A short video featuring Oxford’s Michael Wooldridge has been doing the rounds with a deliberately provocative claim: large language models are a “hack”.
The point is not that they are useless. Far from it. These systems can write, summarise, classify, retrieve, translate, draft and converse at a level that was difficult to imagine only a few years ago.
The point is that useful is not the same as independently intelligent.
A language model can produce a convincing plan. It can also produce a convincing explanation for a plan that does not work. It may perform well when a task resembles patterns in its training data, then fail when terminology changes, information is incomplete or the situation is genuinely unfamiliar.
That does not make the technology a disappointment. It makes it a design problem.
The costly mistake is to treat a fluent interface as proof that a system understands the work, the organisation or the consequences of getting an answer wrong.
The question organisations should stop asking
“Where can we deploy an autonomous AI agent?”
It is an understandable question, but it starts in the wrong place. It presumes that autonomy is the objective and that the model is the centre of the solution.
Most organisations are not looking to acquire artificial reasoning for its own sake. They are trying to reduce turnaround times, make expertise easier to access, remove avoidable coordination, improve consistency and help capable people spend less of their week on routine work.
Those are operational problems. AI may be part of the answer, but it is rarely the whole answer.
A better question is:
Where is human capability being unnecessarily consumed, and what combination of process, systems, information, automation, AI and judgement would make that work better?
This framing changes the work before any technology is selected. Sometimes the right answer is a clearer process. Sometimes it is better data. Sometimes it is a system integration, a workflow rule or conventional automation. Sometimes AI can make a meaningful contribution. Sometimes the work should stay human.
That last outcome should not be seen as a failure. It is often the responsible conclusion.
Fluency is not authority
Language models are good at producing language that fits the context. That can be extraordinarily useful. A model can turn a long policy into a usable first draft, help a service team find the right guidance, or make a complex knowledge base easier to search.
But fluent output has a dangerous side effect: it feels more authoritative than it is.
In a low-consequence setting, that may be manageable. A person reviews the draft, corrects the detail and moves on. In a consequential setting, the same behaviour can be costly. A plausible but wrong answer might affect a customer commitment, a financial decision, a safety issue, a regulatory obligation or an employee’s next action.
The practical distinction is simple: assistance and authority are not the same thing.
An AI system may be allowed to retrieve, draft, compare, recommend or flag an issue. Each of those is useful. The organisation must separately decide when the system can make a decision, trigger an action or communicate externally without human review.
That decision belongs to the organisation, not the model.
Build intelligent workflows, not AI theatre
The most impressive AI demonstrations are often the least useful guide to production.
A demonstration usually has a clean prompt, a narrow task, selected source material and a known route to success. Real work has incomplete information, outdated records, unusual cases, conflicting priorities, missing approvals and people who use different words for the same thing.
An intelligent workflow accounts for those conditions.
It connects the model to current and approved information. It gives the system a bounded task. It defines what evidence an answer requires. It makes permissions explicit. It provides an escalation path when confidence is low or the circumstances do not fit. It records what happened so people can review and improve the process.
The model is one component in that workflow. It is not the workflow.
This is where much of the value sits. Organisations do not need another impressive chat interface that produces variable answers. They need work that moves reliably from request to resolution, with the right information, controls and people involved at the right time.
Test the claim before scaling it
AI capability claims should be treated as hypotheses, not promises.
If a proposed system is meant to save a team time, test it against actual work. Include ordinary cases, awkward cases, ambiguous requests and situations where the right answer is “I do not have enough information.” Change familiar terminology. Remove non-essential cues. Test historical exceptions that experienced staff know to watch for.
Then measure the result.
- How often was the answer correct?
- How much review did it need?
- Did it reduce time or simply move effort elsewhere?
- Where did it fail?
- Could the failure be detected before it caused harm?
- Which decisions still require human judgement?
This does not slow innovation down. It prevents the expensive version of innovation: scaling a polished demonstration into a fragile operating dependency.
A more mature AI agenda
The organisations that gain most from AI will not necessarily be those making the boldest claims about autonomous agents. They will be the ones that are disciplined about the work itself.
They will identify where people are doing work that should be eliminated, simplified, standardised or automated before introducing a generative model. They will be clear about what the system can do, what evidence it relies on and where people remain accountable. They will build proof in a meaningful but bounded workflow before expanding its authority.
That is not a cautious alternative to AI adoption. It is how AI adoption becomes operationally credible.
The goal is not to make work look more intelligent. The goal is to make the organisation more capable, without making it less dependable.
Source: “Oxford’s AI Chair: LLMs are a HACK”, YouTube interview clip: https://youtu.be/7-UzV9AZKeU