Info-Tech: AI transformation needs real expectations, not speed

As artificial intelligence continues to reshape boardroom agendas worldwide, organisations are starting to confront a difficult truth: despite unprecedented investment and enthusiasm, AI deployment is proving far harder, slower, and more expensive than the hype ever suggested.

In a keynote at the Info-Tech Research Group's LIVE event in Montreal, Jack Hakimian, Senior Vice President, Research and Advisory Services, said organisations should "curb their AI enthusiasm" to avoid misplaced expectations and possible low returns on investment.

According to an August 2025 report by the Massachusetts Institute of Technology, 95 per cent of enterprise AI pilots fail to deliver measurable business impact. This has caused some concern as businesses are investing billions of dollars into AI tech as the boom chugs along.

Hakimian said that even the strongest frontier models continue to fall short on multi-step problem-solving. Hallucination rates remain a major challenge in AI adoption. While the issue is widely acknowledged, Hakimian said it is still misunderstood. 

"We tend to believe that hallucinations are a scarce occurrence," he said. "They're not."

Citing OpenAI's own evaluations of its O3 and O4 Mini reasoning models, he noted hallucination rates ranging from 33 to 51 per cent. Even the company's newest large-scale models show non-trivial failure rates, with GPT-5 reporting a failure rate of around 9.6 per cent, according to Info-Tech.

While generative tools have transformed content creation and information retrieval, Hakimian added, their weaknesses become far more visible when organisations attempt to automate complex work.

According to Geoff Nielson, Senior Vice President, Brand, Reach and Influence, the market is now at a turning point, where expectations are beginning to realign with the genuine complexity of the technology.

He sees this shift playing out across countless organisations attempting to implement AI at pace. Many, he says, are feeling pressure from boards to expedite AI adoption, only to discover that integration demands much deeper technical planning, governance, and architectural work than anticipated. 

"There's been this broad belief that AI is easy - just sprinkle AI on top of existing processes and expect 20 or 40 per cent efficiency gains," he explains. "But the reality is that AI sits on the same foundation as every difficult IT project that came before it. It's no easier, and in many cases it's harder."

The limits of LLM reasoning

Info-Tech has listed multi-agent orchestration innovation as one of its tech trends for 2026. Nielson said that while many organisations are still in exploration mode, the sector's innovators have already moved ahead.

"At the innovator level, 81 per cent are currently using some form of agentic AI right now, and another 27 per cent are looking at pursuing growth via AI agents by 2027," said Nielson.

He described multi-agent orchestration as the operational breakthrough, enabling teams to start small with specific, high-value processes and then expand into broader systems where AI agents collaborate. Nielson stresses the need to start specific and build out from there.

"It all comes down to the notion of multi-agent orchestration. What are the business processes we have in place right now that we can automate? And how can we start in very specific, high-value spots?" added Nielson.

Accuracy gaps compound rapidly within automated workflows. Hakimian used a simple example: a single task performed by an AI agent at 90 per cent accuracy may appear strong on its own, but when chained into multi-step processes, minor errors accumulate. 

A three-step workflow at 90 per cent each achieves an effective accuracy of around 73 per cent in the end results. "Consequential processes don't require 70 or 80 per cent accuracy," he said. "They require Six Sigma performance - 99.9997 per cent."

A pragmatic path forward

Hakimian stressed that generative AI remains a transformative technology, noting that progress over the past five years has grown "leaps and bounds" in this decade. But he said the next era of development will likely move away from general-purpose LLMs towards smaller, task-specific models and alternative architectures designed for hierarchical reasoning.

"No single model is good at everything, and therefore the future is probably going to look very different. For AI, you're probably going to have purpose built models. You're going to have architectures that do not necessarily look like today's LLMs, and perhaps not even based on a Transformers architecture."

For organisations navigating the present landscape, Hakimian offered straightforward advice. Teams should be pragmatic, choose projects based on clear value, ignore internal pressure to chase hype, and prototype before committing to full-scale deployment.

"We are bullish on AI - just not on AGI," he said. "If you want to succeed in your AI transformation, remain grounded."

Images courtesy of Info-Tech Research Group. From left: Jack Hakimian and Geoff Nielson.