AI will fundamentally transform agency work.
Already, 98 per cent of agency staff use AI in their work (Agency Happiness Report 2026). They also expect half of all creative work to be taken over by AI within the next three years. We do not yet know whether this crystal-ball prediction will come true.
Either way, however, it is essential for everyone in the agency to truly understand the key fundamentals of AI. The best place to start is with the most important terms, their definitions and how they relate to one another.
Chatbot, agent, MCP, skill, human-in-the-loop – these terms are bandied about in every agency. In this article, we explain the technical concepts simply and precisely.
[.no-toc]AI in the agency: Key terms explained[.no-toc]
LLM
LLM stands for ‘Large Language Model’. The name describes how it works: an LLM predicts text. Based on your prompt, it calculates the most likely next word and repeats this process until the text is complete. This technology underpins AI models such as ChatGPT and Claude. Behind every model is a provider – the company that develops it: GPT comes from OpenAI, Claude from Anthropic, and Gemini from Google.
Why is this relevant for agencies?
Providers regularly release new versions of their models (identified by abbreviations such as GPT-5 or Claude Sonnet 4.6), which differ in terms of speed, cost and strengths.
Depending on the task you want to tackle with AI, it makes sense to find out which model is most suitable. For quick social media captions, a fast, affordable model will suffice – for an in-depth analysis of a complex briefing, a more powerful one is needed. Another detail that is often overlooked: the LLM provider does not, at this stage, disclose where (in which country or data centre) it is operated.
It is therefore entirely possible to use ChatGPT directly from OpenAI – and thus, in most cases, on servers in the US – but there are also providers who operate the same model in the EU and often even in Germany. This makes it possible to use different models in a GDPR-compliant manner, regardless of where the original provider is based.
Chatbot
A chatbot is an LLM with a chat interface: you type in a question, and the chatbot replies in natural language. You’re no doubt familiar with ChatGPT, Perplexity and so on. In addition to the LLM itself, you can set up a system prompt (for example, a fixed role or your tone of voice), and there’s often a memory function that extends beyond individual chats.
A chatbot is useful for anything that’s quick and sticks in your mind: rephrasing a sentence, double-checking an idea, or summarising a meeting you’ve just attended yourself. You provide the context on the spot – that’s all it takes.
Use cases in the agency:
- Running through an ad headline in different tones
- Summarising a meeting transcript from the regular meeting
- Having a technical term explained from a client briefing
- Carrying out a simple search
For tasks involving multiple steps, recurring to-dos or tasks requiring access to external tools or knowledge, an agent is the best choice.
Context
Context refers to the totality of all the information currently available to an LLM for processing a task. For example, the prompt, uploaded files, the entire chat history, connected tools, etc. Generally speaking, the more relevant context there is, the more precise the answer will be. However, every model has a limited context window – that is, the amount of information it can process at any one time. For complex tasks in particular, it is therefore crucial to make the best possible use of this context window and to prepare the information within it as precisely as possible.
Quality is Key
Much like a person’s short-term memory, an LLM also becomes confused when it has to sift through too much varied, contradictory or irrelevant information for a specific task. The general rule is therefore: less context, but more precise, leads to better results than a blind mass of data. On top of that, the entire context provided (i.e. absolutely everything you put into the chat) has to be processed in full again for every single response.
On the one hand, this means you’ll have to wait longer and longer for replies and – perhaps most importantly – your prompts will become much more expensive. It’s also a good idea to always bear in mind that the entire history of a conversation with the chatbot becomes part of the context and is re-read with every reply.
So if you carry on chatting in the same conversation thread for days on end, it becomes slow, imprecise and expensive. You should therefore start new conversations regularly.
Custom GPT & Projects
In most AI environments, you can group chats into projects. The purpose of this is to ensure that you only need to add the necessary context once for all chats taking place within that project (or that custom GPT, etc.). You can also store instructions in a central location (known as ‘system prompts’), which are then automatically sent along with your other prompts every time.
These could include, for example, brand guidelines, tone of voice, etc. If, for instance, something important is learnt in a project chat that will remain relevant in future, the chatbot can incorporate this directly into the project’s memory for all subsequent conversations, and so on. Projects are therefore something like a simple, quickly created precursor to fully-fledged agents (see below).
Agent
In a way, an agent is a chatbot that can take action. Whilst a chatbot responds within the chat window, an agent can also carry out actions outside the chat. The difference is that an agent has various capabilities (known as ‘skills’) and access to tools. For example, it can use a web search to look for information online. Or it can access your tools and systems, if you have integrated them, and carry out actions there.
An agent can therefore carry out much more complex and better standardised tasks than a simple chatbot.
Examples of agency agents:
- Briefing Checker, which checks incoming client briefs for missing information
- Persona Reviewer, which evaluates ideas from the perspective of client personas
- Strategy Drafter, which develops initial approaches for campaigns.
- Further examples in the article: Which AI agents will every agency need in 2026?
[.b-button-primary] Discover more useful agents [.b-button-primary]
MCP
MCP (Model Context Protocol) is a unified standard that allows an agent to access your existing tools – such as your project management tool (hopefully awork 😉), your CRM or Slack – without the need to build a separate, bespoke interface for each connection.
The core idea: the agent and the tool speak a common ‘language’ via MCP, rather than relying on individual, ad-hoc integrations. The standardised format also enables an agent to understand, without further instruction, what they can do within the connected tool. This saves effort and makes integrations interchangeable.
Example:
An agent retrieves the project status from your PM tool via MCP and posts a summary directly to Slack.
[.b-button-secondary] List of MCP connectors [.b-button-secondary]
Skills
In the context of AI, skills are reusable abilities or instructions that an agent or chatbot retrieves as and when needed to perform a task more competently – rather than starting from scratch with every enquiry.
Core idea: Instead of equipping the agent with all the knowledge at once (which would be inefficient, as it takes longer and consumes more tokens – and therefore costs more), it is given access to a library of skills. Each skill contains specialised knowledge, best practices or tools for a specific type of task. The agent dynamically retrieves the appropriate skill when the task requires it – or is instructed directly to use it.
Why this is relevant:
- Makes agents more precise and consistent with recurring task types, rather than remaining generic
- Skills can be customised to suit the specific needs of the organisation (e.g. custom templates, processes, terminology)
- Modular and expandable – new capabilities can be added without altering the core system
Example:
An SEO agency presents its clients with the latest KPIs once a month. A ‘Client Presentations’ skill knows how to structure slide designs according to specific standards. The agent only loads the skill that is currently needed.
Other agency skills:
- Competitor Analysis’ – how to evaluate competitors’ websites according to the agency’s internal criteria
- ‘Quotation Template’ – how to format cost calculations and service specifications
- ‘Customer reporting’ – which metrics are relevant for which type of customer, and how they are presented
Agent Loop
What also distinguishes agents from simple chatbots is that they can take a step-by-step approach to more complex tasks. This ultimately takes place in a cycle known as the ‘agent loop’. An agent receives a request as a prompt and carries out an action. It then checks whether this action has achieved the desired result. If so: display the result. If not: add the interim result to the context and carry out the next action. In this way, an agent works its way through multi-stage tasks that it cannot process in a single pass. Often, an agent even starts by drawing up its own to-do list, which it then checks once during each pass of its agent loop to see if anything remains.
What does ‘Human in the Loop’ mean?
‘Human in the Loop’ (HITL) means that a human remains actively involved in a (partially) automated or AI-controlled process and intervenes at specific points – rather than the system deciding or acting completely autonomously.
For example, to check, correct or approve something before an action is finally carried out.
This can occur at various stages:
- Before the action: The AI makes a suggestion; the human must confirm it before anything happens (e.g. an AI-generated email is only sent once it has been approved)
- After the action (human-in-the-loop): the AI acts automatically, whilst a human monitors and can intervene or correct if necessary
- During training: humans evaluate or correct AI outputs to improve the system.
At a glance: an overview of the terms
- LLM – the brain that predicts text
- Chatbot – an LLM with an interface, but isolated
- Context – what the AI knows; the hinge
- Custom GPT & projects
- Agent – a chatbot that can act
- MCP – docks the agent onto your tools
- Skill – teaches it your way of working
- Agent Loop – multi-stage processing of to-dos
- Human in the loop – the human sign-off
You now have a rough overview of the key terms in the AI universe. And even if you haven’t immediately grasped every aspect of how agents, skills and the like work, that’s no problem. Developments in the field of AI are happening at breakneck speed. It’s not always easy to keep up. Feel free to follow us on LinkedIn for more information.
[.b-button-primary] All about AI in awork [.b-button-primary]
FAQs
What's the difference between a chatbot and an agent?
A chatbot can talk; an agent can act. The chatbot answers your question in the chat window; the agent uses tools and context to actually complete a task in a loop – like reviewing a briefing and drafting the follow-up email.
Do agencies need their own developers for AI agents?
No. Standards like MCP and ready-made agent templates let you connect and adapt agents without coding yourself. What matters is less code than central context, shared skills and clear permissions.
Are AI agents in agencies GDPR-compliant?
That depends on the setup. What matters is EU hosting, keeping client data out of training, clear access rights, and a human in the loop for anything that goes out. In awork, agents run model-independently and GDPR-compliant on servers in the EU.
Where do most agencies stand with their AI rollout?
In the early phases: individuals use chatbots and best practices are shared verbally. Only 16% feel genuinely ready. The next step is central context and shared, specialised agents.
