AI doesn’t fix broken processes – it accelerates them
A bold statement on the face of it but we are in a time of extraordinary change that is happening at unprecedented speed. The explosion of AI over the past two years is changing how people work and how they think. Using AI as a reasoning partner has empowered people to do more faster; changing how things work in manual or system driven workflows has never been easier but, and it’s a big but, accelerating a flawed workflow will only do harm. So, are you sure that your processes are up to scratch? Are they current and reflective of what people actually do? Or are they shelfware that you spent some effort on some years ago to get through a quality audit and then ignored?
Process rarely takes centre stage in strategic conversations. That has worked when the pace of change has permitted complacency. With the acceleration in pace that AI has created that complacency is not just a bad idea; for many organisations it is an existential risk. For a long time, work just happened, customers were served and teams simply adapted when formal processes did not reflect how the business needed to operate. Those adaptations were often pragmatic, even necessary, but rarely documented and almost never analysed.
The advent of AI has changed everything. AI does not enter a blank organisational canvas; it enters existing workflows, draws on existing data, affects existing decisions and depends on (or bypasses) existing controls. If those foundations are unclear, outdated or plain wrong, AI won’t make them better. It will simply deliver bad outcomes more quickly.
In May 2026 Roc Technologies brought together several senior leaders from technology and business roles working across the operational process landscape for a frank roundtable conversation. The challenge that emerged was consistent: organisations are under pressure to move faster, particularly as AI adoption and expectations accelerate, but many are constrained by outdated, fragmented, inconsistent or poorly understood processes.
In fact, this operational reality gap – the accumulated distance between how work is documented and how it actually gets done – has become one of the most significant constraints on organisations looking to innovate. Independent research now puts figures on it: a widely cited estimate, repeated by RAND in 2024, puts AI project failure above 80%, and MIT NANDA’s 2025 study found that around 95% of generative-AI pilots delivered no measurable P&L return within six months.
Process is not the only explanation on offer. The same studies point at unready data, leadership that frames the wrong problem, weak integration, and – in MIT’s case – a six-month payback test few investments would pass; its least-quoted finding is that externally sourced tools reached deployment around twice as often as those built in-house, on self-reported figures the authors themselves caveat. None of this displaces process. It sits underneath them all: each problem is harder to diagnose, and harder to fix, when nobody agrees how the work actually gets done. That said, the constraint is not uniform. Where the work is genuinely new, or the tool is bought whole and bounded, the process question is smaller. It bites hardest where AI is pointed at established, high-volume, cross-team work.
This paper explores why the reality gap matters, how it constrains AI adoption, and what organisations need to do to build stronger foundations for secure, effective innovation.
Every organisation evolves over time – no organisation wants to stagnate. The reality gap opens up when a business evolves but its understanding of how work gets done doesn’t. It’s easy to spot – it shows up in duplicated effort, unclear ownership, excessive approvals, inconsistent data, manual reconciliation, or decisions that depend too heavily on individual knowledge.
These workarounds quickly become the norm, and they stop being questioned – leaders assume they are part of the cost of doing business, and teams compensate through effort, experience and goodwill. But the cost does not disappear: it accumulates, and it compounds.
Left unchecked, the reality gap does not just slow the business down. It changes what the business is capable of: harder to govern, because dependencies are not fully understood; harder to get value from technology investment, because new tools are layered onto old complexity; and harder to improve, even when the appetite for change is high – because the foundations that change would need to build on are themselves uncertain.
All too often, process documentation describes how work is intended to happen – or how people believe it should happen – rather than how it happens in reality. Adaptations begin as practical responses to specific problems, but they become part of how the business operates.
This matters because organisations make transformation decisions based on the process they think they have. They procure technology against the documented workflow, design governance around the formal operating model and measure performance against assumptions that may no longer be true.
The impact in these scenarios is not always immediately obvious – low user adoption of a new technology or ROI that is never quite delivered.
When AI enters that environment, it is working from a map that no longer matches the territory. The danger is not just automating inefficient processes – it is that AI operates at a pace that leaves little room for course correction. By the time a problem becomes visible, the consequences may already be difficult to contain.
AI makes the cost of poor process visible for the first time. But businesses should not see that as a threat. It is an opportunity to fix what was always broken.
If process is to play a more strategic role, organisations need to move beyond traditional process mapping. Mapping still has value, but on its own it produces a snapshot: useful when created, quickly overtaken by operational change.
What organisations actually need is a living understanding of how work happens: connected to systems, data, people, controls, suppliers, performance and risk.
That distinction matters because the reality gap doesn’t just affect efficiency – it affects resilience. Knowing a process exists is not enough. Organisations need to understand what it is for, what data it relies on, where third parties are involved and what happens if something fails.
“An up-to-date process map can show the sequence of activity and support operational resilience through linkages to an understanding of purpose, dependency, risk and impact.”
– Ian Leaver, Head of Automation, Roc Technologies. Roundtable, May 2026.
Process intelligence also changes how organisations approach AI adoption. Instead of starting with a tool and asking where it can be applied, it frames the question as which process needs to improve, what is preventing better performance, what data is involved, what risk needs to be managed and could a process be simplified before it is automated?
Finally, there is the human aspect. If a process is automated using AI, AI cannot be held accountable for the outcomes. People are needed to observe and provide the sanity check: when to challenge an input, when to escalate an exception, when to apply context. As AI becomes more embedded in operational workflows, organisations need to be clear on the role and timing of that judgement.
Taken as a sequence – understand, simplify, automate, observe – this gives leaders a stronger foundation for deciding where technology can create value, where caution is warranted and where human judgement must remain central.
Before investing further in AI, automation or transformation, senior leaders should be able to answer a small number of questions with confidence. The following are a useful starting point – not as a compliance checklist, but as a genuine assessment of whether the operational foundations are ready for what is being asked of them.
Or do they describe an idealised version of the business that people have quietly moved away from?
And what is the risk if those people are no longer available?
And are their dependencies – including third-party and data dependencies – understood in enough detail to manage disruption?
And who is accountable for them – including decisions that AI is now influencing or generating?
If these questions cannot be answered easily, the operational reality gap is likely wide enough to constrain the organisation’s AI ambitions regardless of the quality of the technology it adopts.
That is not an argument against moving forward. It is an argument for combining forward momentum with an honest assessment of where the business really is.
None of this argues for a grand clean-up before anything can begin – those programmes have a poor track record, and the map is stale by the time it is finished. It argues for the opposite: scope tightly, get the specific foundation right for the specific thing you are automating, and use AI itself to reveal the real process rather than the documented one.
AI has changed the urgency of this conversation, but it has not changed the fundamentals. Organisations still need clear work, trusted data, accountable decisions, resilient operations and people who understand how to act safely and effectively. What has changed is the cost of getting those things wrong.
The tolerance for unclear ways of working is shrinking. In the past, the reality gap slowed organisations down. Now, it also makes them harder to govern, harder to automate and harder to protect.
For the leaders closest to this challenge, the starting point is not a technology decision. It is an honest conversation about operational reality: where processes reflect how work actually happens, where they do not, and what it would take to close that gap.
That conversation is harder than identifying an AI use case. It is also more valuable.
The organisations that will get the most from AI will not simply be those that adopt it fastest. They will be those that understand their operations clearly enough to move fast, govern confidently and scale innovation without accumulating the kind of risk that only becomes visible when something goes wrong.
AI may be the catalyst. Process is the advantage.
Roc Technologies works with organisations on the operational foundations that make safe, effective AI adoption possible.
The figures in this paper are drawn from independent, published research and appraised for quality. Where a study’s method is contested or a figure is often mis-attributed, that is noted so the claim can stand up to scrutiny. Evidence was reviewed to July 2026. Roc Technologies advises organisations on operational process and AI adoption and therefore has a commercial interest in this subject; the sources are cited in full so readers can weigh the argument independently.
Frequently quoted for the finding that more than 80% of AI projects fail, around twice the rate of non-AI IT projects. That figure is not RAND’s own: RAND repeats it in its introduction as an existing estimate, footnoted to earlier industry reporting. RAND’s own contribution is a root-cause study drawn from 65 practitioner interviews conducted in late 2023, which named leadership problem-framing, data quality, technology-chasing, weak infrastructure and the limits of the technology itself. Projects using pre-trained large language models were outside its scope. Best read as “the large majority” rather than a precise figure.
Around 95% of generative-AI pilots delivered no measurable P&L return, with failure attributed to integration and workflow fit rather than model quality. Success was defined narrowly, on a six-month P&L basis. The same report found that externally sourced tools reached deployment around twice as often as those built in-house (roughly 67% against 33%), on self-reported figures the authors caveat as not controlling for confounders.
Over 40% of agentic-AI projects are forecast to be cancelled by the end of 2027 on cost, unclear value and weak controls; 60% of AI projects are expected to be abandoned through 2026 where data is not AI-ready. Model capability is not among the named causes.
Organisations reporting significant AI returns are markedly more likely – on McKinsey’s 2025 figures, close to three times as likely – to have fundamentally redesigned their end-to-end workflows. The relationship is reported as a correlation: the survey does not establish that redesign came first, or that it caused the return.
42% of companies had abandoned most of their AI initiatives by mid-2025, up from 17% across 2024. The 42% was reported in June 2025 as a part-year figure, not a full-year result.
Documented “reference” processes routinely diverge from the actual process observed in system logs, and the divergence is measurable – the evidential basis for the reality gap. The foundational study is Rozinat, A. and van der Aalst, W.M.P. (2008), ‘Conformance checking of processes based on monitoring real behavior’, Information Systems, 33(1), pp. 64–95, which established fitness and appropriateness as measures of the distance between a documented model and observed behaviour.
Head of Automation
Ian is responsible for end-to-end process automation for both internal and external business requirements, leading agile delivery from initial discovery to live automated operations.
Universities are fiercely, proudly distinctive. Their academic character, research strengths, student culture and civic relationships are not peripheral to what they do; they are central to the value they create. That distinctiveness has real weight: for students choosing where to study, for researchers deciding where to work, for employers seeking partners, and for the communities universities serve. In a sector where institutions are increasingly required to compete for students, funding, talent and reputation, protecting their differentiation is a strategic necessity.
Across every sector, CIOs are being asked to deliver transformation everywhere, all at once. Modernise platforms. Introduce AI. Integrate systems. Reduce costs. Improve experience. Strengthen security.
It’s a long list, and those who try, rarely succeed.
We talk a lot about collaboration, but much of what passes for teamwork today is really noise. Endless threads, half-attended meetings, and long days spent replying digitally instead of deciding. Call me a rebel but perhaps it’s time to regain some space to think, to act, and to work together with intent.
Over the past decade, the role of the university CIO has steadily evolved from infrastructure leader to strategic partner. But as collaboration across the sector becomes more structured and strategically important, that role is expanding again.
Collaboration has long been part of Higher Education’s institutional DNA. Universities share research infrastructure, co-author papers across continents and form networks of expertise that transcend borders. Knowledge has rarely been confined to a single institution.
For years, cyber security in higher education has been focussed primarily on data – institutional research, personal information and systems. Yet for university CIOs today, the more pressing issue is broader and, increasingly, more human.
End-of-support moments often trigger defensive thinking. Get compliant. Keep the lights on. Minimise disruption. These are valid concerns — but they don’t reflect the potential value on the table.
Partnerships sit at the heart of how most organisations now operate. Whether it’s suppliers, delivery partners, or internal teams, very little is achieved in isolation. Yet partnership is also one of the areas where good intentions often outpace good design.
The best kind of change is often the kind you don’t notice happening. Systems run smoother, processes feel easier, people stop sighing at their screens – and suddenly everything just works a little better.
