Insights

Are AI coding tools killing low-code, or are they transforming it?

Cass Bishop, Director, AI and Automation

AI coding tools and agents have made it much easier to turn descriptions into working software. For many internal tools and straightforward applications, people can now generate usable results faster than configuring a visual builder. This has led some to conclude that low-code platforms are becoming obsolete.

Dramatic claims of this kind attract attention. Hyperbole works well for clicks in AI commentary. It is far less useful when organizations must plan systems that need to stay secure, auditable, and maintainable for years. The practical question is narrower: where pure generation creates problems, and which approaches for adding structure actually work at enterprise scale.

Generating code or agent behavior is one problem. Running reliable systems across real environments is another. Unconstrained generation often produces inconsistent security, missing audit trails, unclear ownership, and logic that is difficult to change later. These problems are real. They do not disappear simply because an AI wrote the code.

DIVIDER

The limits of pure generation

AI tools handle many tasks well: scaffolding applications, connecting systems, producing forms, writing boilerplate, and iterating quickly. For individuals or small teams building simpler systems, this is often sufficient and cheaper than traditional low-code.

Enterprise work frequently involves more. Processes cross multiple systems of record. They handle regulated data. They run for days or weeks with exceptions, handoffs, and decisions that require human judgment. They need consistent identity controls, logging, and the ability to explain what happened after the fact. Code generated without these controls tends to accumulate problems that grow with scale and time.

These issues are not unique to AI. Traditional low-code platforms have long produced their own versions of shadow IT, undocumented logic, permission sprawl, and visual configurations that are difficult to inspect or refactor. Adding agents does not solve those problems by default; weak governance can make them worse.

DIVIDER

How major platforms are responding

Several established vendors are folding AI agents into their platforms. Their approaches differ in emphasis, but the pattern is similar: keep the existing control, data, and audit layers while letting agents perform more of the work.

UiPath has expanded from robotic process automation into broader orchestration. Its Maestro Case capability targets long-running, exception-heavy work such as claims, investigations, and complex onboarding. Cases are treated as persistent records that carry data, participants, and state. AI agents triage work, extract information, draft responses, and suggest next steps. Robots handle repeatable actions. People handle decisions requiring judgment. Reported results from early users show substantial reductions in handling time, though the figures come primarily from the company and early adopters and independent validation remains limited.

Microsoft has made parallel moves across Power Apps, Power Automate, Copilot Studio, and Azure Logic Apps. Agents can operate inside applications. Application capabilities such as form filling and data views are exposed through the Model Context Protocol (MCP) so agents can call them as structured tools rather than inventing their own access methods. Logic Apps supports agent loops inside workflows and can invoke agents built elsewhere. Governance rules, connectors, identity, and logging remain part of the platform.

ServiceNow is taking a comparable path with heavier emphasis on its existing case and workflow strengths. It distinguishes advisory capabilities (Now Assist) from AI Agents that can triage, assign, remediate low-risk issues, update records, and close work within configured guardrails. AI Agent Studio lets teams build and coordinate multi-agent workflows. The platform acts as a system of action: agents operate against ServiceNow’s data model, access controls, and audit trail. ServiceNow has also exposed an MCP server and Action Fabric so external agents can discover and call governed actions. Its Build Agent works inside common coding tools while carrying ServiceNow context and platform governance. The company reports high autonomous resolution rates on internal and customer service workloads, again largely from its own operations and early deployments.

These three vendors are adapting rather than standing still. All of them benefit commercially if customers keep agents inside their controlled environments. The results so far are early and vendor-influenced; they are useful signals, not conclusive proof.

DIVIDER

Broader options and real trade-offs

Governance and structure matter. They do not require these specific platforms.

Many organizations already deliver shared identity, permissions, audit logs, observability, and human escalation using conventional practices: version control, CI/CD pipelines, policy-as-code, identity providers, event logging, and code review. AI coding tools are adding structured tool interfaces (including MCP-style approaches), constrained execution environments, and better logging. Standalone agent frameworks are building their own policy and audit layers. The difficulty is that multi-agent systems introduce new operational problems—prompt and version drift, inconsistent observability of agent decisions, permission sprawl across tools, and the need to evaluate behavior over time—that many teams are still learning to manage.

There is a spectrum of approaches:

• AI agents generating or modifying conventional code inside disciplined engineering processes.
• Agent systems with explicit policy, identity, and logging layers independent of any low-code vendor.
• Selective use of low-code or RPA only where pre-built connectors and visual models clearly reduce cost and risk.
• Established case and workflow platforms (ServiceNow, Pega, and others) that already handle dynamic work and are adding agent capabilities.

Trade-offs cut both ways. Low-code and automation platforms often carry high licensing costs and create switching friction. Visual or configuration-based logic is frequently harder to static-analyze and refactor than ordinary code. Proprietary runtimes limit portability. Unconstrained generation produces its own maintainability and security debt, sometimes accumulating faster and in less inspectable forms. Neither side automatically wins on technical debt. The more useful question is which form of debt is easier to detect, measure, and pay down for a given team and workload.

Market data still shows continued growth in low-code usage for new applications even as AI coding agents pressure the category and force vendors to add agent features. The category is being reshaped, not collapsing.

DIVIDER

A clearer position

AI coding tools are reducing the number of situations where classic visual low-code is the fastest or cheapest option. That pressure is permanent. Pure generation without structure creates real operational and compliance problems in complex environments.

The practical distinction is by type of work. For simple internal tools, scaffolding, and low-risk automation, lightly governed AI generation is often winning on speed and cost. For long-running, regulated, multi-system processes that require consistent escalation paths and auditability, platforms that already supply data models, permissions, logging, and human handoff points remain useful—whether those platforms are traditional low-code/automation systems or more conventional engineering stacks with added agent controls.

The platforms that last will be those that supply useful structure while incorporating AI agents. Those platforms will include some current low-code and automation vendors, but they will not be limited to them. Conventional engineering practices, structured protocols such as MCP, and independent agent frameworks can supply the same controls.

Use AI generation where speed matters and risk is manageable. Use platforms with stronger built-in governance where processes are long-running, regulated, or cross many systems. Treat vendor claims and early results as inputs, not decisions. Structure remains necessary. The useful question is which combination of tools delivers it at acceptable cost, risk, and long-term operability for the specific work being done.

Turn AI experiments into enterprise automation

Generating code is easy. Scaling automation across business-critical processes is harder. Discover how UST helps organizations combine AI, intelligent automation, governance, and orchestration to deliver secure, compliant, and measurable outcomes.

Explore UST Automation Services:
Accelerate Enterprise Automation with UST