According to Bitkom (September 2026), 57 percent of German companies are using AI for the first time – and of those that are already using AI or planning to, 11 percent have AI agents in active use. In agencies, though, AI almost always ends up in the same spot: writing copy, generating images and creating social media posts.
My argument: Most agencies use AI where they sell, rather than where they lose time. The biggest time sink is rarely the creative work itself, but the communication surrounding it – status emails, monthly reports, proposals. That work is recurring, rule-based and tied to data that already lives in your tools. In other words: a perfect fit for an AI agent.
In short: anything that comes around regularly, is built from existing data and can be reviewed by a human at the end. An AI agent is more than a chatbot – it fetches data from your systems on its own, processes it and deposits a result ready for approval. Three tasks pay off in almost every agency first.
Once a week, the agent reads the completed and open tickets from your project tool (Asana, ClickUp, Jira or Trello), plus the latest entries from the project's Slack or Teams channel. From that it writes a status update in your agency's voice: what's done, what's stuck, where a decision is needed from the client. The project manager reviews it, adds a sentence and sends it off – instead of starting from scratch every Friday.
For monthly reports, the agent pulls figures from Google Analytics 4, Search Console or ad accounts and compares them with the previous month. The real value isn't in the chart – any dashboard can do that – but in the two or three sentences of interpretation below it: "Contact form enquiries rose by 18%, most likely driven by the new landing page launched on the 12th." That interpretation is what costs project managers the most time today.
For proposals, the agent works with your past proposals as a template: it takes the brief or meeting notes, looks for similar previous projects, suggests service packages and effort estimates and builds a first draft. For pitches it can research in advance – the client's website, competitors, publicly available company information – and summarise the starting position. Pricing and strategy remain your call, but the blank document is gone.
A worked example, clearly stated as an assumption: suppose an agency in Rosenheim manages 12 active clients. For each project, the project manager writes a weekly status update that takes an average of 20 minutes – that's 4 hours a week. If an agent delivers a finished draft and review takes just 5 minutes per client, the theoretical saving is 3 hours a week, or around 13 hours a month. Add monthly reporting and the effect grows accordingly.
Technically, none of this requires a major overhaul. Two building blocks are usually enough:
If the workflow becomes more complex – for example, if the agent first needs to decide whether a project is behind schedule and then trigger different steps depending on the answer – an orchestration framework like LangGraph is worth considering. For the three tasks above, that's usually overkill at the start.
For an example of what it looks like when CRM, tasks, chat and AI run on a single platform, see our project Agentursysteme. Why a shared data foundation is the prerequisite for any automation is something I explored in Die Zukunft von Agenturen.
An agent is only as good as the data it can find. If tickets aren't maintained and the real decisions happen over the phone, the agent will write a status update that is formally clean but factually wrong. That's more dangerous than no update at all. Before you automate, the process needs to be running in a reasonably disciplined way – AI doesn't fix chaos, it scales it.
Second: anything that goes to the client needs to pass through human hands before it's sent. Especially when the news is bad – delays, budget overruns, falling numbers – tone and timing matter. An agent will get it either too smooth or too blunt. And for proposals, it has no idea what your current capacity looks like, or whether a client feels like the right fit.
Third, data protection is not a minor detail: as soon as client data is passed to a language model, you need a data processing agreement with the provider and should only share the data that is genuinely necessary for the task. This is a technical assessment, not legal advice – when in doubt, that question belongs with your data protection officer.
My advice for getting started: pick exactly one of the three tasks – the weekly status updates work best. Track for two weeks how long they actually take, then let an agent create drafts for two or three pilot clients. After four weeks you'll know whether it's worth it. The building blocks we typically use for this are outlined on our pages for AI agents for businesses and process automation. If you'd like to work through this for your own agency, book a free initial consultation – we'll take a look together at which workflow is eating the most of your time.
We analyze your processes, identify automation potential and show you which digital systems really make sense.
