AI can help agencies complete work faster, but speed alone doesn’t create a lasting competitive advantage. The bigger opportunity lies in preserving client knowledge, learning from previous work, and making that experience available whenever it’s needed.
In most digital agencies, some of the most valuable operational knowledge isn’t documented in a system or shared across the team. It lives in the experience of the people who work with clients and their websites every day. It might be the senior developer who remembers why a particular WooCommerce plugin needs a custom configuration, the project manager who understands which changes a client has repeatedly rejected, or the technical lead who can recognize a recurring performance issue before anyone else understands what’s happening.
That knowledge is enormously valuable. It can turn hours of troubleshooting into a twenty-minute fix, prevent expensive mistakes, and make the difference between a smooth client experience and an unnecessary escalation. The challenge is that much of this expertise remains difficult to access, transfer, or reuse, particularly when the people who hold it aren’t available.
For years, agencies have managed to work around this. Experienced colleagues answer questions, new employees gradually learn how clients operate, and recurring problems are resolved through a combination of documentation and institutional memory.
As AI agents become more involved in development and website operations, however, the limitations of that approach become harder to ignore. An AI agent may be able to complete a task much faster than a person, but if it doesn’t know what happened the last time that task was performed, it can end up repeating the same investigation, overlooking the same constraints, or proposing solutions that have already been ruled out.
The real opportunity for agencies isn’t simply to automate more work. It’s to make sure that every problem solved contributes to the knowledge available for the next one.
The biggest productivity problem isn’t always the work itself
The scale of the knowledge problem becomes clearer when we look at how developers actually work. Stack Overflow’s 2026 Developer Survey, published on October 6, collected responses from 30,903 developers and technologists across 169 countries. Among respondents to its questions about workplace knowledge, 61% identified important context being held in people’s heads as a barrier to finding the information they needed. Another 79% said they discovered important context only after starting or completing a task, at least some of the time. When developers needed answers about their work, 72% turned to colleagues, compared with 63% who consulted code repositories and 60% who used internal documentation.
These figures illustrate a familiar reality: even in highly technical organizations, the most useful knowledge isn’t always where people expect to find it. The problem goes beyond the time spent answering questions.
An engineer who needs to interrupt a colleague for information loses time, but so does the person providing the answer. When the same question is asked repeatedly, that interruption becomes part of the organization’s operating costs.
Atlassian’s 2025 State of Developer Experience report, based on a survey of 3,500 developers and managers across six countries, adds another perspective. Half of the developers surveyed reported losing at least ten hours per week to non-coding work, with finding information, adapting to new technology, and switching between tools among the leading sources of friction. Atlassian has also highlighted that developers spend approximately 16% of their time writing code, which helps explain why making code generation faster doesn’t necessarily address the largest obstacles to productivity.

These are findings from different surveys rather than directly comparable measures, but they point to the same operational challenge. Agencies don’t necessarily lack the ability to complete work efficiently. They often lack a consistent way to make the information required for that work available at the right moment. AI can help retrieve, organize, and apply information, but it cannot reliably use context that has never been captured.
What happens when an AI agent starts without context?
This becomes particularly important as agencies move from using AI assistants to introducing AI agents capable of performing multi-step tasks. A coding assistant can help a developer produce a solution. An agent investigating a website issue may need to examine several systems, review previous actions, identify potential causes, and decide what to do next.
The more complex the workflow, the more important context becomes. An April 2026 study by researchers from Stanford’s Digital Economy Lab and Microsoft Research found that the cost of running agentic coding tasks was driven primarily by input tokens rather than generated output. Agents repeatedly process their instructions and accumulated history as they work. The researchers also found that different runs of the same task could vary by as much as 30 times in total token consumption, without higher consumption necessarily producing better results.
For agencies, this introduces an important distinction between having access to information and managing that information effectively. An AI agent may have the technical capability to investigate a problem, but without a reliable understanding of the client’s systems, previous decisions, and operational constraints, it must spend time rediscovering information that the agency already knows.
That repeated investigation costs resources, whether the work is being performed by a person or an AI system. And it creates a bigger problem when the missing context affects the quality of the outcome. An agent that doesn’t know why a previous configuration failed may recommend it again. One that lacks a client’s history may propose a technically reasonable change that conflicts with an established business requirement.
The difficulty isn’t necessarily that the technology lacks intelligence. It’s that the intelligence has insufficient information to work with.
The hidden cost of knowledge that lives in people’s heads
Consider an agency maintaining 100 WordPress and WooCommerce websites with a team of twelve. One of its senior developers knows that a particular client’s checkout can fail after a specific plugin update because of an unusual interaction with the payment gateway. She discovered the problem several years ago, investigated it, and found a reliable workaround. When she’s available, the issue takes twenty minutes to resolve. When she’s on holiday, two colleagues may spend several hours investigating the same symptoms. If she eventually leaves the agency without documenting what she learned, that additional effort becomes a permanent part of supporting the client.
Now imagine similar situations across dozens of websites, each with different plugins, integrations, customizations, and client-specific requirements. The individual inefficiencies may seem manageable, but together they represent substantial operational overhead.
Introducing AI agents can help teams investigate these problems faster, although automation alone doesn’t guarantee that the underlying inefficiency disappears.
An agent still needs access to the information that makes the twenty-minute fix possible. Without it, the agency simply becomes faster at rediscovering things it has already learned.
This is where agency knowledge management moves beyond documentation and becomes an important part of operational efficiency.
The objective isn’t to record everything that happens or create another system employees have to maintain. It’s to capture the decisions, fixes, and constraints that could materially improve future work. For agencies managing complex website portfolios, that information can become one of the most valuable resources they possess.
Why operational memory matters more as AI becomes widely available
There’s a commercial dimension to this that connects directly to the changing agency economics discussed in the previous article. AI capabilities are becoming more accessible, and competing agencies can increasingly use similar models and automation tools. That makes the tools themselves a difficult source of lasting differentiation. What agencies don’t all share is the experience they’ve accumulated working with particular clients, understanding their systems, and solving problems over time.
At Vistara in 2026, Khushbu Doshi, COO at E2M, described an approach she called an agency AI intelligence layer. According to the example referenced in the original Agentic WebOps series research, the same AI agent produced better results when it was given a client-specific context file documenting previous work, successful approaches, unsuccessful attempts, and decisions the client had already made.
The underlying idea is straightforward: an agent becomes more useful when it can build on relevant experience rather than relying only on its general capabilities. For an agency, this means that operational knowledge can become a reusable asset.
A documented solution might help a new employee resolve an issue without involving a senior developer. The same information could support an AI agent investigating a similar problem, provided the context remains accurate and appropriate for the task.
Over time, that can reduce repeated investigations, improve consistency, and make the agency less dependent on individual employees remembering every detail. There are limits, of course. A previous solution still needs to be checked against the current environment, and not every problem can be handled through an existing runbook. But a team that retains useful context starts from a stronger position than one that repeatedly has to reconstruct it.
The competitive advantage isn’t simply that an agency can perform more tasks with AI. It’s that the agency can increasingly draw on what it has learned from the work already completed.
And unlike access to a widely available AI model, that accumulated client knowledge cannot be reproduced instantly by a competitor.
How agencies can build reusable knowledge without creating more work
For many agencies, the challenge isn’t recognizing the value of documentation. It’s finding a practical way to maintain it without introducing another administrative burden.
A useful starting point is to capture important information while the work is still fresh, rather than launching a large project to document years of historical activity. For example, when a developer resolves a recurring performance problem, the ticket could include a short explanation of the cause, the solution, and anything that should be checked before applying the same fix again. That information immediately becomes useful to the next person investigating a similar issue.
For agencies managing multiple clients, a simple client context file can provide another layer of operational knowledge. It might contain the website’s technical environment, important integrations, known limitations, previous incidents, significant client decisions, and changes that have already been considered or rejected. The file doesn’t need to be extensive. A concise, consistently maintained record is often more valuable than a large knowledge base that quickly becomes outdated. As agencies introduce AI agents into their workflows, making this information accessible in structured, machine-readable formats becomes increasingly relevant.
That might mean maintaining runbooks in a repository, organizing client configurations consistently, or ensuring important decisions can be retrieved through approved internal systems. Access controls and data governance remain essential, particularly when the information includes sensitive client details or operational credentials. Making knowledge available to an agent shouldn’t mean giving every system unrestricted access to it.
Finally, agencies can begin measuring whether these practices are producing results. One useful indicator is the time required to resolve a recurring issue compared with the first occurrence. If the second investigation still requires the same effort, it’s worth examining whether the original findings were captured and made accessible.
The purpose isn’t to eliminate human involvement or automate every decision. It’s to reduce unnecessary repetition and allow experienced employees to spend more time on problems that genuinely require their expertise.
From faster execution to lasting agency value
The previous article in this series explored how cheaper AI execution is changing agency pricing and business models. But efficiency alone doesn’t determine which agencies will benefit most. If every agency can access increasingly capable AI tools, the distinction between them becomes less about who can automate a task and more about who can use those capabilities effectively within a client’s particular environment.
That depends on understanding the client, retaining relevant experience, and making reliable decisions based on what has already been learned.
It also changes how we should think about Agentic WebOps.
Agentic WebOps is the governed operation of websites by AI agents: a backup before any change, a human approving anything irreversible, and a system that keeps what it learns.
That final principle is particularly important here. An AI agent that completes a task but retains none of the useful context may improve productivity in the moment. A system that preserves relevant findings and makes them available for future work creates an opportunity for that improvement to accumulate over time.
For agencies, the long-term opportunity is to build services that become more consistent and efficient as their operational knowledge grows. The agencies that benefit most from AI may not necessarily be those adopting the largest number of tools, but those developing the discipline to preserve what they learn and put it to use.
This article is the third in Servebolt’s Agentic WebOps series, exploring how AI agents are changing the economics, technology, and day-to-day operations of the web.
Coming next:
Retaining operational knowledge makes AI agents more useful, but giving them the ability to act introduces another responsibility. In our next article, we’ll explore what happens when autonomous systems make mistakes in Production environments, and why reliable safeguards require more than instructions telling an agent what it should or shouldn’t do.

