How To Make Decisions While AI Disrupts Itself

So far, strategic corporate decisions about AI have mostly begun with OpenAI vs Anthropic vs Google.

However, once onboard a frontier lab train, it becomes clear that no one knows all the stops on the way to Artificial General Intelligence (AGI), including the labs themselves. The “track” also remains under construction, costing trillions of dollars each year and causing rumours of an investment crash – just like the literal railroads of old.

This has turned out to be a very unusual journey indeed.

Alongside advising our clients, I have been a fellow traveller during the last few months.

We halted a classic SaaS implementation of our own recently to adopt a new agentic stack. However, just as we began constructing it, our strategic AI partner released its own stack in beta, triggering another revision of our already revised plan. These were all good interruptions, but unanticipated destinations, nonetheless.

AI-partner-led disruption is actually relatively linear in that new developments from a single provider interrupt ongoing investments. Messy, but manageable.

But what if these disruptors get disrupted?

What if innovation accelerates outside our strategic partnerships?

What if we have backed the wrong horse?

How do we make any decisions amid this uncertainty?

Emerging disruptions

On these questions, I have news that is good, bad, and uncertain all at once.

I do expect a wave of disruptions deeper than anything we have seen since the launch of ChatGPT in 2022. Moreover, I cannot be sure these will all be led by the dominant frontier labs – at least initially. This is a recipe for complexity.

However, on the upside, I anticipate developments that will address some fundamental issues with current AI, making the technology more effective, efficient, and reliable.

The first category of innovations is technical. Transformer-based Large Language Models (LLMs) are extraordinarily capable, but also computationally wasteful. They repeatedly reprocess vast amounts of information simply to generate the next token in a sequence.

In contrast, world models– most publicly associated with Yann LeCun, the former Chief AI Scientist at Meta, but also under development by Google and NVIDIA - aim to simulate the real (physical) world rather than merely predict language responses.

LeCun thinks these models will eventually replace transformer-based LLMs that he considers “dumber than a cat”, but most think they will augment rather than displace, especially for robotics.

Another threshold about to be crossed is the viability of open-weight models for corporate use. Until recently, OpenAI, Anthropic, and Google provided proprietary systems of unarguably superior breadth and depth.

However, openly available models from Mistral, Alibaba, DeepSeek, and Moonshot are increasingly approaching the performance thresholds of these leading platforms, especially for coding – perhaps the most productive application of AI to date.

Consequently, more companies are likely to adapt, fine-tune, and run these powerful models on their own infrastructure at dramatically lower cost – assuming they have the technical chops. This will considerably diversify and decentralise the generative AI landscape, shaking up frontier model economics in the process.

The leading AI labs are already responding by evolving beyond standalone LLMs towards intelligent orchestration platforms capable of reasoning, coding, searching, planning, and taking action. This move is consolidating hundreds of previously standalone apps, models, and workbenches.

As a result, niche products that were briefly best-of-breed are frequently leapfrogged by new features from leading proprietary models just a few months later. Over time, many previously independent products will survive only as plug-ins.

By moving toward orchestration platforms, the frontier labs are opening a new competitive gap with open-weight models, which can compete on raw capability but not on breadth of offering.

What does this mean?

Over the next 24 months, I believe we will see leading AI models improve in speed, efficiency, accuracy, and breadth – but, for the first time, not only through scaling.

Open-weight offerings will improve choice, economics, and security, while the best SaaS will remain in our organisations in headless form – orchestrated by rapidly maturing agentic orchestration capabilities.

If I am right, there is a lot to like about this direction.

The challenge is navigating the complexity along the way.

What should leaders do?

1. Stay on the frontier-lab train. Currently, the most straightforward way to leverage the innovation curve is through a leading provider. Their rate of development and competitive need for enterprise revenue means they are newly aligned with corporate users' requirements for measurable results.

2. Take an electric bike. Don’t rule out innovations from outside the primary relationship, even if they must be accessed by different means. However, any exception must deliver sufficient strategic value to justify the deviation. Make it exceptional but possible to take this route.

3. Get ready for open-weight models. Even with a leading proprietary platform at the core, most businesses will find specialised reasons for internally hosted models tuned to their specific requirements and secured within their borders. Ask your teams to experiment with current options, so they build experience for 2027 and beyond.

4. Create a clear SaaS strategy. The software that has powered your business for decades is being disrupted. Commission a review of what should survive the AI transition and what needs to go. Negotiate new economics for what stays.

5. Over and above all these points, create a leadership culture that is adaptive. No one knows the AI future, and there will be multiple twists and turns over the next few years. Many plans will not reach their conclusions before a new one is required. Adaptation will be as important as excellence. It already is.

The business embrace of generative AI has been a challenge, but just as we are getting comfortable with the first disruption cycle, the second is underway.

We need to thread the needle between the now and the next while building adaptable teams that can embrace continual innovation.

The good news is that the next generation of generative AI platforms is likely to bring more benefits and fewer risks.

All aboard.

If you'd like to listen to an audio version of this article, please click here.