AI has been fast to arrive and fast to evolve, and even faster to punish the enterprise that chose too early. There is no longer a single model to adopt or a stand-out model to choose from because there’s an entire market of them and they all have their pros, cons and moments in the trendy sun. For many companies, their AI strategy sits either within an ecosystem like Google or Microsoft, or they have invested in one or two market leaders, using one engine to draft, another to analyse and a third to write code while paying a hand over token for the privilege.
The challenge facing most organisations right now is that you want the power and innovation inherent within these models, but you don’t want to be beholden to them and stuck within ecosystems that may yet fall behind. Right now, the models themselves keep leapfrogging each other every few months while the pricing behind them is still being decided. And, on the other side of price, there’s the risk of lock-in – companies need to know that they will have the freedom to change models in six to eight months’ time but the choice available today may not be around in a year’s time.
The numbers capture this strange AI-powered indecision. Worldwide AI spending is already expected to reach $2.59 trillion by the end of the year as companies increasingly widen their wallets to accommodate the cost of the technology. Many still can’t show the return on this spend. The price of a token keeps dropping while the invoices keep rising – a problem brought on by increased demand and usage which is pushing the price up for companies faster than they can save on lowering subscription fees. Capability is cheap, but certainty is not and this is where companies need to find their footing because it is certainty that turns a model into value.
Companies are looking for intelligent ways of adopting multi-modal AI strategies that reroute and reduce costs, especially as the economics of AI keep changing. AI isn’t absorbed into a subscription anymore; it is consumption billed by the token which means that the only question companies need to answer is “What is the return on investment per token spent on a model that’s fit for the business?”. And very few companies can answer that yet.
Microsoft’s answer to the problem has been to stop competing with frontier models and to instead start absorbing them. Instead of pouring everything into building a model that beats the frontier labs and adds to the growing ‘AI war’, Microsoft has started to turn Copilot into an ecosystem designed to route the work to whichever engine handles it more efficiently and then layered this with a degree of autonomy. At the bottom sits the ask-and-answer retrieval most companies already know, but above this sits a semi-autonomous layer where an agent works across applications to carry out a chain of tasks rather than a single instruction. Essentially, Copilot is becoming an integrated space that blends knowledge retrieval, daily artefact creation, semi-automated and automated experiences.
Another aspect of this strategy is the rapid explosion of smaller language models that companies can use and fine-tune on their own data until the model is genuinely their own. A technical process with an unglamorous name – post training reinforcement learning – it delivers a very practical outcome. The frontier models won’t do this, they are too large to train cheaply and too expensive to run, they also don’t necessarily open themselves up to this. The other value add for these smaller models is that they don’t run as high a risk of being switched off by political decisions made in another country or repriced out of the South African affordability bracket.
That said, South Africa is sitting in a uniquely beneficial position. Arriving a step behind isn’t necessarily the handicap it looks like because the expensive lessons being learned in Europe can be learned without paying for them. One of the most important being the need to keep the human in the loop – companies shouldn’t be chopping heads, but rather amalgamating workflows so they have the luxury or choice and the flexibility to change at will.
The end game right now is to invest in an ecosystem approach, much like Copilot with its expansive integration of multiple tools and capabilities. Microsoft’s auto model-routing approach ensures the business can dynamically benefit from OpenAI, Anthropic and its proprietary AI without having to invest in tokens and subscriptions across multiple platforms. It ensures you get value and choice, and that has become very important to AI longevity and sustainability right now.