Oracle or Tool
The AI question many organisations skipped
A few weeks ago I was asked to help design an automated process. Take data from one system, transform it, and output to another. It was a straightforward operational problem. My first thought was to build a Copilot agent.
I didn’t assess the problem first, nor did I evaluate the options. I went straight to the AI tool, opened the Copilot Studio, and started constructing the agent. It took me longer than I’d like to admit to realise I was using completely the wrong thing.
The task needed a deterministic process, a guarantee that the same input would produce the same output, every time. What I was building was generative, a system that could produce a different result from the same data on every run. For creative work, that variance is the point, but for a data pipeline, it is a disqualification.
I scrapped the agent and built an Excel automation instead. It took a fraction of the time and it works reliably. And here is the part that matters: I used AI to help me write it. I didn’t stop using AI. I stopped treating it as the answer and started treating it as a tool to help me build the answer. Same technology, but a completely different relationship.
The question afterwards was not whether I’d made a mistake. It was why I hadn’t assessed the problem before reaching for the tool. The choice was reflexive. I realised I had not evaluated and picked wrong, I had not evaluated at all.
I don’t think that reflex is just mine.
PwC’s 29th Global CEO Survey, published in January 2026, polled 4,454 chief executives across 95 countries. 56% reported that AI had produced no revenue or cost benefits for their organisations. Not underperformed, nothing. Mohamed Kande, PwC’s global chairman, identified the root cause plainly: organisations had skipped the foundational work. They had gone straight to the tool without understanding the problem it was meant to solve.
Gartner is forecasting that over 40% of agentic AI projects will be cancelled by the end of 2027. The reasons they cite are not technical failures. They are escalating costs, unclear business value, and inadequate risk controls. Management failures, not engineering ones. Projects that started with the technology rather than the problem.
Gartner also identified something they called “agent washing.” Of the thousands of companies marketing agentic AI capabilities, Gartner estimates that only around 130 have anything genuine to sell. The rest are existing products rebranded with new vocabulary. The packaging changed, but the product did not.
None of this means AI is failing. It means the way organisations are choosing to deploy it is failing. And the reason is upstream of any individual deployment decision.
There is a framing question that sits beneath all of this, and most organisations have answered it without knowing they were being asked.
What is AI? Not what can it do, but what is it?
There are two positions. I encountered them colliding at a recent event, explained and justified by two people with very different starting points.
One position treats AI as an oracle. You describe the problem and it provides the answer. The human role is to ask clearly and accept what comes back. In this framing, AI is the destination. The measures of progress are how much you can hand over to it and how quickly.
The other treats AI as a tool. A powerful tool, that if used correctly makes knowledgeable people even better. In this framing, the human role is to assess, direct, and evaluate. AI amplifies human expertise, it does not replace the need for it.
Both positions describe something AI can genuinely do. The models are capable of extraordinary things. However, the two framings produce entirely different organisations.
If AI is an oracle then you need people who can prompt well. If it is a tool then you need people who understand the domain well enough to know what problems AI can solve, and when the tool is wrong.
If AI is an oracle then liability is ambiguous, because nobody can say who is responsible when the oracle errs. If it is a tool then liability sits where it has always sat: with the person who chose to use it and the organisation that deployed it.
If AI is an oracle then you invest in AI capability. But if it is a tool then you invest in the expertise that AI amplifies.
The framing choice determines the organisation you build. And the problem is that most organisations are not making this choice consciously. The marketing environment is making it for them.
The evidence that the oracle framing is winning by default extends well beyond the enterprise.
The FCA published the Mills Review on 6 July 2026, its landmark assessment of AI in retail financial services. Around 26% of UK consumers already trust general-purpose AI tools for financial guidance. One in five UK adults, approximately 11 million people, are open to AI making financial decisions for them. Only 40% correctly recognise that there is no formal route to redress when they rely on these tools.
The Review coined a term for what is already happening: “advice-like support.” General-purpose models are providing output that is highly personalised and that would count as regulated advice if a human delivered it. But because AI said it, the advice sits outside the regulatory perimeter and none of the protections apply. The FCA has recommended a perimeter review within three to six months.
This is the oracle framing operating at consumer scale. People are treating a probabilistic tool as though it were an authoritative source. The regulatory framework was not designed for a world in which they would.
The International AI Safety Report 2026 identified the cognitive mechanism. Automation bias is the tendency to rely on automated outputs while discounting contradictory information. It is measurable and persistent in AI-assisted decision-making. Users follow incorrect AI advice more readily when correcting it requires effort. Favourable attitudes toward AI increase the effect. The oracle framing is not just marketing. It operates along the grain of how human cognition works.
My Copilot moment illustrated something I think deserves more attention than it typically receives.
The oracle framing collapses a technical distinction that matters enormously. Generative AI is probabilistic. It produces outputs that may vary each time, even from identical inputs. For synthesis, drafting, creative work, and analysis, that is the feature. For operational processes that require the same result every time, it is not a limitation, it is a fundamental mismatch.
The question “should we use AI for this?” is actually two questions. Does this problem benefit from generative capability? And can we tolerate variance in the output? Most organisations are only asking the first. The second question does not appear in any vendor pitch I have seen.
I use AI daily and the capability is genuine. The models are improving and the practical applications are substantial. The reflexive deployment of AI as the default answer to every problem is itself the problem. Not because AI is not powerful, but because power without assessment is wasteful.
Somewhere in the last two years, the question shifted. It used to be “what problem are we solving?” It became “how are we using AI?” Those are not the same question. The first one starts with the problem. The second is a solution searching for one.
Every organisation has implicitly answered the framing question. If your board is asking “how are we using AI?” rather than “what problems do we need to solve?”, the oracle framing has already won. If your procurement process begins with the AI capability rather than the business requirement, the oracle framing has already won. If your training programme teaches prompting without teaching evaluation, the oracle framing has already won.
The most consequential AI decision most organisations will make is not which model to deploy or which vendor to choose. It is whether AI is an oracle that provides answers, or a tool that makes knowledgeable people even better. One of those framings is being installed by default, by the people with the strongest commercial incentive to see it adopted.
If you have not made that choice deliberately, someone has already made it for you.
I write about AI, cybersecurity, and technology every Friday. Subscribe to get it in your inbox.
Sources & Further Reading
PwC, 29th Annual Global CEO Survey: Leading Through Uncertainty in the Age of AI, January 2026
Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” 25 June 2025
Financial Conduct Authority, The Mills Review: AI and the Future of Retail Financial Services, 6 July 2026
Yonder Consulting / FCA, UK Retail Financial Services Consumer Survey, April 2026
International AI Safety Report 2026