AI for MSPs: why enthusiasm without process burns money

AI for MSPs currently feels like a manual that is rewritten daily. A new pop-up every morning, a new demo every week, and at the end of the quarter the question: what has this actually earned us? Everyone is talking about it. Few MSP owners dare say out loud that they do not yet see a single euro of extra margin. And that is exactly the conversation we are having here, without the stage talk.

Why AI for MSPs is producing so much noise right now

The owner reading this knows the pattern. Your engineers come back from a vendor conference and want to roll out the new copilot feature tomorrow. Your security officer wants to work out first how that sits with data protection law and with data walking through OpenAI. And you are in between, with a diary full of calls and a management team that gets excited and puts the brakes on at the same time.

At the same time you read that NVIDIA is putting billions into OpenAI, that OpenAI is buying chips back again, that AMD and Amazon are joining the same carousel. The money is pumping around a very small club of companies. That makes sharp MSP owners careful without making them drop out. The dotcom bubble is still fresh enough in memory not to pull out a credit card blindly for the newest tool.

The research MIT published two months ago on AI projects in business confirms what you already sense. Nine out of ten projects do not reach the finish. Not because the technology does not work, but because the company underneath was not ready. And that is exactly what most MSPs are running into right now.

What does work with AI for MSPs

What works in practice is less glamorous than the demos on stage. It starts with a simple question: is your data lake clean? If your own processes and your own data quality are not in order, an AI tool merely shines a spotlight on the mess. You pull out your credit card, buy a tool, turn it loose on your sources, and you get visible chaos back instead of usable output.

The MSPs getting results right now do two things differently. They do not build it themselves, and they do not let AI decide on its own. They wait until their PSA or RMM vendor ships the feature, and they put a person in between on every process that reaches the client.

That sounds conservative. It is realistic. An MSP with sixty staff cannot compete with the development capacity of a platform supplier putting a hundred engineers on the same feature. Every time you build something clever yourself, a year later it is in the standard application, more powerful, better maintained, and without you having to do the bug fixes.

What you do invest in yourself is the implementation. Not the tool, but how you set it up. The instructions, the client context, the tone of voice towards end users. That is not secret sauce, that is simply listening to what the client wants. And the client does not want purple, green, blue reports that nobody understands. They call you because something is broken.

Where AI for MSPs earns real money today

The most tangible example right now sits in the service desk. Having incoming tickets read by OpenAI, categorised by priority, the tone of the message assessed, and a draft reply proposed to the engineer based on the content. That is triage that used to cost half an FTE. Now it runs in a workflow tool in the background and the engineer only picks out the outliers. Important detail: the engineer presses send, not the model. As long as that person sits in between, you hold on to the quality your client has an SLA on.

From raw automation to recognition

A second example that works: report automation with Power Automate. Output comes in from an application you do not immediately recognise, the client is matched from your PSA, a flow starts, a ticket is created, all follow-up communication is attached automatically to the same thread. The client gets a response within minutes to something that used to take half a day. Not because AI generated the answer, but because AI recognised the context.

On the engineering side you see where the real gain sits. Engineers who no longer write lines of code, but correct the output of a model. That does not save 10 per cent of the time, it saves hours per week per engineer. And that is the kind of gain that shows up in your margin, not in a management report. Work it out for a team of twenty engineers: two hours a week per person is forty hours, a whole FTE you get back without opening a vacancy.

Why most rollouts still fall over

Where it nearly always goes wrong: the plan-of-action story. The board proudly shows a template that AI fills in ten minutes. Two weeks later you run into an engineer who is still typing their own version with three fingers. The problem is not the tool, the problem is that nobody has agreed who takes over which part of the process. Enthusiasm in the demo phase is free. Adoption in the daily routine costs leadership.

That is why no AI rollout works without what you might internally call a dreaming room, a feasibility room and a constraints room. Where you may fantasise, where you test what is achievable, and where you say a hard no. Without that framework every member of the management team gets excited about something different, and at the end of the quarter nobody has finished anything. In an MSP with 40 to 80 staff in particular, that is the most important brake you have to build in yourself.

What helps is naming one owner per workflow, with a quarterly target you can express in euros or hours. Not “we are going to use AI for reporting”, but “we are taking two hours a week out of the reporting line for client X, measured in tickets that are no longer created by hand”. That is the difference between an innovation project that quietly evaporates and an improvement you see back in your P&L.

The MSPs that still have margin from AI in two years are not the MSPs that switch on every tool fastest. They are the MSPs whose data is in order, who ride along with the vendor instead of building against it, and who dare to put a person between model and client. In the MSP Late Night performance club, owners share every month which workflows really earn money and which experiments were quietly shot down. Joining in saves you a year of working it out yourself.