We spend a lot of time talking to technology companies about what is coming next.
In most sectors, “next” might mean a development expected in a year or two. With the advent of AI, it can arrive halfway through a research project. A feature that felt distinctive during the initial briefing, may be widely available by the time the findings are presented.
That speed is exciting, but is also becoming one of the industry’s biggest risks. Something that we saw first-hand in our AI agent security research with Akeyless.
The question is no longer simply what AI can do. It is whether our ability to understand, govern and safely deploy it is developing at the same pace.
In our view, AI companies do not need to stop innovating. But they do need to stop assuming that maximum speed is the right speed for every product, capability and use case.
AI capability is accelerating quickly
The UK AI Security Institute has found that performance in some areas of frontier AI is doubling roughly every eight months. Its testing also showed that, by mid-2025, advanced models could complete hour-long software tasks more than 40% of the time, compared with less than 5% in late 2023.
This is not just improvement in the quality of written answers or generated images. AI systems are becoming better at coding, using tools, solving scientific problems and completing longer sequences of work with less human involvement.
Adoption is moving quickly too. Stanford’s 2026 AI Index found that 88% of organisations were using AI in some form. At the same time, documented AI-related incidents increased from 233 in 2024 to 362 in 2025, while reporting against responsible AI benchmarks remained inconsistent.
The gap between capability and control is becoming difficult to ignore.
The threat to people is not limited to sci-fi movies…
Discussions about AI risk can quickly move towards machines becoming more intelligent than humans and acting beyond our control.
Those scenarios remain uncertain and contested. But uncertainty should not distract us from harms that are already easier to see.
AI is being used to support fraud, scams, cyberattacks, manipulation and the creation of convincing false content. Overreliance on AI can also lead people to accept incorrect outputs without sufficient scrutiny, particularly when a system sounds confident or appears more capable than it really is.
The 2026 International AI Safety Report divides the risks into malicious use, system malfunctions and wider systemic effects. It also notes that the pace of progress could change dramatically if AI systems begin to accelerate AI research itself.
That does not mean a catastrophic outcome is inevitable. It does mean the potential consequences become harder to dismiss as systems gain more autonomy, access and influence.
So, should AI companies slow down?
A blanket slowdown is unlikely to be practical. AI development is global, competition is intense and increasingly capable models are becoming more widely available.
There is also a genuine cost to moving too slowly. AI can support medical research, improve accessibility, strengthen cyber defences and remove repetitive work. Delaying beneficial applications carries risks of its own.
The better answer is selective slowing.
Where an application is low-risk, transparent and easily reversed, businesses should be able to experiment quickly. Where an AI system can access sensitive data, change software, make decisions about people, move money or affect critical infrastructure, the pace should be set by evidence rather than the launch calendar.
Before deployment, businesses should be able to answer some fairly simple questions:
- What genuine user need does the new capability address?
- How could it fail, be misused or behave unexpectedly?
- Where is human approval still meaningful?
- Can the system be monitored, reversed or shut down?
If those answers are unclear, slowing down is not a failure of innovation. It is a sign of responsible product development.
What working with AI and technology companies has taught us
One of our observations from researching technology markets is that companies can become understandably focused on what their product is capable of doing.
Customers often look at the same technology differently. They want to know whether it is reliable, where the data goes, how much control they retain and what happens when something goes wrong.
More capability does not automatically create more confidence. In an emerging market, it can create more uncertainty.
Research helps identify where that confidence gap sits. It can show which applications customers are ready to adopt, where they expect human oversight and which safety claims they actually find credible.
This matters commercially as well as ethically. As AI buyers become more experienced, companies that can demonstrate control, transparency and an understanding of real customer concerns are likely to be more trusted than those competing on speed alone.
Speed should be earned
AI companies do not have to choose between progress and caution.
They can move quickly when the risks are understood and the consequences are limited. They should move more carefully when systems become more autonomous, harder to reverse or capable of causing significant harm.
The most responsible model may be fast learning, staged deployment and slower independent scrutiny where the stakes are highest.
AI development will continue to accelerate. The businesses most likely to succeed will not necessarily be those moving fastest. They may be the ones that know when to apply the brakes.
To explore how attitudes towards AI adoption, trust and responsible innovation are developing in your target market, contact Callum Budd at callum@mra-research.co.uk.
