Is MCP Dead? Is Saas dead?

Let us get nerdy as always, we will talk about MCP & Agent Skills
Followed by what on earth is going to happen to all the SaaS software we've spent the last decade or so building our businesses on.

I will go slow I know this AI train is hard to keep up with.

First, what an "AI Agent" actually is?

Before we get to the jargon, we need a mental model (not AI this time).

An AI agent is simply a software that can receive a goal set in plain English, figure out the steps needed to achieve it, and then go and do those steps. While doing so it will be calling APIs, reading files, sending messages, querying databases, the internet, automating what you or you and your coworker would have to plough through many screen to do yourself, and much faster.

Everyone loves to anthropomorphise AI, so lets see it as a capable (sometimes useless jr), new colleague who never sleeps, always does what they are told.

The interesting question is: how do you give an agent the knowledge it needs to do a job well, and how do you give it the tools it needs to actually do the job? In real world you would put the employee through and induction process and maybe let them read your outdates SOPs, then shadow someone to see how the business processes actually work... see well this is where MCP and Skills become relevant.

Agent Skills -teaching the Agent how to think

So on their first day, you put the new employee onto one of the tasks in the department. You hand them the single Standard operating Procedure out of the folder of documents that contains your company processes etc. They read the folder matrix skill style, seconds later, now they know, how you do things here.

That guide document is essentially what an Agent Skill is.

Technically, a Skill is a file in markdown format (most often called SKILL.md). Sometimes with a few supporting scripts or templates with it too. It encodes knowledge and workflows. In contains conventions, expertise and, when you encounter X, do Y. Skills is an open standard, and many modern AI tools have support baked in now.

a Skill doesn't actually do anything by itself. It contains no execution ability. It can't call an API. It can't write to your database. It can't send a Slack message. It just tells the agent how work should be approached.

This makes Skills extremely lightweight, easy to use, share and maintain. You simply write a markdown file, the agent reads it, and suddenly it understands your team's conventions as well as someone who's been working with you for six month (if only it were like that).

MCP -Empowering the agent to actually do

Understanding what to do and being able to do it are two entirely different things. You need to give the new employee a desk/mobile phone, a computer, keys to the filing cabinets etc.

MCP the Model Context Protocol is what gives agents the ability to act in the real world.

MCP is a standardised protocol ("USB-C for AI", such an over used analogy) that lets an agent connect to external tools and services
Your database, your CRM, email server, your file system, your internal APIs. When an agent needs to query your Dynamics ERP data, it sends a properly authenticated request through the MCP layer, gets real data back.

MCP is now a open source standard.

In my view I think it is vital to take away that MCP handles the critical production environment stuff:
authentication, permissions, audit trails, error handling, rate limiting, persistent connections. Your deterministic ERP and CRM systems, where you need 100% accuracy, need a rigorous, permissioned interface to the probabilistic AI layer that doesn't like to stick to the rules. This is the role of MCP.

Is MCP dead because of skills?

I don't recognise this question. When Skills came along last year (this stuff moves so quick), people saw some MCP servers get replaced by skills and may have concluded that MCP is murdered by Skills. As I understand it, this is nonsense! What were replaced were specific badly designed MCP implementations, that didn't need to exist in the first place.

The agent needs both: the knowledge of how to do something, and the ability to actually do it.

Right now I don't see MCP going anywhere for integrating to real systems, involving real data, with real permissions etc.

Well, is SaaS dead then?

I don't for the immediate future see AI replacing all software, and its much more powerful than just a chat window. So the future is somewhere in the middle.

AI is not replacing SaaS. It is replacing the human who operates SaaS.

Your CRM still holds your customer data.
Your ERP still runs your financial processes.
Your HR platform still manages your people records.

Those systems have years of integrations, compliance certifications, audit histories, and business logic baked into them. They're not going anywhere very soon, well not in the enterprise and SME/SMB space anyhow. The data they hold is becoming so much more valuable going forward, think about it. AI agents will be able to unlock the data they hold, in ways never seen before, and at scale.

What is being disrupted is the UI layer. The dashboard, the many menus and pop up windows all just to approve an expense report. The SaaS UI is how humans interact with business software, AI Agents ideally needs things CLI access, they need tools to go to stuff direct.

Tell an AI agent "approve expense reports" and it orchestrates the workflow across your HR system, your finance platform, and your approval trail behind the scenes, via APIs, following the rules encoded in Skills and enforced and enabled through MCP. Saas is there in the background, AI will front it.

Think of it as unbundling and re-bundling, the interface being stripped away. The underlying data and business logic layer becomes stronger as it needs to, because it's now the engine that AI agents are running against.

What is going away today's licencing

SaaS tends to work on per-seat billing. You have 200 employees, you pay for 200 seats.

In the new world, those 200 employees each have a team of AI agents doing research, drafting documents, processing data, updating records. One person is doing the work of five, but through one licence seat. The revenue model is not fit for use anymore.

So what works, which allows the Saas providers to get paid for the value they bring to the table. Outcome-based pricing? Consumption-based pricing? Pricing around the value delivered to the organisation rather than the number of humans sitting in front of screens? Per-seat model was designed for a world where software was operated by humans, with AI Agents popping up in all platforms, that era is already departing.

Summary

The data and compliance layer, your systems of record, your governed data, your audit trails. This becomes more important, not less. AI agents are only as good as the data they can access.

The deterministic SaaS backends, your CRM, ERP, HR platform. They don't go away. They get pushed further into the background, accessed through APIs rather than GUIs. Their job is to hold the business logic and the data with 100% reliability.

The MCP orchestration layer, this is the safe, permissioned, audited bridge between AI agents and your real systems.

The Skills layer, that tells your agents how to behave your company's conventions, your workflows.

The agent interface, the thing your employees actually interact with. Plain english objective orientated control.

AI is adding another layer, bringing more power and efficiency.

Are they prepared?

Most organisations just dont have the level of data integrity nor documented business processes in the level of detail required to make agents truly useful. There needs to be a big effort before they can fully harness AI power in the core of the business. Anyone involved in ERP or CRM implementations will know this and also have scars to show for it.
With AI there remain real challenges still around governance, hallucination, determinism in critical processes, that need to be solved, but leaps forward are being made every few months, so it won't be long coming.

This post was inspired by
Martin Olsen who posted "MCP is dead"- which caught my eye, as MCP being dead is something I have not seen in my haunts -see what I did there?

Martin's post is here: "MCP is dead" based on work by Robert Porter