BlogResearch notes

Four Ideas From the ETL Journal Fall 2025 for Agentic AI in the Enterprise

Four ideas from IT Revolution's Enterprise Technology Leadership Journal Fall 2025 that matter for enterprise agentic AI, and how we at XALT read them.

Philipp Göllner6 October 20263 min read

IT Revolution's Enterprise Technology Leadership Journal, Fall 2025 (Volume 2, Issue 2) collects ten papers from technology leaders on how AI changes software delivery and leadership. The journal is a free download.

Most of the papers are about AI-assisted development. We read them with a different question in mind: what do they mean for enterprises that let AI agents act in their systems? Here are the four ideas we found most useful, with our view on each.

Verification becomes the bottleneck

The paper "Revenge of QA" by Elisabeth Hendrickson, Jeffrey Fredrick, Joseph Enochs, John Willis and Kamran Kazempour argues that when AI produces work faster than people can review it, checking that work becomes the constraint. Producing more, faster, does not help if review and validation cannot keep up. The authors call for building quality into the whole process instead of inspecting at the end.

Our view: for agents this is even sharper than for code. An agent does not only produce text, it changes tickets, sends messages and updates records. Every action needs a check that fits its risk. We see two practical consequences. First, approval should be part of the action flow, not a review step afterwards. Second, every action must be recorded so that a mistake can be traced and undone. That is why Northstar requires approval for every agent action and keeps a full activity log.

Someone is on call for what the AI did

"No Vibe Coding When I'm On Call!" uses a story about an incident in the middle of the night to show the operational side of AI-assisted work. The question it raises is simple: who owns the outcome when AI-generated work breaks in production?

Our view: the same question applies to every agent. Before an agent goes live, it needs a named human owner, clear limits on what it may do, and a way to see what it did. Without these, the first incident turns into a search for who is responsible. With them, it is a normal incident.

Adoption is a governance, measurement and money question

"Lessons from Enterprise GenAI Adoption Journeys" sums up experience from large-scale deployments. IT Revolution's summary names six domains: governance, strategic tooling, measurement, financial planning, workforce enablement and organisational change.

Our view: this list matches what we see in practice. Pilots rarely fail on the model. They fail because nobody defined who may approve what, nobody measured a baseline, and nobody owned the budget. We treat these as part of the build, not as tasks for later. Our posts on measuring AI impact and choosing models per task go deeper on two of them.

Human factors decide in critical systems

"Leading the Human-AI Revolution" looks at human-AI interaction in safety-critical systems and at the psychological safety teams need to adopt AI well. Where errors are costly, the design of the handover between human and machine matters as much as the technology.

Our view: most enterprise processes are not safety-critical, but the principle carries over. People accept an agent when they can see what it proposes, change it and say no. An agent that acts without that visibility creates resistance, even when it is right. Design the human step first, then automate around it.

What we take from it

The common thread across the journal: the hard part of AI in large organisations is not producing output. It is checking it, owning it and governing it. For agentic AI, we draw one conclusion: control and traceability are the product, not an add-on.

Sources

Related: Regulated AI – Operational AI for the regulated enterprise.

For enterprise leaders

Apply these ideas to your first agents

A conversation for leaders who want agents to act in their systems: verification, ownership and governance for your first agents.