The question any organisation has about AI agents is the same question it asks about any technology investment: what is the ROI and what is the benchmark for success?
AI agents, software systems that autonomously perceive, decide and act to achieve defined goals, have moved from pilots into live enterprise operations faster than most technology waves in recent memory. However, there is now data that tells a more complicated story than the hype suggests.
KTSL’s 2026 State of AI Agents in the UK research found that 88% of UK enterprises are actively deploying AI agents. However, only 20% have reached measurable business impact. That gap between deployment and proven return is where most organisations are sitting right now. Closing it depends less on the technology itself than on understanding what the end goal is.
What did the report tell us about Pharma specifically?
The most surprising finding from the research is that cost savings ranked last among the expected benefits of AI agents, with only 29% of respondents citing it.
Instead, the top priorities were:
-
Faster incident resolution
-
Improved customer and user experience
Both were cited by 44% of organisations overall.
What these figures show is that the dominant narrative around AI, automation as cost reduction, is framing the issue incorrectly. Pharma enterprises have moved from asking “what can we cut?” to “what can we improve?”, and speed and responsiveness are the answer.
This changes how AI agent deployments should be scoped, how success should be measured, and how a business case should be constructed.
Compared with retail and finance, pharma identified faster resolution of incidents (50%) as the key metric for adopting AI within IT Service and Operations Management.
Why is Pharma pushing for faster resolution and how can it be measured?
The challenge for pharma is ensuring that speed of development is not compromised by the underlying IT infrastructure. It is a key component of successful development and deployment.
Capturing clinical trial data is often carried out via an app, while analysis depends on the speed and quality of IT. If IT issues prevent that speed, the business cannot move forward efficiently and it ultimately affects the bottom line.
The most critical metric in any IT organisation is Mean Time To Resolution (MTTR): how quickly staff and colleagues can get what they need to do their jobs. While incidents are often the focus, requests are equally important because both contribute to the total delivery of the service.
MTTR can be broken down into key components:
-
Identify
-
Triage
-
Isolate
-
Diagnose
-
Fix
-
Verify
AI agents can play a role in some or all of these stages, with agentic capability moving across the lifecycle of an incident.
For example, an agent could capture a log at the point an issue occurs, perform a level of triage and identify the error. It could then locate an article pointing to a script to resolve the problem, run the script and rerun a test to confirm the fix. This allows AI to automate the whole MTTR lifecycle.
It is important to understand where delays exist within the current operation.
Pharma environments are typically complex, spanning cloud and on-premise systems, which makes issues harder to identify and isolate quickly.
Before and during any AI agent deployment, organisations must understand what the current metrics are in order to measure the result.
For example, if the current MTTR is 18 minutes per incident and the time taken to identify the issue accounts for the first 10 minutes, this would indicate a strong case for leveraging AI to identify the exact problem. A metric for success would be reducing this by 50%.
Where AI agent deployments are succeeding, and where they are not
The research found that 71% of enterprise AI agent deployments are meeting or exceeding ROI expectations, representing a positive and meaningful majority. However, the one in four that fail to meet expectations is also significant, especially given that 26% of enterprises say they plan to pause further deployments over the next 24 months as a result.
The reasons businesses are falling short are revealing:
-
Skills gaps
-
Insufficient business cases
-
Data quality issues
-
Lack of a suitable technology partner
None of these are technology problems. They are deployment and strategy problems, which means they are solvable with the right approach.
In terms of the one-in-four failure rate, Pharma sits at 29%, the highest in the study. The highly regulated and complex nature of the sector almost certainly contributes to this.
Pharma is also unique in that AI skills within the business are often highly developed, but these skills typically sit within development teams while operational IT struggles to keep pace. This creates unrealistic expectations and increases the risk of disappointment within the sector.
The human factor and perception of value
It is important to consider the human interaction as well, both in terms of perceived value for the customer and whether agents are seen to enhance or improve the existing support model.
Agents need to be positioned as improvements rather than alternatives, and they need to be faster than the current methods on offer or they will never be perceived as a success.
They need to remove effort rather than simply shift it elsewhere, and engineers themselves need to feel the benefits. These should be key considerations when building the business case and will help close the expectation gap.
The governance gap
Despite rapid adoption, only 27% of enterprises have a formal and comprehensive security policy covering their AI agent deployments, while 88% are running autonomous systems with access to sensitive data and the ability to make high-impact decisions with minimal human oversight.
This is often a structural problem. Legacy security architectures were designed for human-operated systems, not AI agents making autonomous decisions across distributed environments.
The lesson from previous technology waves, SaaS sprawl, shadow IT and ungoverned cloud adoption, is that governance retrofitted after the fact is harder, more expensive and more disruptive than governance built in from the start.
The complexity of integration
Most enterprises recognise that AI must work with what they already have. The research found that 73% identified integration with existing systems as a top priority, a hard-won lesson from previous transformation cycles.
Only 35% rely solely on public cloud, while 48% favour private hosting and 17% operate hybrid environments.
Each hosting decision adds integration complexity and, in combination, creates environments requiring careful architectural thinking that places additional pressure on IT teams already managing legacy transitions.
This is often seen as a blocker to AI adoption. However, AI also has a role to play in identifying key legacy issues and highlighting where integration will add the most value.
Insight is key to making the right integration decisions, and AI can streamline this process as long as it is positioned earlier in the value chain.
Build correctly rather than deploy hastily
For organisations on the wrong side of the ROI figure, the question should not be how to deploy more AI, but how to close the gap between deployment and demonstrable return.
That means getting the foundations right first: well-defined use cases, clean and well-governed data, proper integration, and security policies that reflect the reality of autonomous systems operating at speed.
Slow down, build them correctly, and the returns will follow.