What is an AI agent in PR?
An AI agent is an AI that does not only produce text but also performs actions. You give an instruction in plain language, the agent picks the steps that go with it and runs them in the systems it has access to. In PR that means searching a journalist database, assembling a media list, drafting a release, preparing a send and reading the results back later.
The difference with an ordinary AI writing assistant lies in those actions. A writing assistant hands you text you paste somewhere yourself. An agent works inside your system, with your data, and shows you what is ready. Your own work shifts from doing to instructing and checking.
How an agent uses your PR system
An agent can only do what a system offers it. That connection now often runs over the Model Context Protocol, an open standard describing how an AI client calls tools from other systems. The system runs a server with tools, the client connects with a token and uses them.
- Tools. Every action is a separate tool with a clear job: search journalists, save a media list, generate a draft, schedule a send.
- Permissions. The token decides what is allowed: read only, write as well, or send. No permission, no action.
- Context. The agent works on your own data: your clients, your earlier announcements, your results. Not on a general picture of your brand scraped from the web.
Want to see what that looks like in practice? The MCP server page lists every tool and shows an example conversation.
What agents are already good at
Agents are strong at work that takes many steps but little judgement. That is exactly where most PR time disappears.
- Research: filtering journalists by topic, outlet and reach, and turning that selection into a media list.
- Repetition: preparing the same announcement for several brands or clients, each with the right list.
- A first draft based on your input, earlier announcements and the tone of voice you use.
- Overview: which campaigns are scheduled, who opened the last mailing, which client still has no media list.
- Preparing a send, including an upfront check on recipients and blockers.
What stays yours
The line sits at judgement and relationships. An agent does not weigh whether something is news, and a journalist does not build a relationship with a tool.
- Deciding whether your news carries news value. An agent will not turn a product update into a story.
- Checking the substance: figures, quotes, names and promises you cannot keep.
- The contact itself. A pitch that works comes from someone who reads the outlet.
- The decision to send, and the moment you send.
- What you share. Embargoes, sensitive client information and internal figures stay your call.
How to build that contact is covered in Approaching journalists.
Setting it up safely
An agent is only as safe as the permissions it gets. Agree on those before you create a token, not after.
- Start with read access. Let the agent search and report first, write later.
- Keep writing and sending separate. A send is prepared, a human confirms.
- Give every token one purpose and revoke it once it is no longer needed.
- Spot-check what the agent produces, especially media lists and figures in the copy.
- Agree on what never goes into a prompt, such as client information under embargo.
How to start
- Pick one task. Take the work that costs you most time today, for example assembling media lists.
- Connect with read access. Create a read-only token and let the agent shadow that task for a week.
- Compare the result. Put the agent's selection next to your own list. What is missing, what does not belong?
- Extend with write access. If it holds up, let the agent prepare drafts and sends as well.
- Keep the confirmation. Sending stays a human action, even when everything before it is automated.
The MCP server page shows how the connection works; Sending a press release covers when to send.
