TL;DR:
- AI is limited when it stays inside a chat window.
- To hand over workflows, it needs access to tools and context.
- MCP servers expose actions AI can execute without knowing each system’s internal details.
- Skills provide reusable instructions for how work should be done.
- The combination makes AI much more practical in day-to-day work.
Why chat is not enough
When using AI only in chat windows, it might not be straightforward why so many people talk about it changing white-collar work. While chatting with an AI can be helpful, the way AI can do one’s work is fairly limited this way. Even when satisfied with an answer, one still has to copy and paste it somewhere or, worse, do the task themselves. However, outside of a very small bubble this has been and still is the case for many use cases of AI.
Where it breaks down
The problem with this approach is that it is not scalable enough to change the way we work in general, as it is so often promised nowadays (whatever that’s supposed to mean anyway). Generating text alone is not enough to hand off complete workflows to AI. Also, most tools AI agents can use are not yet powerful enough. There was, and in many cases still is, a lack of integration of AI into the systems we work with day to day.
An AI agent is an AI system that can choose from a set of tools to complete a task. Instead of only answering with text, it can search, edit, click, run commands, or call external systems when those capabilities are available.
What is missing
To allow AI to be helpful to us, we have to allow it to access the tools and information we need to complete our tasks. At the same time, we have to instruct it on how the work should be performed and what it should take care of for each task. How can we integrate AI into our systems in a scalable and reliable way?
Tools
The answer lies in packaging the tools and information we have into small modules that can be consumed by AI. There are different ways to do so. On the one hand, there are MCP servers.
MCP stands for Model Context Protocol. In simple terms, an MCP server exposes tools or data sources in a structured way so an AI system can use them. Instead of hardcoding every integration into every chat app, a tool can be provided once and then reused by compatible AI clients.
They are mostly used to provide tools to an AI that it can then use to interact with the world outside the chat window. This way, AI does not need to know how a customer follow-up is entered into a CRM system or how a meeting invite is sent in the background. It can just execute the right tool and does not have to worry about the implementation details.
For example, a tool provided through an MCP server might look like this:
{
"name": "create_follow_up",
"description": "Create a follow-up task for a customer conversation",
"inputSchema": {
"type": "object",
"properties": {
"customer_name": { "type": "string" },
"next_step": { "type": "string" },
"due_date": { "type": "string" }
},
"required": ["customer_name", "next_step"]
}
}Skills
Then there are skills.
Skills are reusable instructions for AI. They usually include a short description, a clear trigger, the tools involved, and a few constraints so the model knows when to use them and what to avoid. If MCP servers provide the hands, skills provide the playbook.
They can be used to provide standardized instructions to AI, for example which tools to use in which order and what to take care of when executing a task. They instruct AI on which tools to use in which way and in which order when working on a task.
For example, a skill for a sales manager could look like this:
---
name: Customer Follow-up
description: Draft and create customer follow-ups from CRM records and meeting notes.
---
# Customer Follow-up Instructions
When I ask you to "Create a customer follow-up":
- **Information Check:** Ask for the customer name, next step, due date, and source notes if any of them are missing.
- **Tool Selection:** Use the `get_customer_record` and `list_follow_up_types` tools from the MCP server before drafting the follow-up so you use the correct customer and format.
- **Drafting:** Write the title as a short action. Keep the body to a short summary, the agreed next step, and any open questions.
- **Output Format:** Show the draft first. If it looks right, then use the `create_follow_up` tool from the MCP server to create it.By combining skills and MCP servers, AI can be integrated into a wide range of systems, limited mainly by the availability of tools and skills, whose variety is only growing. Step by step, those capabilities will be introduced into the mainstream and therefore provided to people outside of the AI bubble. This will mean that suddenly, customer follow-ups for other people to work on or even a weekly action plan can be created by steering an AI agent and giving detailed instructions on what to include and what to leave out at each stage of the process. While using a skill as orientation, an AI agent can then ask a team lead for relevant information and use tools provided by MCP servers to execute any number of tasks, limited mainly by human judgment and compute.
The same applies outside project management. A recruiting team could define a skill for screening incoming applications and use MCP tools to read CVs, update an applicant tracking system, and prepare interview summaries. The people in the loop would still make the hiring decision, but much of the repetitive coordination work could be delegated.
In this setup, the human suddenly becomes the bottleneck, having to be worried about not having the perfect strategy or not steering AI in the most effective way. Only when looking at AI from this perspective can one fully understand its potential. The technology is there. What comes next is a broader phase of integration and transformation across how we work.