Enterprise teams trust AI more in build than release

Gearset has published research showing that enterprise software teams often review AI-generated code much as they do developer-written work. The findings point to a cautious but widening use of AI across software delivery.

According to the data, 43% of enterprise teams review AI-generated code and configuration changes in exactly the same way as human-written work, while 33% apply greater scrutiny. The figures suggest many teams now treat AI output as a contributor to delivery work, while still requiring close oversight.

Trust in AI varied sharply across the software development lifecycle. At the build stage, 82% of teams said they trust AI involvement, but that share fell to 58% at release, when changes reach live production systems.

This suggests teams are more comfortable using AI where errors can be identified and corrected before they affect customers or business operations. By contrast, release and operational stages prompted greater caution, with 42% of teams excluding AI entirely from release activity.

Trust also remained relatively strong in validation and observability work. Gearset found a clear appetite for autonomous agents in planning and observability, where gains in speed and efficiency were seen as meaningful and the consequences of errors easier to contain.

The gap between build and release reflects a broader distinction in how engineering teams view automation. In lower-risk parts of the process, AI is used as an assistant that can help produce code, support planning or surface problems. In production-facing stages, many teams appear to prefer established automation tools and deterministic continuous integration and continuous delivery pipelines.

Review standards

The findings suggest AI use is moving beyond experimentation in enterprise software teams. Rather than creating separate standards for machine-generated output, many teams appear to be folding that work into existing review practices.

That may matter for managers deciding how to govern teams that include both human developers and AI systems. If AI-generated changes are subject to the same checks as junior developers' work, adoption can expand without a wholesale redesign of review structures.

At the same time, the share of teams applying greater scrutiny to AI-written changes suggests trust is still conditional. AI may now be accepted as part of day-to-day engineering work, but many organisations still want stronger safeguards before accepting those changes into production workflows.

The research sits within a wider software industry debate over where AI adds measurable value. Coding assistants and autonomous agents have drawn heavy interest from developers and investors, but enterprise buyers have generally moved more carefully than consumer users because security, compliance and reliability issues carry higher costs.

That caution is especially visible in sectors that rely on large business platforms and strict release controls. Salesforce environments, which often handle customer data and core sales processes, clearly show how software changes can have immediate operational consequences.

Gearset, which focuses on Salesforce DevOps software, said it supports more than 39 million deployments worldwide and is used by more than 3,000 enterprises. Customers it cites include Deliveroo, Zurich, Cision, McKesson and IBM.

Risk balance

The findings also underline that enterprise AI adoption is not a single decision but a series of smaller choices made at each stage of delivery. Teams appear willing to use AI where it reduces time spent on repetitive work or speeds up checks, but less willing to hand over control when failure would be hardest to reverse.

That distinction could shape how software vendors position AI tools for large organisations. Products aimed at planning, code generation, testing and observability may face fewer barriers than tools that seek direct control over production releases or live operations.

Kevin Boyle, Chief Executive Officer at Gearset, commented on the findings: "What our data shows is that enterprise teams have developed a clear-eyed view of where AI earns its place in the delivery pipeline. They're embracing it enthusiastically for building, planning, validating and observing, exactly where it can save the most time without adding risk. And they're being appropriately cautious at release, where deterministic automation already works and the stakes are highest."