The AI Reckoning

Why AI Is Here to Complement Your Organisation, Not Save It

Let’s be honest about something, the AI conversation inside most organisations right now is happening in one of two directions: breathless excitement or quiet dread. The board wants a strategy, the IT team is trying to work out what that even means and the frontline staff have been given a tool they weren’t asked about, a login they haven’t used and a vague instruction to be more productive. Somewhere in the middle of all of it, a senior leader is presenting a slide that says “AI-Enabled Transformation” without being entirely sure what the next twelve months are actually supposed to look like.

This piece is for that organisation; the real one, not the case study version, the one with the messy data, the paper files in an off-site warehouse, the legacy systems held together with goodwill and institutional memory, and a workforce that has heard about the next big thing before and is, quite reasonably, waiting to see whether this one sticks. The argument here is not that AI is overhyped, it genuinely is a remarkable set of capabilities, the argument is that how most organisations are deploying it or thinking about deploying it, is setting them up for disappointment. The fix is less about the technology than it is about honesty: about what you actually have, what you actually need and what AI can and cannot do when given the reality of your organisation rather than the idealised version.

The numbers are telling us something important

You don’t have to take anyone’s word for this, the data is in and it is strikingly clear. Deloitte surveyed more than 3,200 business and technology leaders around the world for their 2026 State of AI report and found that 74% of organisations want AI to grow their revenue, yet only 20% have seen that actually happen.^[1] That, is not a rounding error, that is a structural gap between what organisations expect and what they are experiencing. PwC’s 29th Global CEO Survey, drawing on 4,454 chief executives across 95 countries, found that more than half, 56%, have seen neither higher revenues nor lower costs from their AI investments and just one in eight CEOs report both positive impacts.^[2] These are not obscure metrics from a sceptical research house, these come from two of the Big Four and firms most actively selling AI transformation services. When the consultancies are publishing data showing the sector isn’t delivering, it is worth sitting up and paying attention.^[3]

None of this means AI isn’t working anywhere, it clearly is for some organisations, but the consistent picture emerging from serious research is that those organisations are the exception, not the rule, and the gap between them and the majority is not down to the technology itself, it is down to the foundations that either exist beneath it or do not.

What happens when things go wrong

Abstract warnings about AI risk are easy to dismiss, a concrete story with a price tag attached is harder to ignore. In 2025, Deloitte Australia found itself in the uncomfortable position of issuing a partial refund on a AU$290,000 government report after it was found to contain fabricated academic references and an invented quote attributed to a federal court judge.^[4]

The 237-page report had been prepared for Australia’s Department of Employment and Workplace Relations using Azure OpenAI, a fact not disclosed in the version published on the government’s website in July. A revised version appeared quietly two months later after a University of Sydney researcher, Chris Rudge, read the document and noticed something was wrong. Rudge had spotted a reference to a non-existent book attributed to a real law professor, one with a title that was outside her field of expertise and that he had never heard of, he knew immediately that it had been fabricated. Further investigation revealed multiple invented citations, references to papers that did not exist and a quote conjured in the name of a real federal court judge. Australian Greens Senator Barbara Pocock described the errors as “the kinds of things a first-year university student would be in deep trouble for“.

What makes this story important is not that Deloitte made a mistake: organisations make mistakes and fundamentally, humans make mistakes. What makes it important is the mechanism, AI language models hallucinate: they generate fluent, authoritative-sounding content that is simply untrue and they do so without any awareness that they are doing it. That is not a bug that will be fixed in the next software release, it is a fundamental characteristic of how these systems work. The responsibility for catching it sits with the human beings in the workflow “every single time” and if the incentive structure of your organisation, the time pressure, the billable hours, the imperative to ship, means that human review layer is thin or absent, then you are not using AI to produce better work, you are using it to produce more work, faster, with errors that are increasingly difficult to spot. As Connor Deeks of AI advisory firm Codestrap put it plainly: everyone has believed the fairy tale that AI is already perfect.^[5] The Deloitte Australia case is a reminder that it is not and the organisations discovering that lesson through a lawsuit or a public refund are learning it the hard way.

The data problem nobody wants to talk about

Here is the question that tends to get skipped in most AI strategy conversations, because the honest answer is uncomfortable: what is the actual quality of the data you are planning to use? AI does not create knowledge. It works with what it is given and produces outputs that reflect the completeness, consistency and accuracy of that input. Give it good data, well-structured, well-maintained, current and it has a reasonable chance of producing useful outputs. Give it the data most large organisations actually have, partial, inconsistent, scattered across systems that were never designed to talk to each other, maintained by whoever happened to be responsible at the time and it will produce outputs that reflect exactly that; confidently, fluently and without flagging any of the problems.

PwC’s research confirms that data readiness is one of the clearest differentiators between organisations seeing real returns from AI and those that are not.^[2] The companies achieving both revenue growth and cost savings are the ones that have invested in proper data foundations: integrated environments, clear governance, data that has been curated rather than merely accumulated. Most organisations have not done that work, they have years of technical debt, CRM records that were never cleaned, financial systems that do not reconcile with operational ones and customer data spread across platforms that were acquired, imperfectly integrated and never fully rationalised. Deploying AI on top of that is not a solution to the underlying problem, it is a way of producing wrong answers with impressive confidence.

But there is a further dimension to this that tends not to make it into the AI strategy presentation at all. For many larger organisations and for virtually every public sector body of any age, a significant proportion of the most valuable institutional knowledge has never been digitised, it sits in archive boxes in off-site warehouses, filed under systems that made sense to the person who designed them decades ago and are largely inaccessible to anyone who wasn’t there at the time. A local authority dealing with a planning dispute may have relevant precedents in paper files going back forty years. A university trying to understand long-term student outcomes has academic records that predate any digital system it currently runs. A financial services firm carrying out a regulatory review may find that the documentation trail simply stops at the point where paper records were never migrated. The institutional memory, the context, the nuance, the decisions made and the reasons they were made exists, but AI cannot touch any of it.

Nowhere is this more visible, or more consequential, than in the NHS. Ironically, I fall victim to this as the majority of my medical records still sit in paper form, somewhere in off-site storage locations managed by my previous GPs over the years, namely in the form of Lloyd George records. Lloyd George records, named after the Liberal Prime Minister who introduced a national health insurance scheme in 1911, are the small brown paper envelopes that served as the primary record of a patient’s medical history for over a century of NHS and pre-NHS care, containing everything from childhood vaccination records and handwritten clinical notes to historical diagnoses and test results, they represent an irreplaceable archive of patient data covering tens of millions of people.^[6]

New Lloyd George envelopes were finally discontinued in January 2021, but the records themselves, spanning more than 6,700 GP practices in England alone, remain largely undigitised. The NHS Long Term Plan, published in 2019, committed to digitising all primary care paper records and set a target for the majority of health and social care services to have electronic records in place by March 2025, that target was not met. By 2022, the national programme had shifted to a ‘Digitise on Transfer’ model, meaning records would only be scanned when a patient moved practice rather than comprehensively across the board.

As recently as 2024, NHS England was still extending its procurement framework for scanning services, with a maximum contract value of £75 million and a supplier agreement running through to 2026.^[7] As of early 2025, some individual GP practices were still awaiting the start of their nationally-funded digitisation.

The implications for any AI strategy in healthcare are significant, a GP using an AI-assisted clinical decision tool today may be working from a digital record that begins sometime in the mid-2000s with no information about what came before. The patient’s childhood illnesses, their early diagnoses, their family history as documented across decades, all of that may exist only on paper, waiting to be scanned. The AI is not wrong to work with what it has, but the clinician and the organisation commissioning the AI system needs to understand that what it has is incomplete in ways that are not always obvious and could, in some circumstances, genuinely matter.

This is not a criticism of the NHS digitisation programme, which faces a genuinely enormous undertaking, it is an observation that the ambition for AI-enabled healthcare and the infrastructure on which that ambition depends are not yet in the same place, and that gap needs to be acknowledged honestly rather than glossed over in a procurement document or a political speech.

This pattern is not unique to healthcare, it is replicated across local government, central government departments, legal services, insurance, education and any large private sector organisation with a history longer than its digital systems. The archive boxes exist, the unscanned files exist, the data that never made it into any system because it was created before the system was built, that exists too. Any honest AI strategy has to account for all of it, which means being explicit about what the AI can see and, equally important, what it cannot.

The tools are there, the people aren’t using them

Setting data quality aside, there is a deployment problem that most organisations are not being straight with themselves about. Deloitte found that while access to AI tools has grown from 40% to nearly 60% of workers in a single year, fewer than 60% of those with access actually use the tools regularly as part of their day-to-day work. Among non-technical staff, just 13% describe themselves as genuinely enthusiastic and actively choosing to engage with AI, while 21% would prefer to avoid it altogether.^[1] These are not people who tried AI and found the technology wanting, these are professionals who have looked at the tools available to them and made a rational judgement that the effort of changing how they work is not, right now, worth the benefit they can see. That is valuable feedback and it is feedback that most organisations are ignoring, because the strategy was decided above them and the assumption is that adoption will follow if you provide access.

Part of the problem is a design one. Ali Sarrafi, CEO of enterprise agent platform Kovant, makes the point that most enterprise AI tools are a year or two behind consumer products in user experience. When someone can do the same task more easily and intuitively with a tool they use in their personal life than with the one their employer has licensed, they will use the familiar one, or they will simply do the task the way they have always done it. Habit is a powerful force, and novelty only wins when it is genuinely, demonstrably better. The fix is not to mandate usage or run another training session, it is to find the places where AI genuinely makes someone’s working day better, where it removes something tedious or genuinely difficult and build from there, letting people feel the benefit before asking them to change their habits at scale.

Nobody has a playbook, even if they are pretending they do

One of the most refreshing things said publicly about enterprise AI in recent months came from Dorian Smiley, co-founder of AI advisory firm Codestrap, speaking to The Register. He said simply: “No one knows right now what the right reference architectures or use cases are for their institution. A lot of people are pretending that they know. But there’s no playbook to pull from.“^[5] That honesty is rare.

The consultant presenting the five-step AI transformation framework is not lying, they are pattern-matching from a small number of early cases onto your organisation’s specific context. The vendor whose platform promises to unlock productivity across your workforce is not lying either, their tool may well do what they say it does, in the right environment, with the right data, for the right use case. What neither of them can tell you is whether your environment, your data and your use case meet those conditions.

The organisations doing best are the ones that are comfortable not knowing, they are running genuine experiments, measuring what actually matters, building feedback loops and treating each deployment as a real test rather than a foregone conclusion. They are asking whether AI actually improves the outcomes they care about, not whether it produces more output, but whether the output is better in ways that matter to real people.

The stakes of pretending to know are rising, Connor Deeks highlights emerging pricing pressure as clients start realising that AI is being used to produce deliverables and begin asking why they should pay full human rates for work a machine assisted with. Insurers are already quietly lobbying to remove AI-generated outputs from business liability policies, the market is beginning to price in risks that many organisations are still treating as theoretical and a reckoning is coming, not for AI itself, but for the gap between what has been promised and what has been delivered.^[5]

What actually working looks like

For all the caution in this piece, there are organisations getting this right and the common threads are worth paying attention to, PwC’s data identifies the roughly one-in-eight companies achieving both revenue growth and cost reduction from AI.^[2] What distinguishes them is not the sophistication of the models they are using, it is the strength of what is underneath: clean, integrated data environments; AI deployments aligned to specific business outcomes rather than general capability; responsible AI frameworks in place before deployment; and cultures that genuinely support people in changing how they work. Among these leading organisations, 44% have integrated AI directly into their products, services and customer experiences, compared to just 17% of peers. They are not using AI as a shortcut, they are rebuilding how value is created and delivered.

The manufacturing example from Kovant is a useful illustration of the right mental model. A company with 7,000 suppliers had a restocking process that involved enormous amounts of manual coordination, emails, confirmations, follow-ups and approvals across a vast supplier network. AI agents now monitor stock levels, initiate contact with suppliers when replenishment is needed and generate summary reports. A human planner reviews and approves every purchase order before it is sent. The AI has absorbed the volume and the tedium; the human has kept the judgement and the accountability. The result is a reported 95% reduction in manual effort on that workflow, not because humans were replaced, but because the friction consuming their time was removed. That is the model: clear scope, defined outcomes, human oversight at the points where it matters and data that has been prepared to support the workflow rather than simply hoped for the best.

Jim Rowan, US Head of AI at Deloitte, captures it well: “The organizations succeeding with AI aren’t just investing in automation and algorithms, they’re investing in their people“.^[3] You cannot automate your way to a capable, motivated workforce, but you can use AI to give a capable, motivated workforce more room to do the work that actually requires them.

Questions worth asking before you deploy

PwC finds that the question most occupying CEOs right now is whether their organisation is transforming fast enough to keep pace with AI,^[2] that is an understandable anxiety, but speed without direction is not an advantage, it is a way of making mistakes faster and at greater scale. Before any significant AI deployment, the questions that actually matter are these:

• What data are we giving this system to work with and what is its quality? Is it complete, consistent and current, or are we hoping AI will paper over the gaps?
• What proportion of our institutional knowledge has never been digitised? Where does it live, in off-site storage, in filing rooms, in the memories of people who will retire in the next five years and what are the implications of AI making decisions without access to any of it?
• What specific problem are we solving and how will we measure whether we have actually solved it? Not outputs, outcomes.
• Who is accountable for reviewing AI-generated outputs before they reach a client, a patient, a decision-maker or a regulatory body?
• Have we designed human oversight into the workflow or are we relying on people to exercise it voluntarily under time pressure?
• What is our liability exposure if the AI produces an error and have we checked whether our insurance actually covers it?
• Are our people being genuinely brought along with this or are we deploying from above and hoping adoption follows?

None of these are anti-AI questions, they are the questions any reasonable organisation would ask before any significant operational change. The fact that they are routinely skipped in AI conversations says something about the peculiar mixture of excitement and anxiety that surrounds this technology at present.

A more honest conversation

Here is the version of the AI story that is actually true, this technology can do remarkable things; It can take the tedium out of work that nobody should have to do manually, it can surface patterns in data that humans would never find by looking, it can draft, summarise, translate and explain at a speed and scale no team of people could match. For organisations that have prepared well for it, it is already delivering real and meaningful returns, but it is not magic. It is not a substitute for good data, for the 110 years of NHS patient records still waiting to be scanned, for the clear governance that stops a fabricated court quote making it into a government report or for the human judgement that knows when something does not look right. It cannot read the files in your off-site archive, it cannot tell you when it is making something up, it cannot repair an incentive structure that rewards speed over accuracy, it cannot build the trust with your workforce that makes change actually stick.

The organisations that will look back on this period as a genuine turning point are the ones that resist the pressure to perform transformation and instead do the quieter, harder work of actually building something. They will start with an honest assessment of what they have, what they are missing and what AI can realistically do within those constraints, they will digitise the records that need digitising, they will clean the data that needs cleaning. They will deploy carefully, ask difficult questions and have less impressive slides at the board presentation, they will also be the ones with results to show for it in three years’ time. The rest will have the strategy deck, the tool licences and the growing suspicion that something has gone wrong, followed, eventually, by the reckoning. Some of them, like Deloitte in Australia, will be writing the refund cheque, most will simply have very little to show for a very large spend.

AI is not here to save your organisation, it is here to work alongside the people in it, on the foundations you build for it, within the limits you are honest enough to define. That is a less exciting proposition than the one being sold from most conference stages right now. Based on the evidence, it is also the one that holds true today.

Sources

Deloitte (2026)
State of AI in the Enterprise 2026.
Deloitte

PwC (2026)
29th Global CEO Survey: Leading through uncertainty in the age of AI.
PricewaterhouseCoopers

Claburn, T. (21 Jan 2026)
AI hasn’t delivered the profits it was hyped for, says Deloitte.
The Register

Claburn, T. (17 Mar 2026)
AI still doesn’t work very well, businesses are faking it, and a reckoning is coming.
The Register

Paoli, N. (7 Oct 2025)
Deloitte was caught using AI in $290,000 report to help the Australian government crack down on welfare after a researcher flagged hallucinations.
Fortune

NHS England (2019–2026)
Digitisation of Lloyd George (GP Paper Records) Programme, various updates and procurement notices.
NHS England

---

References

1. State of AI in the Enterprise 2026 Deloitte[↩][↩]
2. 29th Global CEO Survey PwC[↩][↩][↩][↩]
3. AI hasn’t delivered the profits it was hyped for, says Deloitte The Register[↩][↩]
4. Deloitte was caught using AI in $290,000 report to help the Australian government crack down on welfare after a researcher flagged hallucinations Fortune[↩]
5. AI still doesn’t work very well, businesses are faking it, and a reckoning is coming The Register[↩][↩][↩]
6. Digitisation of Lloyd George records NHS England[↩]
7. NHS England to extend digital purchasing system for digitisation of Lloyd George records HTN[↩]