For the better part of a decade, the enterprise IT world ran on a small set of trusted platforms. ServiceNow handled your incident queue. Jira Service Management tracked your sprints and escalations. ManageEngine or ServiceDesk+ kept the help desk alive. These were not bad tools. They brought order to chaotic, high-volume operations, and for structured workflows inside clean environments, they still hold up.

But something has shifted: Unstructured parts of the enterprise workflow have entered the crosshairs of automation firms. Emails that don’t match a template or vendor portals with no API. Rules-based automation was never capable of handling unstructured, legacy junctures in a workflow. Large language models can, and they’ve changed automation as a result.

What RPA Platforms Were Built for

Traditional robotic process automation works by mimicking human actions at the interface level. A bot follows a predefined script. If the interface stays exactly the same and the inputs stay exactly the same, it works great.

But since enterprise environments do not stay the same. Unstructured inputs require human judgment before the bot can proceed. According to research from EY and Deloitte, somewhere between 30 and 50 percent of RPA projects fail outright, and only three percent of companies successfully scale their initiatives. That is not marginal. It is evident that there is a structural problem built into the architecture of the technology itself.

The deeper limitation is data. Studies consistently estimate that between 80 and 90 percent of enterprise data is unstructured: emails, chat logs, scanned documents, support transcripts, contracts. Rule-based bots cannot read those things. But now, with LLMs, we can.

How Modern AI & LLMs Differ

A large language model does not follow a script. It interprets context. Given an invoice that does not match a purchase order, an LLM-powered agent can flag the discrepancy, draft a supplier message, and route the exception through the appropriate approval chain. It can do that across different layouts, formats, and phrasing in the source data, because it is reasoning about the content rather than pattern-matching against a fixed template.

The research is still early, but a comparative study published through arXiv in 2025 found that while traditional RPA outperforms LLM agents on execution speed in stable, repetitive environments, LLM agents required significantly less development time and adapted far more flexibly to dynamic interfaces. That development time gap matters enormously at the enterprise scale, where building and maintaining RPA scripts is an ongoing cost center.

This is the core distinction: RPA automates a known sequence of steps. LLM agents automate against a goal.

How the Incumbents are Responding

None of the major ITSM and workflow platforms missed this. They are all moving (with varying degrees of urgency) towards adoption.

ServiceNow has been the most aggressive. Its Now Assist suite layers generative AI capabilities directly into the Now Platform, handling case summarization, flow generation, and virtual agent interactions. Their chief product officer has described the company’s strategy as building task-specific small language models rather than competing with foundation model providers, arguing that focused models serve enterprise use cases faster and at lower cost. By 2026, ServiceNow is positioning agentic AI not as a feature but as the primary layer through which work is routed and resolved. The platform’s advantage is its unified data model, which means AI agents can draw on context from across the enterprise without requiring a data migration.

Atlassian has taken a similar path with Rovo, its AI layer for Jira, Confluence, and Jira Service Management. Rovo agents can triage incidents, clear backlog, and take actions across connected tools like Slack, Google Docs, and Workday. Atlassian has also built out hosted LLM infrastructure within its own cloud boundary, which addresses the data residency and compliance concerns that have made enterprise AI adoption slow in regulated industries. Rovo Chat uses a hybrid approach, drawing on GPT, Claude, Gemini, and Mistral depending on the task.

ManageEngine and its ServiceDesk+ product have moved more conservatively, integrating AI-assisted features at the ticket and workflow level, but the direction is consistent across the category: every major ITSM vendor now treats LLM integration as a survival requirement, not a differentiator.

The platforms are adapting. But adaptation from inside a large existing product is different from building for intelligence from the start. There is a meaningful gap between bolting generative AI onto a workflow engine and designing an automation system around a model that can reason.

The Governance Gap

The platforms are also dealing with a problem that does not show up clearly in product demos. As ServiceNow’s EMEA president Cathy Mauzaize put it, 2026 is about embedding AI into the operating fabric of business, which makes governance “not a brake, but an engine of trust.” That is a smart reframe, but it reveals the underlying anxiety.

Agentic systems make decisions. Traditionally, if a bot fills a field incorrectly, you roll it back. But when an LLM agent misroutes a high-priority incident or drafts an external communication with an error in it, the exposure is very different. The organizations that are scaling AI automation successfully are those treating observability, auditability, and human-in-the-loop design as first-order requirements.

For IT service management specifically, this matters a great deal. When AI is integrated deeply into information and knowledge workflows, the same reasoning capabilities that make LLMs powerful can introduce inconsistency if the system is not well-governed. The model needs context boundaries. It needs clear escalation paths. And it needs someone who understands how the pieces connect.

Small, Specialized Firms Have an Unusual Advantage

This is the part that tends to get lost in coverage of the big platform wars.

The enterprise software vendors are building horizontally. Their products have to work for a hospital, a logistics company, and a bank, all at once. That means their implementations are general by design. The AI capabilities in Now Assist or Rovo are impressive, but they ship as defaults that require significant configuration to deliver real operational impact.

Small, specialized AI services firms work differently. They come in with a specific operational problem, map the unstructured processes, identify what the LLM layer can absorb, and build something that fits the environment. They are not locked into one platform’s data model or one vendor’s model preferences. They can combine tools. They can build against open APIs. And they can move considerably faster than a traditional enterprise software engagement, often in weeks rather than months.

The cost argument matters too. Hiring a team of technical staff who know how to configure ServiceNow’s AI layer, manage model integrations, and build governance frameworks around agentic workflows is expensive. Enterprises that are not already deep in the ServiceNow ecosystem often find that the licensing costs and implementation overhead make the ROI case hard to close. A focused external engagement, scoped to a specific workflow or department, frequently delivers a cleaner return at a fraction of the cost.

The Transition

The question enterprises should be asking is not whether to move toward LLM-augmented automation. That direction is settled. The useful question is which workflows to prioritize, and how to introduce intelligence without destabilizing the operational systems that already work.

A few things tend to matter immediately. First, the target processes should have clear, measurable outcomes so that results can be validated against baseline performance. Second, the automation should not begin in high-stakes, low-tolerance environments. Third, hybrid approaches, where LLM agents handle reasoning and exception management while existing RPA or ITSM tools handle structured execution, reduce implementation risk considerably.

The platforms are catching up. But the window where a well-scoped, intelligently implemented AI automation engagement delivers outsized value ahead of a full platform migration is real, and it is open right now.