LLMs are rapidly becoming the frontier for organisational transformation.
The believers are already acting. Salesforce cut 4,000 support staff declaring that AI now does half the work; Klarna shrank its headcount by some 40% in AI's name. Markets have gone further still — around $2 trillion has been wiped from software stocks in what Wall Street calls the “SaaSpocalypse”, on the bet that AI agents will make per-seat software obsolete. The vision is this: the first single-person, billion-dollar company.
And yet the evidence is stark. MIT's NANDA initiative studied 300 enterprise deployments and found that around 95% delivered no measurable impact on the P&L, with only about 5% achieving real revenue acceleration.
The Context Cap
The technology is astonishingly capable, and the limitation preventing transformation is not raw intelligence. Frontier models now score above PhD-level experts on graduate science exams and reached gold-medal standard at the 2025 International Mathematical Olympiad.
And it's not hallucination, either: on Vectara's industry benchmark the best models now summarise documents with hallucination rates as low as 0.7%, and in one clinical-summarisation study the error rate fell to 1.47% — at or below typical human levels. Anthropic's own CEO has suggested frontier models may already hallucinate less than people on some factual tasks.
The issue is memory, in the very human sense of remembering what is going on: the organisational context, the instinct for coherence with the direction of the business, and the preferences of the customers.
In an LLM, this memory is called the context window — the volume of text the model can hold in mind at once, measured in tokens (each roughly three-quarters of a word). Today's frontier models advertise windows from 200,000 tokens to a million or more. But the model holds nothing but that context: each session starts blank, and whatever isn't in the window may as well never have happened.
Missing context leads to some of the most egregious missteps. Starbucks Korea's “Tank Day” promotion in May 2026 is the cautionary tale: an AI system proposed the campaign's slogans, which — lacking any sense of local context — landed it on the anniversary of the 1980 Gwangju Uprising, where the military killed hundreds using tanks, and echoed a notorious line from a 1987 torture cover-up. The local CEO subsequently left the business.
Organisation-Scale Context
Organisational memory, or institutional knowledge, is the tacit knowledge held in people, culture and process that outlives the individuals. To truly transform an organisation with AI, this tacit information must become tangible text for the context of an LLM.
And here is the key point: organisational memory is vast — far beyond what current LLM technology can hold. A senior leader spending 30 hours a week in meetings and handling a few hundred emails a day generates, conservatively, more than 150 million words over a decade.
Assuming a tenth of that is genuine, decision-relevant memory, we can put a single leader's decade of experience at north of 15 million tokens: already beyond the largest context window any model has announced, and many times larger than anything you can deploy. A top-100 leadership team multiplies it into the billions.
Moreover, the vast majority of that is simply not written, nor could it be easily transferred to text in many cases.
Winning Applications Respect the Context Cap
Look at where AI is genuinely succeeding, and one pattern runs through all of it: the applications that work are the ones that respect the context constraint.
These are the tasks whose full context can be supplied up front; none assumes years of accumulated institutional memory.
AI-drafted legal work has already won in court: in 2026 Garfield, the first AI law firm approved by the Solicitors Regulation Authority, won what its founder called the first trial won by an AI law firm anywhere.
It works precisely because the task is memory-bounded and the context tangibly written — the solicitor's pre-trial work (witness statements, filings, correspondence) is largely self-contained, with everything the model needs fitting inside its context window.
The same pattern explains the explosion in developer “tooling”. Coding context — the codebase, the ticket — is explicit and retrievable, so it can be loaded into the window on demand rather than carried as vast and tacit institutional knowledge. That is why adoption has run fastest here, with around 85% of developers now using AI tools.
The Handover Test
The tempting response is to wait for the context problem to be solved — it is an active area of work. Subquadratic's SubQ made waves with a (contested) claim of a 12-million-token context window. Yet as we saw earlier, this is still short of the institutional memory even a single experienced leader carries, let alone a whole team.
A better, practical response is to consider context. When weighing a deployment, think of the LLM as you would a brilliant new hire on their first day: enormous raw talent and broad experience, but no specific institutional memory — more than capable of following instructions, yet with no awareness of organisational purpose or history.
Could this talented hire, on their first day, take over this task, process, decision or customer interaction? Do you have the clear instructions on hand to brief and guide them, to account for the typical scenarios? If both tests pass — bounded and written context — the project is viable.
The final and most ambitious response is to reconstruct the organisation itself. The largest impact — and the greatest competitive advantage — will come from “agentic transformation”: being first in your industry to rebuild workflows and restructure whole functions around the context constraint.
WM Commercial helps leadership teams identify the workflows where AI can safely take on work today — and redesign the operating model where it cannot.