The measured size of it
- 61%
- Leaders feeling more pressure to prove ROI
- up since last year
- 6%
- UK businesses with positive ROI across all AI
A service desk copilot can be live in a fortnight and leave the same engineers triaging the same tickets by hand a year later. The go-live date proves that something was configured; it does not prove that the business now works differently.
Deployment is a procurement and configuration event. A model, an agent or an assistant is made available inside a system the business already runs, with a date, a user count and an invoice. Readiness is the operating condition around it: the workflow the tool feeds, the definitions its inputs depend on, the person who owns the number it is meant to move, and the evidence that behaviour has changed.
Compare the two on one attribute and the confusion clears. A deployment question is answered by Friday: is it on, who has a licence, which team can reach it? A readiness question requires a measurement: has turnaround fallen, has rework dropped, does an exception now arrive with a recommendation attached, and is the change still there next month? Most UK programmes answer the first question and present it as though it settles the second. The result is expensive technology sitting beside an unchanged process.
The distinction matters now because buying has become easy. UK firms can purchase capability faster than they can reorganise around it. ITWeb's report on the organisational side of AI puts the point directly: "Changing how a business performs because of AI is much harder." ITWeb
The readiness test is ownership, staging and a stop rule
The first reliable test of AI readiness is whether a named person can lose their budget when the benefit does not arrive. ITWeb's report on the value architecture discussion sets out three tests for any AI idea, and all three are organisational before they are technical. ITWeb
A credible business case states the problem, the baseline, the measure that must move and the value at stake. It does not start with the tool. An organisation ready enough to execute has access to the process, the data and the decision rights for that specific use case, plus the people capacity to change the work. A named benefit owner is the person whose operational number moves when the tool works and who has the authority to stop the spend when it does not.
Staging follows from that. Prove is the gate that establishes the problem is real and the data can support a result. Pilot is the gate that tests the changed workflow with real users and a baseline that will survive scrutiny. Scale is the gate that commits money, integration and training only after the first two gates have produced evidence. A programme that cannot name its stop points is not a programme; it is an option on hope.
The pressure to skip those gates is measurable. Consultancy.uk reported that 61% of leaders feel more pressure than a year ago to prove ROI, 48% of CEOs say their organisation's culture stifles innovation, and 35% of organisations not yet seeing positive ROI cite integration as a top challenge. The same pattern appears in the adoption data. Startups Magazine reported that 64% of UK businesses have integrated AI into half or fewer of their workflows, only 6% report a positive ROI across all their AI initiatives, and 51% of UK workers say their company's workflows are not well documented and depend on institutional knowledge.
Ownership is where the gap becomes concrete. Supermetrics reported that 52% of surveyed organisations place data strategy with a team outside marketing while marketing remains accountable for AI outcomes, and 66% say they cannot prove AI ROI or can only show activity disconnected from business results. That is not a marketing problem. It is the general shape of an AI programme in which accountability has been issued without the authority or the data to discharge it. The discipline that fixes it is not anti-experiment. It is anti-unowned experiment.
Trusted data is a thin slice around a decision, not a clean-up programme
AI-ready data is not the same as data that exists. It is data whose meaning, ownership, access and quality have been settled for a specific decision, and that work is usually narrower than a data lake programme.
ITWeb's example is the antidote to dashboard optimism. ITWeb describes an executive asking an AI assistant how sales are going and being told they are up 13%, before a salesperson points out that the rise is the annual price increase and there are no new deals in the CRM. The number was present. The meaning had never been checked. Data being available does not make it ready for a model, an agent or a board pack.
That is why the useful unit is a thin slice. ITWeb describes a thin slice as selecting one meaningful outcome, checking what data and controls it needs, building it and proving the result inside 90 days, then using the result to make the next outcome easier to prove. ITWeb The slice is not a pilot for its own sake. It is a controlled way to fix only the definitions, access rules and controls that the outcome needs, and to stop before the work turns into an enterprise data clean-up with no commercial end date.
The scale of the foundation problem is easy to underestimate. Celonis reported that 82% of business leaders agree AI cannot deliver unless it understands how the business runs, yet only 9% of companies have made all their organisational data usable for AI. That gap explains why so many pilots look competent in a demo and fragile in production. The model is not short of intelligence. It is short of context.
Workload placement belongs in the same discipline. Retain keeps a regulated or latency-sensitive workload where it is, with the controls it already satisfies. Connect exposes data or a service through governed interfaces without moving the system of record. Modernise changes an application or data model in place so it can supply reliable context. Move relocates a workload to a platform that gives the scale or service the outcome requires. Retire removes a system whose data or process no longer earns its maintenance cost. The question is not whether the organisation has finished a cloud journey. It is which destination each workload needs, and what governed bridge lets identity, interfaces, monitoring, security and human review work across the mix. ITWeb makes the case that hybrid is often the fastest responsible route to a first outcome. ITWeb
Value arrives when an insight enters a workflow and changes behaviour
An AI capability that answers a question has not yet earned its place. It earns it when the answer reaches an approval, a transaction, a system update or a front-line decision, and when someone can show that the work changed.
ITWeb describes the shift from answering to doing: an agent may identify an exception, interpret the information and prepare a recommendation, while automation executes the next step and a person remains accountable for the decision that matters. ITWeb Supplier onboarding that checks continuously for missing information is a workflow design. Finance exceptions that arrive ready for review are a process design. Maintenance information that moves from a potential problem into the work order that acts on it is an integration design.
The same principle applies on the operational side. In a plant, a mine or a refinery, an insight is only valuable if it reaches the people and systems that run the operation. ITWeb's framing of the IT and OT bridge is useful because it refuses to flatten the difference: each side keeps its safety requirements, disciplines and owners, while data and decisions cross in a governed way. ITWeb Much of the industrial value of AI will be won or lost on that bridge.
Adoption then has to be measured as a business variable. Available means the capability exists and can be reached. Relevant means it fits the role and the decision that role actually makes. Repeated means it has become part of the workflow. Valuable means the operational number has moved and the movement survives a second month. A licence count cannot tell you any of that, and neither can a successful launch event.
Workforce readiness tends to erode before the ROI shortfall becomes visible. Ankura cites MIT's NANDA initiative finding that 95% of 300 generative AI pilots assessed failed to deliver value, and it argues that the adoption work is a change programme. That is consistent with what the ROI evidence keeps showing. Celonis reported that only 15% of business leaders have formal metrics in place for measuring AI returns. Without a baseline, a defined behaviour change and evidence that the behaviour is sustained, an organisation cannot distinguish a useful tool from an expensive habit.
The advantage does not go to the earliest buyer or the largest spender. It goes to the organisation that becomes reliable at converting a business problem into a measured outcome, then repeating the move. The question for leadership is not a usage count. It is where operating performance has changed, which workflow produced the change and who owns the number. If that answer is weak, the next purchase is unlikely to be the remedy. The next step is to fix the operating condition that made the last investment inert.
Sources
- AI is easy to start, the harder question is whether the business is ready to benefit from it
- The AI readiness gap and the path to ROI
- New research on the AI readiness gap, and why ownership is where it breaks down.
- Culture and operational readiness are holding back AI roll-outs
- Prove AI ROI with Measurable Results | Celonis
- Workforce Readiness Is the Missing Link Between AI Investment and AI Returns | Part 2 - Ankura.com
Source: AI is easy to start. The harder question is whether the business is ready to benefit from it, ITWeb





