AI experimentation to enterprise impact: What’s slowing engineering?

The AI bottleneck is no longer the model: it is the organisation.

That was the underlying conclusion emerging from a recent HotTopics Food for Thought roundtable debate in partnership with Nearform. Technology and engineering leaders, spanning financial services, consulting, infrastructure, advertising, and industrial sectors, came together to discuss by now a familiar question: why are so many enterprises struggling to move from AI experimentation to enterprise impact, particularly within their engineering departments?

Beyond the answer, what unfolded was really a commentary on institutional readiness itself.

Legacy processes present structural hurdles

For all the attention given to models, agents, and copilots, the leaders in attendance repeatedly returned to a simpler observation: most organisations are still trying to graft AI onto processes, governance structures, and operating assumptions designed for a different era. 

“We thought the problem was [just] data,” one participant noted. “But often the problem is also the process itself.”

That distinction matters. Across sectors, leaders described organisations rushing to deploy AI into environments where workflows remain undocumented, fragmented, or dependent on tacit institutional knowledge accumulated over decades—what I have reported on elsewhere as crystallised intelligence. In some firms, engineering teams discovered that the process maps they were automating had not been updated for 10 or 15 years. Others described sprawling estates of legacy systems where nobody could confidently explain what certain applications still did, yet nobody was willing to turn them off.

The result is that leaders are subject to a peculiar contradiction. Boards increasingly demand visible AI progress because, in many cases, investors expect it, while the underlying organisational infrastructure remains unresolved.

“If we don’t say AI at investor day, the share price suffers,” one attendee remarked, ruefully.

That pressure is reshaping behaviour throughout organisations. 

Engineering teams have been experimenting aggressively with generative coding tools, agentic workflows, and rapid prototyping systems. Several participants described developers who can now produce weeks of work in days; others admitted they could no longer imagine returning to entirely manual coding practices. In fact, engineering teams have been some of the first to experience the broad changes brought about by generative and agentic AI given the speed with which these tools have mastered their responsibilities. They are also technically proficient and curious by nature, tinkering and tweaking with AI, often without direct oversight or formal orders. As such many teams look to this function for clues as to how the jagged edge of AI will impact their own workflows and structures.

Yet this pioneering productivity surge has introduced a second-order problem: governance capacity is not scaling at the same rate as experimentation.

Shadow AI and the C-suite

One recurring concern was the rise of what might be called “unmanaged AI enthusiasm”. Teams are building prototypes with little oversight, using unfamiliar languages, external APIs, and poorly governed datasets—shadow AI. In several cases, participants described executives themselves independently experimenting with no-code or low-code AI tooling before attempting to deploy the resulting applications into production environments.

“The functionality was good,” one leader said. “But the way it was built meant we had to rewrite the whole thing.”

This tension surfaced repeatedly throughout the discussion. AI systems are lowering the barriers to software creation while simultaneously increasing the importance of engineering discipline. The traditional craft of software engineering is not disappearing; if anything, it is becoming more strategic. But the role itself is evolving.

“Developers are becoming validators,” one participant observed.

Increasingly, engineering value lies less in manually writing every line of code and more in defining requirements, validating outputs, monitoring behaviour, and designing systems resilient enough to absorb probabilistic technologies.

That final point cascaded into a general debate on institutional readiness given that departmental budgets, headcount, and return-on-investment are front-of-mind for the CEO. Several participants argued that many executives (think: the CFO) still misunderstand the fundamental nature of generative AI itself. Traditional enterprise software was largely deterministic: identical inputs produced identical outputs. Generative and agentic systems do not behave that way.

“This is not a deterministic technology,” one attendee noted. “It is probabilistic by definition.”  

Piloting proof-of-concepts effectively

That observation explains why conversations around hallucinations remain so unsettled. In some contexts, non-deterministic behaviour is beneficial: creative ideation, exploratory analysis, and strategic scenario generation may benefit because systems generate unexpected possibilities. But highly regulated environments—financial services, compliance, healthcare, insurance—cannot tolerate ambiguity in the same way.

The challenge for the C-suite is in understanding where variability is acceptable and where it becomes operationally dangerous.
What does this mean for leaders when organisations move from proofs of concept to production systems? Participants repeatedly stressed that pilots often succeed because they operate within controlled environments. The real difficulties emerge once AI systems encounter messy operational reality: inconsistent data, fragmented permissions, undocumented workflows, and unpredictable human behaviour.

“POCs are not out there with the children,” as one participant put it.

Several leaders also argued that enterprises are now entering a more sober phase of AI adoption. The earlier excitement around simply deploying copilots or purchasing licences is giving way to more difficult operational questions. How should agents be governed? What permissions should they hold? How should decisions be monitored? How should failures be audited? Who supports systems after deployment? What happens when external dependencies change?

One financial services leader described growing concern around organisations granting broad administrative access to AI agents without fully understanding the sensitivity of the underlying data environments. Others pointed to the growing operational burden of monitoring AI systems over time, particularly as underlying datasets drift or evolve.

“You have to monitor the AI continuously,” one participant warned. “It is not a case of deploying it and assuming it stays correct.”

How leadership needs to mature

That operational maturity gap now extends to leadership itself. 

Multiple attendees questioned whether many boards currently possess sufficient literacy to govern AI transformation effectively; this view was supported by one attendee who themselves is a serial board member. In some organisations, in fact, leaders remain caught between fear of falling behind competitors and fear of introducing uncontrolled risk into the business.

“There’s a difference between understanding the technology and understanding the consequences,” one participant noted.

This may prove one of the defining executive challenges of the next five years: AI strategy is not a technology conversation but one that increasingly touches legal liability, organisational design, governance, procurement, compliance, workforce planning, and investor communication, simultaneously.

And beneath all of it, I noted, sits a deeper cultural tension: how much failure are organisations willing to tolerate in pursuit of innovation?

Engineering leaders repeatedly defended the importance of experimentation. Many argued that meaningful AI capability can only emerge through iterative learning, failed attempts, and operational discovery. Yet boards and CFOs, particularly in regulated sectors, often remain uncomfortable funding initiatives where outcomes cannot be guaranteed in advance—think back to the aforementioned deterministic-probablistic tension. 

“Experimentation requires failure,” one attendee argued. “But boards want certainty before the learning has even started.”

Perhaps the clearest signal from the discussion was that enterprises are gradually abandoning the fantasy that AI is a simple plug-in productivity layer. More sophisticated organisations are beginning to recognise that AI adoption requires redesigning operating models themselves.

Closing thoughts

AI investments are far, far more demanding than deploying a tool. Nor do the largest models or the most aggressive AI branding neatly predict the most successful AI launches; they certainly do not yet predict value for money. Yet what these debates do reinforce is that the careful combination of technical experimentation and institutional discipline provides a foundation from which governance strong enough to manage risk, engineering mature enough to support rapid iteration, and leadership literate enough to distinguish genuine transformation from performative adoption, can thrive.

As one executive summarised towards the end of the discussion: “The winners won’t be the companies that simply use AI. They’ll be the ones that know where AI actually changes the business.”

In partnership with