Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
Rightmove’s profit warning today — to paraphrase: ‘AI is going to rewire how people shop for houses in ways we don’t yet understand and can’t yet explain but we’ll be throwing money at figuring all this out and you’re paying’ — reminded us of something.
About a week ago, Jefferies tech analyst Surinder Thind reported back from a visit to last month’s Gartner’s 2025 IT Symposium in Orlando. His note’s interesting in the universe of AI sell-side research because it doesn’t talk about 13-digit hyperscaler capex and humanoid workforces. Instead, it talks about “a disconnect between business leaders’ expectations of what AI can do and reality”.
We’re going to quote at length from Thind’s note because it’s good. It starts with an observation that, without “a complete deconstruction and reimagination of the enterprise”, there are probably no quick-hit returns on investment:
What we found at the conference was many AI projects that were being undertaken seemed like they were being done in silos. While this might generate measurable productivity gains in this first generation of initiatives, it will likely not work for the next generation. Many software vendors that were advertising AI solutions, including AI agents, seemed focused on single use cases/workflows. This is not entirely unexpected as the past decade plus has allowed software companies that solve single pain points to thrive. Looking ahead, we believe the real value from Agentic AI will come from an agentic mesh—a decentralized architecture that allows multiple autonomous AI agents to collaborate and act across different systems, tools, and language models. But the technology isn’t quite there yet.
Extrapolating from this, we get the sense that procuring all the different AI solutions from software vendors may end up being a significant waste of enterprise spend. We acknowledge there is an “AI arms race” underway, but we’re not sure if it is winnable at this point.
The big takeaway of the note is that it’s very hard to make AI deliver anything useful.
According to a McKinsey survey cited by Jefferies, 80 per cent of companies have deployed generative AI in some form for at least one business function. Of the AI adopters, 80 per cent still report no material contribution to earnings from the deployments.
Jefferies also cites Gartner estimates that, on average, an AI deployment costs $1.9mn upfront. However, for every 100-day AI deployment there’s an extra 25 days of staff training once the system’s in place, followed by 100 to 200 days of “change management” to make sure everything’s working as promised.
Gartner estimates that, on average, for every one AI tool purchased, an organisation will see 10 ancillary, hidden costs they did not anticipate (examples include licensing, legacy integration, managing access credentials for AI agents, comparison testing, security, etc). This means that business leaders who go into the AI procurement process thinking it will cost $1.9mn and take 100 days to implement are actually seeing it cost much more and take much longer, making ROI difficult to achieve.
Gartner also estimates that, by 2027, 40 per cent of AI projects will have failed. The unstated corollary is that, for projects started in 2025, the attrition rate will be much, much higher.
Another underestimated problem is data quality, says Thind.
In many conversations we sensed frustration at the inability to scale from pilots due to concerns around data (and governance) issues. According to an AWS presenter, approximately 70% of IT budgets are spent on managing legacy systems, legacy systems cause a 6-18 month delays for rolling out new features in software products, and around 40% of software developer time is spent on managing technical debt.
The above becomes a much more serious issue when companies try and implement Agentic AI, because if an organisation does not have clean data and quality data governance policies, AI agents cannot be trusted to execute autonomous functions and make real business decisions based on data that the humans themselves deem untrustworthy.
Then there’s the problem of turning hypothetical efficiencies into actual revenue growth and cost savings. A gen-AI deployment in the sales department might allow salesfolk to spend less time on the phone, but to deliver tenable business value the deployment has to be cheaper or better than humans.
Jefferies cites a Gartner survey that finds 74 per cent of CFOs are seeing productivity gains from AI, but only 5 per cent have managed to cut costs and just 6 per cent saw any kind of revenue uplift.
“With so many executives struggling to understand the technology, specifically its potential and its limits, we come away from the conference a bit more confident that the AI disruption narrative will take longer to play out, and perhaps in ways that we may not currently understand or appreciate,” writes Thind:
Because of the top-down pressure from Boards that we heard about in some conversations, it’s hard not to assume significant resources are being spent on products or solutions that ultimately may not further the needs of the business.
The Jefferies analyst argues that, eventually, companies wanting to extract value from AI will give up chasing ROI gains through on one-simple-trick stuff and bring in consultants, IT outsourcers, etc, who’ll completely rebuild their tech stack and transform how they function.
Maybe that’s what Rightmove’s doing? It’s certainly going to be spending outsourcer money, with an extra £12m investment in product going through the P&L and £6m of capitalised expense. And at pixel time, with Rightmove’s investor day just starting, the shares are down 16 per cent.
Techbros, be careful what you wish for.
