Five takeaways from The AI Summit London 2026

Two years ago, after The AI Summit London, I wrote that generative AI was being talked about as ‘an iPhone moment’. I also promised to come back and see how those five takeaways aged.

Back then, the excitement was that AI could chat. This year, it was that AI can do the work.

Some of the 2024 themes have held up well. The balance between people and machines still matters. Responsible use and governance have not gone away. Multi-modal interaction, which felt fresh then, has now become the norm.

So, what’s changed? These are the five things that stood out for me this year.

1. AI has stopped chatting and started working

Unsurprisingly, the event was dominated by one word, ‘agents’.

AI started with predicting, then creating, chatting, and now it’s working. Indeed, IBM say they expect the average enterprise to be running more than 1,600 agents by the end of 2026 (Source: IBM, London AI Summit, 2026)!

Why the rush? Ironically, it’s speed. Faster and better decision-making is becoming one of the clearest advantages of AI adoption. Air Chief Marshal Sir Richard Knighton, Chief of the Defence Staff for the UK, put it plainly: “AI removes cognitive biases that haunt humans, and processes data far faster than any human team can replicate.”

There was a healthy dose of realism too. Building an agent is the easy bit, but fundamentally, an agent is really just a model running in a loop, with ‘hands’. The hard part is everything around it, i.e. provisioning, security, observability and governance once those agents start to multiply.

Running one agent is not the same as running a thousand. That is the difference between looking after one server and running a data center.

The agent harness, the orchestration and the unglamorous plumbing, is where the real work comes, and where the value opportunity and the risk are also the highest.

2. Make someone accountable for the loop

In 2024 I argued human oversight should never drop to zero, which appears to be holding, at least for now. This year London Business School’s Michael Jacobides and Catriona Campbell, author of AI by Design, highlighted a key challenge though. “Human in the loop”, they said, has quietly become a phrase we hide behind.

It sounds reassuring, but in practice, it can mean no one person is actually accountable. Their suggestion was blunt but useful; put a named human in charge of the loop. Not a committee. A person.

That person is not only accountable for the process, the inputs, and the outputs, but critically,they should also be enabled and accountable for the cost to deploy and run those agents. It’s all very well automating processes, but if you’re burning tokens at a higher cost than humans, then ROI crumbles fast. Tokenomics (i.e. token economics) is increasingly becoming both a challenge and an opportunity for organizations to invest in, much the same way as FinOps has been for cloud infrastructure.

So, if the job itself is moving up the stack, from doing the task to checking the output and designing the process, then the recurring question throughout the event was: “Will this cost jobs?” The general view was no, or at least not in the simple way the headlines suggest. Economist Torsten Slok published an analysis in May 2026 titled ‘Zero evidence of AI-related job losses’. Demand for software engineers is still rising, even as machines write more code.

[That demand however, is for senior talent, which does leave junior engineers in a challenging position in the short term.]

This is Jevons paradox hard at work again, i.e. make something cheaper and we use far more of it.

Kingfisher gave the clearest example. Its B&Q marketplace has grown from 55,000 products to more than 2.5 million in four years. No human team could possibly manage and validate all that data by hand. Agents now do the first-pass checks every day. A named person is still accountable for how those checks are designed, and for what happens when one gets it wrong.

For employees, the message is simple; if your role is one repeatable task, pay attention. If your role is a mix of judgement, context and execution, expect parts of it to change. The work won’t disappear overnight, but the responsibilities will change, and that includes being accountable for the output of your ‘digital colleagues’.

3. Your data and context matter more than your model

LLMs are becoming a commodity. On most benchmarks, the gap between frontier models is narrowing, with Chinese and open-weight models close behind. If that is right, and I think it is, then betting your business on one model or one vendor is a mistake.

So where is the advantage? In your data, and in the context wrapped around it. You have to design that context deliberately.

Humans need context to do good work. Machines do too. Without it, asking your agent to ‘bring the customer up to speed’ might have it checking their Wi-Fi bandwidth…

As we’ve talked about many times before, data gravity and data inertia mean it’s critical that IT organizations have a data strategy, which becomes the foundation for any serious AI plans.

4. Production is the only thing that counts

AstraZeneca’s Chief Digital, Data and AI Officer, Anne-Claire Gerbaldi, gave us the line of the event: “Pilotitis is a disease.” More pilots will not save you. Getting something into production, with a return on investment that you can actually measure, is what counts.

Sit down and cost out what it really takes to move from proof of concept to production. Define the return for your lighthouse projects. Pick the handful of workloads that genuinely affect revenue, cost or risk, and start there.

Also be honest about timing. In many cases, development is now the fastest bit! Rollout, adoption, change management and trust take much longer.

The most important question for users is still the simplest one: What is in it for me? If you can’t answer that, do not be surprised when the pilot stays a pilot.

5. Digital sovereignty has gone mainstream

We’ve been talking to clients about digital sovereignty for many years at SoftwareOne. We were also a launch partner of the AWS Digital Sovereignty competency, so it was good to see the topic move from side-room discussions to main-stage agenda.

The questions boards are asking now are consistent:

• Where is our data held?
• Who can get to it?
• How?

HPE’s Matt Harris summarized national capability as a five-layer cake: applications, models, infrastructure, chips, and energy. Real sovereignty means deciding which of those layers you actually need to own.

Ukraine is a great example of this. The government moved the majority of its data to the cloud just days before its data centers were hit in 2022. That decision helped make Ukraine one of the most digitally resilient nations in the world.

For most organizations, sovereignty comes down to three things: control, compliance and continuity.

You don’t need to build everything from scratch. You need to know where your data is, who can get to it, and how you keep operating when things go wrong. The key to balance is protecting what matters while still being able to use the best technology available.