Innovation without Disruption
Why enterprise AI needs stable infrastructure, commercial discipline and the right foundations.
Almost every organisation I speak to is exploring how AI can create business value. But behind every new initiative is an IT team that still has to keep everything else running.
That tension isn’t new. What’s changed is the pace at which expectations are growing, the pressure to adopt new technologies and the mixed results organisations are seeing as they try to turn innovation into measurable outcomes.
The Honest Numbers
MIT’s Project NANDA reported last year that 95% of the organisations in its research were getting no measurable return from generative AI. It’s worth noting that this was a preliminary study with a limited sample, but the findings raise an important question about how effectively organisations are turning AI investment into business value.
Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value or inadequate risk controls.
The challenges don’t end once AI reaches production. In Cloudera’s 2026 survey, 98% of IT leaders said they had pulled back an AI deployment at least once because of failures, hallucinations or unexpected behaviour.
And while organisations are still working out how to adopt AI formally, employees may already be using it. MIT found that only 40% of companies had purchased an official AI subscription, yet workers at more than 90% of companies were using personal AI tools for work.
That presents two challenges at once: the opportunity to innovate, and the responsibility to govern how that innovation happens.
What Successful Projects have in Common
The good news is that successful AI projects tend to share a few important characteristics. MIT’s research points to projects built around a specific workflow, with measurable outcomes and a clearly accountable business owner.
Two customers come to mind, and they made very different choices.
A logistics group we work with focused on one process: quoting. In their market, responding quickly can make the difference between winning and losing business. Their tool reads incoming requests, including complex PDF and Excel attachments, and drafts the quote. It’s designed to reduce a task that previously took 20 to 30 minutes to around 30 seconds.
They chose to keep the solution in-house to maintain control over their data. It runs on their own server at a site with its own solar capacity, avoiding the need to pay for colocation.
A large infrastructure group took a different approach to hardware. Rather than owning servers and carrying the associated data centre costs, they rent GPU capacity from local providers and pay for what they use. They also wanted to avoid becoming dependent on a single AI model vendor. Before scaling up, they dedicated the year to putting the necessary governance and controls in place.
Neither approach is inherently better. The right decision depends on the organisation’s infrastructure, operational requirements, data considerations and commercial objectives. What matters is understanding those factors before committing to a solution.
MIT also found that AI tools acquired through external partners succeeded about two-thirds of the time, compared with roughly one-third for internally built systems. That doesn’t mean organisations should never build their own solutions. It means being realistic about the expertise, resources and ongoing support required to make them work.
Keeping the Lights on while you Experiment
Here’s how I’d approach the balance:
Ring-fence the experiments. Keep AI pilots in a separate environment with their own budget. This limits the risk of an unsuccessful test affecting production systems and makes it easier to track what each experiment actually costs.
Make the basics boring. Backups you’ve actually tested, identity and access under control, and a security stack you trust. In my experience, the teams that innovate fastest are the ones whose foundations are dull and dependable.
Know your data. Understand where it lives, who can access it and what your obligations under POPIA are before introducing it to any AI model. Data governance shouldn’t be an afterthought.
Watch the running costs. A pilot might look affordable, but costs can change significantly as usage grows. Factor in token and inference spend, licensing, infrastructure and ongoing support before deciding whether a solution is commercially sustainable.
Size the infrastructure properly. Whether AI runs in the cloud, on premises or across both, it needs the right compute, storage and power capacity. GPU availability and procurement lead times should also form part of the planning, particularly where dedicated infrastructure is required.
Where to Next?
Every AI initiative starts with a question. The important part is knowing what to explore, what to consider and what it will take to make it work.
If there’s an AI idea sitting in your backlog, let’s Xplore it. Bring us the use case and we’ll help you assess the infrastructure, understand the commercial implications and identify the risks before anyone signs a purchase order.
Because innovation should move your business forward, not compromise the technology keeping it running.