AI At A Crossroads: Why Trust, Talent, Cost And Delivery Will Define The Next Phase Of AI

According to Dr John Bates, CEO of enterprise document intelligence pioneer Doxis, the next phase of AI adoption will be defined not by ever-larger models, but by how responsibly businesses deploy them

One of the biggest misconceptions about LLMs is that they behave like conventional software like our ERP and accounting packages. But, says Doxis’s Bates, “ChatGPT is, by definition, a non-deterministic system: you can never be sure it will give you the same answer to the same problem at 2pm that it did at 9am.”

In fact, this is true of all LLMs, where identical inputs can produce different outputs. While that flexibility underpins their creative capabilities, it also creates challenges for enterprise use. In regulated sectors, unpredictable outputs can lead to poor decisions and undermine confidence in AI systems.

“Where inaccurate outputs can lead to poor decisions, this is actively dangerous in business contexts,” Bates believes. “But it’s also a challenge for automation, especially in regulated industries, where predictability is critical,” he says.

The best solution, he believes, is to move to a hybrid approach, where traditional rule-based systems are combined with LLMs. “Ultimately, even the most advanced AI should be viewed as a tool,” he points out, “and organisations must ensure they are deploying the right technology to the right task.”

Avoiding Danger And Enhancing Trust

Some of the most successful techniques Bates has seen in client deployments involve a composable approach that uses the best of AI with the best of what came before.

This should include bolstering the reliability of answers from agents by dynamically switching between models, combining emerging techniques retrieval-augmented generation (RAG) with vector search for advanced analytics, and using rules-based systems where determinism is required, Bates points out.

Another key area where support for AI is essential is contextual AI, using historical data and temporal patterns, which LLMs are not good at.

The Reality Check On AI Spending

As AI adoption accelerates, organisations are also grappling with rapidly increasing compute and token costs. “Without a doubt, token usage is a major issue,” says Bates. But, he adds, experience from previous technology cycles can help us become more realistic in our expectations and so, better plan for how to maximise the opportunity AI presents.

“As Gartner says, we are at the peak of inflated AI expectations, with heavy spending on AI experimentation being the consequence,” he notes. However, he expects businesses to become far more disciplined over the coming years. Rather than measuring AI success through usage (token spend) alone, organisations will increasingly focus on tighter governance and cost controls as businesses gain a clearer understanding of the real underlying economics and ROI of practical enterprise AI.

Now is the time, he suggests, for CIOs to start trying to apply ‘outcome maxing’ metrics, where AI use is tied to measurable business results.

This shift towards cost optimisation could have implications beyond enterprise IT budgets. As businesses become more selective about AI spending, Bates believes today’s lofty valuations across major AI companies and even broader markets may face increasing scrutiny, potentially leading to some market correction as a result.

Moving Beyond The Hype

Bates believes the industry is entering a more mature phase of AI adoption. While today’s systems are remarkably capable, they should not be mistaken for any kind of AGI (artificial general intelligence) or a Technological Singularity.

“I think it’s clear to both the line of business manager and the CIO and indeed the information manager that AI in 2026 is powerful and transformative, but its best chance for sustained ROI is to be used carefully, especially because of trust and the black-box nature of these systems,” Bates notes.

Ultimately, the challenge for organisations is not chasing the latest model, but deploying AI safely, efficiently and strategically. That means, he predicts, that the winners of the next phase of AI will not necessarily be those with access to the biggest models. “They will be the organisations that  deploy these capabilities responsibly and effectively.”

The author is global CEO of European-headquartered Doxis, the document intelligence company