Organisations are jumping on the agentic artificial intelligence (AI) bandwagon, but they may be overlooking key components that can result in hidden costs, such as delays.
Enterprises often believe they have abundant data, but this may not necessarily be the case when deeper assessment work is carried out, said Fan Ho, Asia-Pacific executive director and general manager of Lenovo’s solutions and services group, during an online media roundtable held on Monday.
They will realise that the data they have may be fragmented across their organisation, Ho said, in response to FutureCIO’s query on hidden costs that organisations often overlooked in agentic AI deployments.
Access controls for the various business units and disciplines typically are not uniformed, and companies lack a single plane of visibility, she said.
Such issues will impact their ability to execute their agentic AI projects, resulting in delays, she noted.
These companies then realise they need to go back to the drawing board and plug the gaps, she said, adding that this is a common challenge organisations face.
The cost of agentic AI initiatives comprise three broad categories, encompassing model selection, infrastructure, and development and maintenance, said Debdut Maiti, Lenovo’s Greater Asia-Pacific director of solutions and services group.
Careful management across these areas can help companies better manage their rollouts, Maiti said.
In model selection, for instance, he advises organisations to look at their use cases and determine what kind of AI models will best fit each use case. Such decisions can depend on how much latency the workflow can tolerate and how accurate the AI-powered result needs to be.
Establishing these baselines will ensure token consumption is optimised and, hence, cost is more efficiently managed, he said.
Similar cost assessments should be made for infrastructure and development platforms, so organisations can decide which workloads are more cost effective to run on public or private clouds or on-premises, he noted.
Potential to prevent disruption
If deployed thoughtfully, agentic AI can double productivity where it is implemented, Ho said, citing research from IDC.
By 2027, 40% of Global 2000 companies are expected to adopt agentic AI workflows within three years. Generative AI (GenAI), specifically, can generate returns of 3.7-folds on average for every dollar a company invests, she said.
In its 2026 predictions for technology infrastructure and operations, Forrester suggests that agentic AI workflow could prevent a major outage autonomously.
The research firm noted that tech leaders faced growing pressure to improve resilience and uptime amidst an increasingly complex tech environment. This would push more IT teams to turn to AI for help.
Some 43% of AI decision-makers already were using AI in IT operations, it said, while 55% of those working on AI for operations were testing AI for predictive maintenance.
Noting that agentic systems could make decisions or take next best actions, Forrester said: “We predict that an agentic workflow will prevent a system outage through a complex mesh of predictive and causal AI as well as a combination of AI-initiated action and proactive notification with the system and relevant support teams.
“However, agentic systems don’t always make the correct decision," it said. "To maintain trust and see ongoing success, IT teams must institute regular testing of agentic guardrails to mitigate decision-making drift and keep humans in the loop to orchestrate the agentic layer.”
Proper framework needed first
It is critical that organisations first establish the governance framework, including data management policies, that they will need to manage their agentic AI workflows, Ho said.
This should encompass fundamental policies, such as who can access what data, where the data will reside, what kind of platform it should be on, and the different use cases in which agents and employees can have access depending on their role and function.
And they cannot simply let AI run their AI governance framework, she said.
“There needs to be a rulebook or reference guide of some sorts, before you can actually jump into agentic AI,” Ho said.
It is a key component that Lenovo itself is putting together, to provide a step-by-step guide for its customers, she said.
It will help ensure organisations are not jumping ahead too quickly and ignoring basic processes, which may backfire in the later stages of their agentic AI implementation, she added.
In fact, less than 15% of companies will turn on the agentic features in intelligent automation suites due to key challenges, Forrester said, in its 2026 predictions on automation and robotics.
It noted that organisations will struggle to test model-driven decisions within complex workflows. Governance frameworks also will be incomplete, with pilot projects showing limited payback.
“High failure rates reinforce enterprises’ preference for deterministic automation over unpredictable AI agents in critical workflows,” the research firm said. “ROI challenges and insufficient testing capabilities will keep most organisations running traditional rule-based automation through 2026, despite vendor pressure to adopt agentic features.”
Forrester anticipates that organisations instead will adopt hybrid orchestration that encompasses both deterministic workflows and agentic AI. They also will bring in new components, such as skill catalogues to inventory human competencies, workflow maps to identify agent insertion points, and human-agent collaboration frameworks to define partnership levels.
It added that most AI failures stem from training agents on idealised processes that do not match the realities of operating them. This will result in inconsistent behaviours and unreliable outcomes.
“Process intelligence will rescue 30% of failed AI projects, [mapping] how the work actually gets done, not how it’s supposed to happen,” Forrester said.
It said real-world process data and insights will become the foundation for autonomous or semiautonomous operations.
It further recommends that AI project leads mandate process discovery before deploying agentic initiatives and to regard process intelligence as a foundational capability.
According to Forrester, process intelligence applications provide analytics about a company’s processes that are operated by humans and implemented within IT systems, enabling decision-making on how to improve process performance.
