In the first post in this series, I explored Seven Assumptions AI Is Challenging in Libraries and Knowledge Management. We then considered Why Every AI Project Becomes a Learning Project, before turning to The Knowledge AI Can’t Find, and why some of that knowledge may never have been captured in the first place.
Now the series moves into more technical territory.
AI agents, autonomous workflows, and self-learning knowledge systems feature increasingly heavily in professional discussions. The claims are impressive: systems that understand organizational context, anticipate what people need, and act without waiting for instructions at every step.
But how much of this is happening in practice? Drawing on academic research, industry reports, and product documentation, this article examines what these systems do, where assistance becomes action, and how far the evidence supports the claims being made.
What Does It Mean for a Knowledge System to Act?
Ultimately, people want answers and explanations. Search engines have traditionally helped us locate the information from which those answers can be built. Answer engines now attempt to produce them directly, which is a significant shift in the evolution of knowledge retrieval.
Systems can influence what happens next—a move towards what John Delaney calls the “action engine.” Recommender systems suggest relevant content or choices, while automated workflows perform predefined actions, such as sending an alert or routing a request.
Agentic systems go further. They can assess a situation, select from the tools or actions available to them, and decide how to move toward a defined goal.
The important boundary lies between providing information and having permission to act. That might mean updating a record, routing an enquiry, running code, sending a message, or initiating the next stage of a workflow.
If you sense this all sounds familiar, you are right. Libraries have used automation and machine learning within defined workflows for years. The US National Library of Medicine, for example, uses automated indexing to assign medical subject headings to biomedical literature. The system performs a specific task within an established workflow.
The more recent development is the amount of discretion given to the system. Instead of following a single fixed route, an AI agent may interpret the current situation, choose an appropriate step, and examine the result before continuing. This creates exciting new opportunities but involves potential risks.
From AI Assistants to Selective Autonomy
The difference between an AI assistant and a more agentic system lies largely in how much authority it has. An assistant responds to a person’s request and prepares something for them to consider, such as an answer, summary, draft record, or piece of code. The person reviews the result and decides what happens next.
Under selective autonomy, a system may complete certain actions without prior approval, but only under defined conditions. This can move work forward more quickly, while increasing the consequences of an error.
Examples of AI Assistants
Clarivate’s Alma AI Metadata Assistant can analyze information about a library resource and generate draft metadata. A cataloger then reviews and amends the proposed record in the metadata editor.
Kresge Law Library at the University of Notre Dame tested large language models on metadata extraction and processing tasks for a collection of continuing legal education materials. The models performed most successfully where the collection was highly structured. However, human review remained necessary.
A More Agentic Example
In 2026, T-Tech researchers published a peer-reviewed account of an AI system used in customer-support workflows. The system drafts responses and performs certain actions, such as transferring conversations and closing completed chats. High-confidence actions proceed automatically, while uncertain cases return to a human operator.
In tests, it automated 45% of sessions and reduced operator active time by 39%. Although the employee-authored study covers one proprietary environment, it offers a rare example of selective autonomy in practice: limited, measurable authority rather than an independent digital colleague.
What Does the System Need to Know?
The T-Tech example brings us back to the central question running through this series: what counts as organizational knowledge?
Knowledge exists in documents, databases, repositories, and people’s experience. It is also present in how work gets done: the sequence of steps people follow, the corrections they make, the permissions attached to different roles and the circumstances that require escalation.
An agent can only use this operational knowledge if some form of it has been made explicit and accessible. T-Tech’s system drew on structured workflow information, previous interactions, and corrections from human operators. Its available actions were clearly defined.
This structure helps explain why operators accepted suggested clicks, transfers and session closures in approximately 85% to 92% of cases, while accepting only 51% of customer-facing messages. Procedural work offered a limited range of actions and an observable outcome. Appropriate language could depend on tone, ambiguity, emotional context, or information the system had not received.
The case points towards several conditions for selective autonomy:
- a clear goal;
- reliable and accessible information;
- a limited range of permitted actions;
- observable success or failure;
- a route back to a person when exceptions arise.
Missing, outdated, or poorly organized knowledge already weakens search and retrieval. An incomplete document may produce an incomplete answer. An outdated procedure could lead to an incorrect action. Ambiguous permissions may allow the system to go further than intended.
As a system gains authority, knowledge quality becomes an operational concern.
Knowing When to Stop
Much of the discussion around AI agents concentrates on what they may eventually do independently. The evidence suggests that recognizing when to stop is just as important.
A system needs a route for dealing with insufficient information, unexpected situations, and actions that require human authorization. This cannot depend entirely on the model recognizing its own limitations, so boundaries must be built into the workflow.
The appropriate level of control will depend on the consequences. Routing a routine request presents different risks from changing a legal record, approving expenditure, or communicating sensitive advice.
Examples remain limited in scope and number. The strongest ones tend to involve narrow processes, structured information, defined actions, and clear routes back to a person. Whatever the technical arrangement, responsibility for defining and overseeing those boundaries remains human and organizational.
How Much Autonomy Do Knowledge Systems Actually Have?
At the beginning of this article, I asked how much of the excitement around AI agents and autonomous knowledge systems is reflected in current practice. The answer is: less than some claims suggest, but enough to represent a meaningful change.
What emerges from the evidence is a continuum. Libraries have used automation within defined workflows for years. AI assistants can now prepare answers, records, and other work for human review, while selectively autonomous systems may be permitted to complete particular actions under controlled conditions.
The strongest examples remain narrow and structured, with clear goals, limited permissions and routes back to a person. As systems gain greater authority, the quality of the knowledge and processes behind them becomes more consequential. An incomplete answer is one problem; an incorrect action is another.
Libraries and information professionals already have practical experience introducing technology into discovery, cataloging, and service delivery. That experience is directly relevant as they decide where limited autonomy could add value, what knowledge a system would require, and where its authority should end.
These decisions also raise deeper questions about privacy, transparency, accountability, and institutional responsibility. In the next post in this series, I will explore what happens when AI begins to challenge the values knowledge organizations are built to uphold.
0 Comments