Are We Ready to Work With Agents?

AI agents are quickly becoming one of the most important shifts in enterprise software. Unlike a chatbot that waits for a question and returns an answer, an agent is designed to pursue a goal: interpret intent, gather context, use tools, complete steps, check progress, and hand work back to a human when judgment or approval is needed.

For scholarly publishing and education, this shift is significant. These industries are built on complex knowledge workflows: manuscript submission, peer review, research integrity checks, metadata enrichment, content production, learning design, assessment, compliance, and reporting. Many of these workflows are not a single task. They are sequences of decisions, handoffs, exceptions, policies, and quality checks.

That is exactly where agents may create value. But it is also where many agent projects may fail.

The risk is that organizations treat agents as another software feature: add an assistant, connect a few tools, and expect productivity to improve. In reality, successful AI agent applications require more than model capability. They require the right organizational design, platform architecture, workflow boundaries, data discipline, evaluation methods, and human accountability.

I see five elements becoming critical.

The first is a clear task boundary. Agents work best when the goal is specific enough to measure and constrained enough to manage. “Help with peer review” is too broad. “Summarize reviewer availability, draft invitation options, and escalate policy conflicts” is more useful. In education, “support learning” is too vague. “Recommend revision activities based on assessment gaps and approved learning objectives” is a better agentic task.

The second is a trusted action space. Agents should not be given unlimited freedom across enterprise systems. They need permissioned tools, defined data access, approved actions, and clear stopping points. In publishing, an agent may suggest a reviewer, flag a missing ethics statement, or draft author correspondence. But final editorial judgment, policy decisions, and irreversible actions still need controlled approval.

The third is context quality. Agents are only as useful as the context they can access and interpret. For publishers and education providers, that means policies, metadata, rights, content structure, submission history, learning objectives, assessment rules, accessibility requirements, and customer-specific configurations must be organized in ways machines can access and use. Poor context creates poor actions, even when the model is strong.

The fourth is evaluation. Agent quality cannot be measured only by whether the output looks fluent. Organizations need to know whether the agent followed the right steps, used the right sources, respected policy, avoided unsupported claims, escalated correctly, and improved the workflow outcome. This means evaluation becomes part of day-to-day operations, not a final QA step before launch.

The fifth is ownership. Agents cross product, engineering, operations, editorial, learning design, legal, and customer-success boundaries. If nobody owns the full workflow outcome, the agent becomes another disconnected tool. Successful organizations will define who owns agent behavior, who approves changes, who reviews incidents, and who decides when automation should stop.

This shift has major implications for the SaaS platforms widely used across the industry. For many years, SaaS platforms have been built around a scalable, standardized, and largely deterministic workflow model. They are organized through screens, modules, permissions, forms, rules, and workflow states. Users log in, navigate to a function, complete a task, and then move to the next step. This model has worked well because it brings consistency, repeatability, and operational scale.

Agentic software changes that pattern. Instead of asking users to follow a fixed click path, agentic systems start from user intent. They interpret the goal, gather context, use approved tools, complete intermediate steps, and return work for review, correction, or approval. As a result, the platform becomes less of a place users visit for every individual action and more of an orchestration environment where work is initiated, monitored, checked, and completed across systems.

This does not mean SaaS platforms disappear. It means their interaction model and value proposition change. Future platforms will combine deterministic workflow control with probabilistic AI capability. The deterministic layer will still manage identity, permissions, workflow state, policy rules, compliance, audit trails, and final approvals. The probabilistic layer will help interpret intent, retrieve context, generate drafts, recommend next actions, detect exceptions, and support decision-making.

In other words, the platform becomes the operating environment for agents. Its value will no longer come only from standardizing workflows at scale, but from safely coordinating people, AI agents, data, policies, and tools around trusted outcomes. For scholarly publishing and education, this shift is especially important because the work depends not only on efficiency, but also on quality, integrity, evidence, and human accountability.

This could reshape products around manuscript readiness, integrity triage, reviewer operations, production quality, metadata completeness, and author support. For education, it could reshape products around adaptive content transformation, learner support, assessment workflows, CPD evidence, and skills-based progression.

The commercial model may also change. Customers will not only ask whether a product has AI. They will ask whether it reduces editorial burden, improves quality, shortens cycle time, increases compliance, supports learning outcomes, and provides evidence when something goes wrong. SaaS value will move from feature access toward trusted workflow performance and from usage driven to value driven.

This is why agent strategy should not begin with the question, “Where can we add an agent?” A better question is: “Which workflow has enough structure, data, repetition, human oversight, and measurable value to be safely improved by an agent?”

AI agents may become powerful colleagues in research, publishing, and education. But they will not succeed simply because they can act. They will succeed when organizations design the work around them carefully.

The next competitive advantage may belong to those who do not just deploy agents, but know how to govern, evaluate, and scale them inside real knowledge workflows.

Dr. Hong Zhou is Vice President, Product Management at KGL, where he leads product strategy and AI innovation for solutions serving researchers, librarians, and publishers. He can be reached at info@kwglobal.com.