AI adoption and the business of keeping up

There is something rather appropriate about discussing artificial intelligence in the House of Lords. Surrounded by centuries of history, participants spent the evening considering technology that apparently becomes obsolete somewhere between the starter and the main course. Even the menu proved unreliable: the advertised fish arrived as chicken. For once, the unexpected output had nothing to do with AI.

 

The conversation was frank, occasionally uncomfortable, and refreshingly willing to challenge itself. Beneath the excitement sat a familiar concern: businesses want the benefits immediately, while the people responsible for managing the risks are still trying to establish what, exactly, has been deployed.

 

Check Point’s 2026 Cloud Security Report provided the backdrop. In the survey of 1,042 cybersecurity and IT professionals, 70 per cent reported GenAI in production, but only 14 per cent had policies both enforced and audited. Just 5 per cent reported full visibility into AI tools and services (these are separate survey findings, not percentages of production users). They demonstrate a considerable gap between enthusiasm and operational control.

 

The figures prompted concern, but they do, of course, require interpretation. A lack of fully enforced and audited policies does not necessarily mean an organisation has no governance whatsoever. The practical question is whether its controls cover the activities taking place and whether anyone can demonstrate that they work.

 

The discussion quickly moved beyond simply demanding more policies. Annual reviews struggle when capabilities change continually. For instance, a carefully managed and approved corporate service offers limited reassurance if employees also upload information into personal AI accounts. Personal devices have acquired an enthusiastic younger sibling in the workplace: personal AI tools.

 

Questions about data location also exposed a need to distinguish contractual protections, model training arrangements and actual information flows, rather than simply assuming safety.

 

Retrospective restrictions create another problem. Once people become accustomed to a useful tool, withdrawing it can encourage them to find less visible and more dangerous alternatives. The answer needs to include an approved route that actually works for employees. Otherwise, governance becomes something people navigate around while getting their jobs done.

 

Before deciding how to govern AI, however, organisations need an agreed definition of the technology itself. Traditional automation, generative tools, internal assistants and autonomous agents are repeatedly bundled together, despite presenting different problems. An employee summarising a document and a system influencing a customer’s access to finance deserve different scrutiny and controls, while calling everything “AI” does very little to help.

 

Participants distinguished between tools used by employees and AI embedded within products or business processes. The latter can affect customers directly, making service reliability, fairness and potential harm central considerations. Governance needs to follow the purpose and consequences of a deployment, rather than simply follow its marketing label.

 

The same difficulty applies to incidents. An inappropriate chatbot response, confidential information escaping and disruption to a critical service are hardly interchangeable. Definitions determine escalation, accountability and the evidence needed to understand what happened. Without them, committees can spend considerable time discussing risks while talking about entirely different things.

 

Ownership was equally untidy. Data teams, security, legal functions and business leaders all have legitimate interests, but distributing involvement does not automatically establish responsibility. The board needs to understand the risks it accepts, and those governing deployments need sufficient authority to act throughout the business. Giving someone AI governance because they already manage data is purely an allocation of work, rather than proof of readiness.

 

Business value received some welcome scrutiny too. Counting licences, token usage, prompts or newly created agents demonstrates that something is being used but does not establish that the organisation is better off. Participants questioned whether productivity claims translated into measurable savings or improved outcomes and whether total costs, future pricing and energy demands were receiving enough attention. A dashboard full of adoption metrics can look wonderfully impressive while answering remarkably little.

 

The commercial pressure was nevertheless real. Organisations face expectations to reduce costs, improve services and keep pace with competitors. Security teams therefore need to explain how controls support those objectives, with evidence that decision-makers can use when weighing opportunity against exposure.

 

There was genuine disagreement about the security implications. One argument held that familiar controls still matter most: classify information, restrict access and understand where sensitive data belongs. If an AI assistant finds information that an employee should never have accessed, the permissions problem may have existed long before the assistant arrived.

 

The counterargument was that speed, autonomy and connections between systems change the consequences. Weaknesses that were previously difficult to discover can become much easier to exploit. Individual safeguards may work as designed while leaving dangerous gaps between them, and both arguments deserve attention. Sound fundamentals reduce exposure, but they need testing against what AI systems can actually do.

 

Model updates also complicate that assurance. A replacement model can produce different behaviour even when the surrounding application appears unchanged. The discussion questioned how organisations could test those changes and maintain confidence over time. Approval at launch cannot settle every subsequent risk; that work needs attention between deployments.

 

Thankfully, the evening offered more than an increasingly elaborate list of things to worry about. Good news examples of AI automation included analysing large volumes of medical information, improving threat intelligence, simplifying access to HR guidance and reducing repetitive governance work. These were useful reminders that there are worthwhile applications beyond an executive instruction to “do something with AI”.

 

However, even useful applications raise uncomfortable questions. An AI assistant that makes workplace information easier to find can also reveal employees’ private concerns through their questions. Predictive capability creates choices about what an organisation should do with information, as well as what it can do. Ethics belongs inside the design conversation, before an apparently clever feature becomes somebody else’s regulatory or legal problem.

 

Education emerged as perhaps the strongest practical theme. People, including senior leaders, need to understand what these systems do, where their limitations lie and how to question confident answers. Training should help cautious employees recognise sensible opportunities while giving enthusiastic users reasons to pause. However, it was agreed that neither blanket fear nor unquestioning trust is a particularly useful operating model.

 

There was also concern that outsourcing every summary and answer could weaken the thinking needed to evaluate the results. Producing meeting notes is useful; understanding the meeting still requires attention. Organisations should be careful about confusing the possession of information with having actually learned anything from it.

 

The closing conversation returned to practical steps: define the use case, understand the data, assess the risks, establish ownership and test before expanding. Some organisations can begin with lower-risk applications and build confidence deliberately. Others are already dealing with widespread adoption and need to recover visibility. Their starting points differ, and so their roadmaps should, too. Clear frameworks could also give recurring AI meetings direction, connecting proposed uses to agreed risks, responsibilities and decisions.

 

Shared learning matters. Aviation provided an analogy for investigating failures and making lessons available to others, although participants also recognised the additional challenge of active adversaries in cybersecurity (something not particularly relevant in the aviation industry). Interestingly (and perhaps unsurprisingly), nobody around the table had a finished playbook. That makes collaboration among peers and partners vital, particularly when commercial pressure encourages everyone to pretend they are further ahead than they really are.

 

A central takeaway was that responsible adoption starts with people capable of making informed decisions, as the technology will continue moving unnaturally quickly. Organisations must understand AI’s purpose, challenge its claims and create conditions in which it can deliver value safely – preferably before the next dinner course arrives and all our assumptions are out of date again. 

 

---

To learn more, please visit: www.checkpoint.com