Why rising AI adoption can hide a bigger question about business value, risk and competitive advantage
Across the organisations I advise, AI adoption is now relatively easy to point to. There are more licences, more pilots and more teams experimenting with the technology in their day-to-day work. Those numbers can look reassuring, but they still leave leadership with a harder question. What has materially changed in the business as a result?
For years, I have worked across both sides of this problem, building AI systems in regulated industries while separately advising organisations on emerging AI, adoption and where the technology can create value. Since early 2023, I have repeatedly seen firms with plenty of AI activity but far less evidence that the work itself has changed in a meaningful way.
Most adoption dashboards are designed to show access and activity. They can tell you how many people have a licence, how frequently a tool is being used and whether usage is rising. They tell you much less about whether an expensive process now works better, whether an important decision is being made with better information, whether a service has improved or whether AI has made previously uneconomic work practical.
When high usage hides shallow adoption
I think of this as surface adoption. AI is clearly present inside the organisation and people are using it, but much of that use still sits around the edges of existing work.
Recent research is now giving scale to something I had already been seeing through my advisory work. Deloitte UK’s 2026 GenAI Workforce Survey of 25,000 workers found that searching for information and writing emails were the two most common uses of generative AI, both at 43%. Half of GenAI users reported no time saving, while only 28% said their organisation had a policy they could find, understand and believed was current.
There is value in writing an email more quickly or producing a summary in less time, particularly when those small gains are repeated across a large workforce. I become more cautious when rising usage is treated as proof that AI investment is already creating substantial business value.
Two firms can buy the same AI and get completely different results from it. One may use the technology mainly to make existing tasks a little faster. Another may use the same underlying capability to remove a costly bottleneck, rethink specialist work, improve the quality of an important decision or offer something that previously made little commercial sense.
That difference becomes more important as capable AI gets cheaper and easier to access. If competitors can buy many of the same models and tools, the technology itself becomes less distinctive. Advantage starts to come from the choices made around it, including where AI is applied, which work is worth changing, what proprietary knowledge the organisation brings to it and where professional judgement still improves the outcome.
This is where I see asymmetry becoming important. If every firm has access to similar AI, matching competitors is unlikely to create much advantage. The opportunity comes from deciding where your organisation should advance differently, and where those different choices can create a capability or outcome that others cannot easily replicate.
The AI your dashboard cannot see
There is another part of adoption that formal metrics may never capture. Some AI use is happening outside approved systems altogether.
Deloitte found that one in three GenAI users were using the technology without their employer’s knowledge. Shadow AI sounds slightly dramatic, but the behaviour behind it is often quite ordinary. Someone needs to analyse information, prepare for a meeting or complete a task, and the approved technology either does not meet that need or the person is unclear about what they are allowed to use.
The consequences can be much more serious. The UK’s National Cyber Security Centre has warned that shadow AI can expose sensitive information, increase the risk of data breaches and intellectual property loss, and leave organisations with less visibility and control over their information.
For regulated and client-facing firms, I think this also raises a wider question around AI consent. If client, employee or commercially sensitive information enters an AI system, leadership needs to know whether that use was authorised, whether the information should have been processed in that way and what happens to it afterwards.
A policy sitting somewhere on an intranet will only take an organisation so far. The person doing the work has to understand what that policy means when they are faced with an actual decision. They also need to feel able to say, “I used AI here and I am not sure about the answer,” because that gives the organisation a chance to see uncertainty before it becomes a larger problem.
Start with the work
When I advise leadership teams, I prefer to begin with the work itself and the problem worth solving. Where is something unnecessarily expensive, slow or constrained? Where could better information materially improve a decision? What could the organisation do differently if the cost or effort involved in that work changed?
Once those questions are clearer, it becomes much easier to decide where AI belongs. You can then look at what information a system should access, what people need to understand to use it well, where professional judgement remains important and who owns the outcome.
My perspective on this was shaped by building financial AI systems after my co-founder and I had run our own private wealth management practice. That work took us into the FCA Advice Unit, where we worked on conversational AI models, and we were later invited to share our work on agentic AI in regulated environments through the FCA’s AI Spotlight.
Working in regulated environments teaches you quickly that an answer can sound entirely plausible and still lead to the wrong outcome. As AI makes information, analysis and output easier to produce, the ability to decide what deserves to be trusted, what needs to be challenged and where human expertise changes the decision becomes more valuable.
The same applies to AI training. An investment professional, compliance officer, operations team and client-facing employee may all have access to the same underlying technology, but the information they handle and the consequences of getting something wrong are very different. Generic AI training can introduce the technology, but people still need to understand what good use looks like in the reality of their own work, role and tasks.
For leaders, I think successful AI adoption now needs a broader test. I would want to know which important work has changed, whether the outcome is better, what new risks have appeared, where AI is influencing decisions and whether accountability remains clear.
Those answers are harder to fit neatly onto an adoption dashboard. They are also where I think the more valuable leadership conversation begins.