- 2. October 2026
- Posted by: Uwe Seebacher
- Category: NEWS
AI agents need more than just good prompts
Generative AI has made its way into many communications departments. However, its practical value often falls short of expectations. The text sounds plausible but is too general. The responses do not take into account internal positioning, brand guidelines, or specific target audiences. The reason often lies not in the language model, but in the data set.
Anyone who wants to use AI agents effectively for corporate communications must therefore start earlier. Not with the prompt, not with the model, but with the question: What knowledge should the AI actually be able to access?
From a Document Repository to a Useful Knowledge Base
Communications departments rarely lack information. The real problem is how it’s distributed. Press releases, key messages, positioning papers, brand guidelines, websites, intranet content, product information, and marketing materials are often scattered across different locations and exist in various versions. It is precisely this fragmentation that prevents AI from reliably working with corporate knowledge.
The first step, therefore, is to identify and consolidate relevant communication data. It is important to note that “the more, the better” does not apply here. What matters is which sources are actually needed for the task at hand.
A file repository only becomes an AI-ready knowledge base when its content is organized in a structured manner. Documents must be made machine-readable, broken down into smaller units of information, and tagged with metadata. Source, URL, date, or version help to locate specific content later and assess its currency. The key insight, therefore, is this: Data must not only be available, but also discoverable, up-to-date, and unambiguously assignable.
The target audience and channel must be integrated into the logic
Corporate communications do not follow a “one-size-fits-all” approach. A LinkedIn post follows different rules than a press release, and an internal newsletter follows different rules than an external website. Similarly, global, regional, and local audiences differ from one another.
These differences should not have to be described anew in the prompt every time. It is recommended to systematically incorporate them into the architecture of the AI assistant. This includes, for example, options for internal or external target audiences as well as regional or national contexts. The selection can then determine which data sources are prioritized.
Similarly, fixed guidelines can be set for individual channels: structure, length, style, hashtags, calls to action, or other formatting rules. When a request comes in, the agent can first determine which channel the content is being produced for, what topic should be covered, whether additional research is needed, and which templates or examples should be considered. This shifts part of the prompt-writing work to the system logic. This increases consistency and reduces dependence on how well individual users craft prompts.
Not all data belongs in the same pot
Another important principle is: Different data sources require different access methods.
Static or relatively stable content, such as guidelines, position papers, or approved communication materials, can be indexed centrally and made discoverable through semantic or vector search. Current or dynamic data, on the other hand, should be retrieved directly from the source system whenever possible.
This applies, for example, to market, reputation, or survey data. Instead of regularly copying this data into a knowledge base, an AI agent can recognize the intent behind a query and retrieve the necessary values in real time via an API. The language model then takes over the task of explaining or summarizing the structured data.
For communications departments, this raises an important architectural question: What information belongs in a central knowledge base—and what information should be retrieved directly from source systems at runtime?
Data Ownership Becomes a Communications Task
Technically sound data access solutions solve only part of the problem. The key factor is who is responsible for the content. Especially in international or decentralized organizations, information is maintained locally. A central communications department cannot keep every country website or every product document up to date on its own. That’s why clear responsibilities are needed. If the underlying data is outdated, even the best AI agent will operate with outdated knowledge.
Data governance should therefore not be viewed solely as an IT or compliance issue. When it comes to communication, at least three questions must be addressed: Who owns the information? Who updates it? And how does the system determine which version is valid?
This makes data ownership a prerequisite for reliable AI communication.
Automate updates whenever possible
Manual data maintenance does not scale well. The more sources an AI assistant uses, the more important automated processes become.
Websites can be crawled regularly, SharePoint repositories can be automatically reindexed, and new or updated files can be transferred to central repositories via workflows. APIs are ideal for situations where data is already structured and up-to-date.
The goal should be to ensure that data remains up-to-date without relying on manual, one-off actions. The more automated data maintenance is, the lower the risk that the agent will work with incorrect or outdated information.
Communications and IT Must Work Together
As soon as AI agents access corporate data, authentication systems, interfaces, and search architectures, the issue becomes technically complex. Communications departments can hardly develop such systems in isolation.
It is therefore advisable to collaborate with the IT department early on. However, this collaboration requires translation in both directions. The IT department must understand which communication sources are relevant, which content is authoritative, and why different target audiences require different contexts. The communications team, in turn, must learn how data structures, interfaces, and technical dependencies work.
This collaboration does not happen automatically. It develops through joint projects, clear use cases, and ongoing communication. In the case study described here, this learning process was an essential part of the development.
Don’t start with the agent
For communications departments, this leads to a simple sequence of steps: First, define the use case; then, determine the required data; next, structure the data sources and clarify responsibilities. Only then should the agent be built.
Taking the reverse approach quickly leads to a powerful language model with insufficient business knowledge. The result is generic responses to specific communication questions.
Similarly, measuring success should not be narrowed down too early to supposedly exact savings figures. It is often difficult to clearly isolate time or cost effects. It makes more sense to start by asking about usage intensity, satisfaction, shorter commutes, common use cases, and perceived improvements in quality.
From Assistant to Agent Architecture
The next step in development goes beyond individual writing or research functions. In the future, specialized subagents will be able to take on various subtasks and collaborate within multi-step workflows.
Possible examples include agents for initial crisis reporting, interview training, research, image search, or channel-specific text production. An overarching orchestration system distributes tasks to specialized components.
However, this also places greater demands on the database. The more autonomously agents operate, the more important it is to have reliable sources, clear lines of responsibility, and transparent decision-making processes.
The most important investment lies ahead of AI
The key conclusion, therefore, is less technological than it might initially seem. Companies do not need more AI capabilities first and foremost, but rather better conditions for AI.
These include curated data, a centralized or at least integrated knowledge architecture, metadata, clear data ownership, automated updates, and defined rules for target audiences and channels.
Work on it does not end once an agent goes live. Data changes, new sources are added, structures evolve, and new use cases emerge. Data management thus becomes an ongoing task for the communications function.
Those who lay this foundation not only improve the output of AI agents; they also create greater transparency, consistency, and order in the communication itself . This is precisely where the strategic value lies: The agent is not the starting point. The starting point is a robust database.
Interview: What AI Agents Can Do in Corporate Communications
When AI Provides Information: Who Offers Guidance?
