When Johnson & Johnson revealed a shift in its AI strategy last year, it signalled the end of a period of experimentation for the pharmaceutical giant and the beginning of a more disciplined, results-driven era.
At its peak, J&J employees were pursuing nearly 900 individual AI-related use cases. But as the WSJ reported at the time, this was a position that was simply unsustainable. Instead, the company decided to change tack and home in on high-value generative AI use cases in areas such as drug discovery and supply chains.
For many in the tech sector, this was a high-profile example of a business deciding to call time on pilots and trials in favour of a more structured approach where AI finally has to earn its keep.
Fast forward to the beginning of 2026, and the State of AI in the Enterprise report by the Deloitte AI Institute suggests that many companies are also following a similar path, with business leaders becoming impatient and wanting to see a return on their investment.
AI is shifting from a one-off investment to an ongoing operational expense
And it’s easy to see why. Amid the current AI boom, a recent report by Mavvrik, a US-based IT financial management platform, found that 80% of enterprises missed their AI infrastructure cost forecasts by more than 25%. It claimed that AI costs are “crushing margins” and that AI-related overheads are harder to manage than cloud costs because AI introduces new variables where “minor changes in usage can spike spend by 100x”.
It’s a good point. Unlike traditional software, which tends to follow predictable instructions, AI systems explore multiple possible paths to reach a solution. Sometimes that could mean resolving an issue in five minutes. At other times, it might mean AI systems retrying, looping and testing alternatives for hours, leading to additional expense.
But there are other unforeseen costs as well. The dash to embed AI can often expose shortcomings in existing IT infrastructure. Legacy systems and on-prem environments, for example, that might have been perfectly suited to a pre-AI world may now lack the necessary processing capacity required.
What rightsizing AI adoption really looks like
As a result, this may force organisations to bring forward plans to upgrade their technology. But at what cost? It’s not simply a case of writing a cheque and flicking a switch. In practice, it means being clear about the task AI is meant to improve, and defining measurable indicators and a baseline against which to judge impact.
But rolling out AI isn’t just about the technology. One of the lessons we’ve all learned over the last couple of years is that if workflows are unclear or the quality of data is poor, then AI will expose those shortcomings. Without a proper assessment of whether the company is ready for AI, such a project could quickly turn out to be unsuccessful, becoming a time and money sink.
Equally, if AI is embedded into structured environments – with defined responsibilities and feedback loops – then we also know that AI can genuinely improve outcomes.
For instance, a recent report by SolarWinds sought to understand how generative AI (GenAI) has been incorporated into ITSM (IT Service Management) workflows. It focused on those areas within helpdesks designed to reduce manual effort, such as automatically suggesting ticket responses, sourcing relevant knowledge base articles and generating incident summaries.
The report wanted to discover whether AI was genuinely making life easier for IT teams or merely adding another layer of complexity. The findings showed helpdesks saw a significant decrease in average incident resolution time after enabling the GenAI features.
Savings and Benefits
On its own, that saving of just under five hours per ticket – or just under 20% of the time spent per support request – is impressive. But it’s only when you multiply that efficiency gain across the hundreds, if not thousands, of tickets generated each year that the true scale of AI’s impact becomes clear.
While the savings clearly point to GenAI’s benefits, the report also highlighted a common thread around that all-important pivot from pilot to operations. Instead of viewing GenAI as a test or side project, those teams that integrated AI tools into their daily service desk workflows appeared to have better outcomes.
“These results show what’s achievable when AI adoption is combined with effective change management and a focus on process improvements,” said the report. “They also act as a benchmark for other organisations evaluating the potential impact of similar tools,” it said.
Proper safeguards are key
But perhaps the biggest lesson learned from the last couple of years is that if we’re to truly maximise ROI, then we need to look at the bigger picture and develop AI systems in a much more methodical way. In effect, we need to develop AI by design. What does that mean? Well, in terms of privacy and security, it means establishing clear rules around not just the use of data but also how AI tools act and behave.
We also need to look beyond bias checks during model training to ensure that fairness is woven into AI right at the start, while ensuring that humans have the final say. And to build confidence in a new generation of tools, we need a proper paper trail to ensure that AI decisions can be traced and analysed so that, in the event something goes wrong, people can understand not simply what happened but why. It’s a good practice to ask an AI vendor which models are in use, where the training data came from, and what happens to the data users feed into the AI.
It’s too soon to say definitively whether Johnson & Johnson’s pivot signalled the end of AI’s pilot phase and the beginning of a more disciplined era. What we do know, though, is that the organisations most likely to succeed are the ones that are able to successfully manage the transition from AI experimentation to delivery while keeping a keen eye on outcomes and cost.
Learn more at solarwinds.com
- Data & AI
- Digital Strategy