LLM, chatbot, agent, MCP, skill, human in the loop are flying around every agency – but hardly anyone explains how they connect.
In short: an LLM is the brain, a chatbot wraps it in an interface, an agent can act with it – once context, tools/MCP, skills and a human in the loop are added.
This article explains the chain, agency-specific and without the tech jargon – promise 🤞
Why does AI fail in so many agencies?
It’s rarely down to the technology. The models are here, and they've been in use for a while.
But AI fails on the missing shared context: everyone prompts on their own, on their own data, so results stay inconsistent and can't be improved as a team.
The numbers from the Agency Happiness Report 2026 are clear:
98% of employees use AI every day, roughly half of all creative work will soon be AI-assisted, and 32% of clients demand AI in new projects.
Even so, only 16% of agencies feel genuinely ready.
[.b-testimonial]So the problem isn't usage – that's been happening for a while. The problem is the rollout across the team.[.b-testimonial]
How do LLM, chatbot and agent connect?
An LLM predicts text but can't act.
A chatbot is an LLM with an interface – handy, but everyone uses it on their own.
An agent can act, once context, tools and sign-offs are added.
The four terms build on one another, and each one solves the previous one's problem:
- The LLM is the brain – the model, like GPT-5, Claude Sonnet 4 or Gemini 2.5.
It does one thing: based on your prompts, it outputs the next most likely word, over and over, until the text streams out.
It doesn't know your agency processes and can't actively do anything.
- The chatbot brings the interface – like Perplexity, Claude or ChatGPT.
You type, it answers. You can store system prompts and memories. But in your agency everyone chats in their own window, on their own context: ten people, ten prompts, ten results.
- Context is the hinge.
A model only knows what's in its context window. A chatbot is only ever as good as what you paste in with your prompt each time – and no one pastes in all agency information. Long chats in particular get fuzzy fast.
What matters: agency, client, brand and project context.
- The agent is the leap to acting.
Give the chatbot tools – like web search or deep research – and it works in a loop until the task is done: from text to real groundwork in the project.
This is why more tools never fixes it. What's missing is shared context – not another chatbot.
What turns a chatbot into an agent?
Four building blocks. Tools & MCP so it can act and connect. Skills so it does things your way. A human in the loop so a person signs off before anything goes out.
- Tools & MCP: without them, the agent can reason but not reach.
Tools are web search, data access and predefined actions. MCP is essentially USB-C for AI and docks it onto your CRM, ERP or Slack. Without both, it falls back to being a chatbot.
- Skills: without them, the agent acts somehow – but not your way.
A skill is the packaged instruction "here's how we do it". Without skills it delivers fast, bur generic and off-brand.
- Human in the loop: without it, speed becomes a liability.
The checkpoint where the agent waits for sign-off before anything leaves the building. The more capable the agent, the more damage one unchecked step can do. No one wants to just throw automated emails at their clients.
At a glance:
An agent = LLM (reason) + context (know) + tools/MCP (act) + skills (consistency) + human in the loop (sign-off).
What does an AI agent look like in an agency?
Take a briefing agent:
a new briefing starts the agent, it reads the context, checks against a skill, acts through tools and presents the result to a person for sign-off.
- ⚡ Trigger: a new briefing lands in the project.
- 🗺 System prompt: the role – "briefing reviewer".
- 👀 Context: agency and client knowledge from the docs.
- 🪄 Skill: your briefing template with the review criteria.
- 🧰 Tools & 🔌 MCP: access to the project, web search, Slack.
- 🙋 Human in the loop: the account lead signs off the follow-up email.
The result in minutes rather than hours:
- a reviewed briefing
- documented open points
- a finished email draft
AI rollout in agencies
Understanding the chain of all these AI elements is the basis of your AI rollout. Getting it running across the whole team is its own topic: agencies can roll AI out in phases, from individuals to a fully agentic agency – each phase needs its own structure.
Conclusion
The models are here and the tools are here – what decides the outcome is whether the pieces connect.
An LLM without context is blind, an agent without skills is generic, and any agent without a human in the loop is a risk.
At a glance: the terms in context
- LLM – the brain that predicts text
- Chatbot – an LLM with an interface, but isolated
- Context – what the AI knows; the hinge
- Agent – a chatbot that can act
- MCP – docks the agent onto your tools
- Skill – teaches it your way of working
- Human in the loop – the human sign-off
👉 Get the chain right and AI stops being a jumble of individual prompts and starts working as one system.
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.
