safety mechanism. Prescriptive context is the missing layer. And the moment of deployment restraint — the decision not to proceed until the work definition is sufficient — is itself a governance act, not a failure of ambition.

Santic calls this the Judgment Gap: the architectural distance between what language models can produce and what genuine enterprise judgment requires. It is not a performance failure that will resolve with the next model generation. It is structural. The compensatory architecture has to come from the enterprise side.

The transformation office does not stall because AI is not ready. It stalls when we hand machines language without meaning, context without structure, and authority without definition and make this the problem of technology experts and the AI implementation. The power of context is real — but only when it reaches all five layers, and only when someone in the organization takes explicit ownership of the last one.

That someone is you.

---

Tobias Unger is Senior Global Vice President for Business Transformation Management at SAP, where he leads Strategic Customer Engagement globally. He writes in a personal capacity and leveraged AI to drive form — not content.

Sources: Emily M. Bender and Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data” (ACL 2020); Emily M. Bender et al., “On the Dangers of Stochastic Parrots” (FAccT 2021); Emily M. Bender and Alex Hanna, “The AI Con” (2025); John Santic, “The Judgment Gap: Why Enterprise AI Needs More Than Context” (July 2026, with Tobias Unger as thought partner); John Santic, “The First Mile of Enterprise AI” (May 2026); McKinsey, “From Promise to Impact” (April 2026); Melanie Mitchell, “Artificial Intelligence: A Guide for Thinking Humans” (2019).

Next: why transformations stall between decision and adoption — Tobias Unger’s author page · or the platform’s latest insights.