Previous

The first was pace. Generative AI has been adopted faster than any previous technology at the same stage, large language model (LLM) capability is improving exponentially, and the unit cost of intelligence is falling more rapidly than computing or electricity costs did at similar points in their development.

The second was who captures the value generated by AI: the providers of frontier models or the companies that turn raw intelligence into commercial outcomes for their customers?

Our view is that AI will make the strong stronger. The businesses that benefit most will be those with a defensible position built on proprietary data, deep domain knowledge and products that are model agnostic.

Three of our portfolio companies shared what this looks like in practice.

Trust beats intelligence

Customers will only delegate important work to AI when they can hold it to account and trust and verify what it produces.

Altia, a leading provider of intelligence and investigation software, has instructed a specialist, independent legal team to test the accuracy of its AI agent. The team will also help Altia build legal checks into the product, so that users can review AI-generated output in confidence and in full before accepting it.

Permissions, audit trails and escalation processes must be designed in from the start if automation is to be introduced safely into high-value work.

“AI can't hold a state of mind. It can't hold a belief. So, a human has to stay in control.” 

Sarah Saxton-Jones, Director of Product, Altia

AI agents operate, people govern

AI agents are now users of software alongside people. That raises new questions for product teams: which data and functions to make available through APIs, how to manage access, and how to charge for work done by an agent rather than a person.

As the team at commercial finance software provider Lendscape explained, there are no definitive answers to many of these questions. Those making progress are investing in stages, gathering evidence as they go to help boards make informed decisions about the next steps. They are also clear about where AI adds value – introducing a “learning” step into each cycle of agent behaviour so that performance improves over time.

“If it's deterministic, it's a rule. You don't need an LLM for that. The context has to be hard to gather. If it's easy to gather, our customers will do it themselves.”

Steve Taplin, Chief Product Officer, Lendscape

The bottleneck has moved

Generating code is now the easy part, but faster coding does not mean faster delivery. The bottleneck has moved in both directions: upstream, to agreeing what the right product should do; and downstream, to reviewing, testing and releasing far more code than before.

The challenge is now engineering discipline: agreeing intent, establishing consistent standards across a team and defining a focused brief so that AI produces useful results. It also means redesigning review processes to keep pace with new ways of working, while ensuring the end product meets the original business objectives.

In regulated markets, building is rarely the constraint – deploying safely is.

"AI can write the code, but it can't make up for a weak organisation or a broken SDLC, it just amplifies the issues you already have. Getting that right is what makes AI work, and it's not easy."

Stephen Hender, Chief Product and Technology Officer, WSD

The value is in what you build around the model

The three sessions highlighted a common message: the value of AI in software lies not in the raw intelligence that LLMs provide, which is fast becoming a commodity, but in what businesses build around it – products customers can trust founded on industry-specific data and knowledge, systems that let AI agents do work of real value, and the engineering discipline that turns faster code into reliable releases.

AI can create value through new products, changing what customers pay for, and through productivity, changing how businesses operate. Both rest on the same disciplines of reliability, governance and measurement.

We are very grateful to our speakers and portfolio companies for sharing their experiences and contributing to the discussion.