HOW TO BUILD AN AI PRACTICE INSIDE YOUR MSP
Most MSP AI practices fail for the same three reasons. Nobody owns it. The data underneath is a mess. It gets sold before anyone knows how to deliver it. Here is the order that avoids all three.
Every vendor at every conference is telling you to launch an AI offering. None of them will tell you what it costs to deliver or whether it makes money. That is not because they are hiding it. It is because they genuinely do not know your delivery cost, and their program was never designed to find out.
You cannot sell what you do not run. Build it inside your own business first.
Inventory what you are already paying for
Before you buy anything else, find out what AI you already own. Copilot seats bundled into Microsoft 365. AI features switched on inside your PSA, your RMM, your documentation tool, your phone system. Individual subscriptions your team expensed. List every line item, what it costs, who owns it, and what data it touches. Most MSPs I do this with find overlapping tools and at least one subscription nobody remembers buying. Consolidation usually pays for a chunk of the work.
Write the policy before someone needs it
Your techs are already pasting client data into chat windows. That is not a hypothetical. The policy answers who has access to which tools, what data can go in and what cannot, how client data is handled, what you say when a client asks about your AI use, and what happens when something goes wrong. Write it against your actual stack and your actual client contracts. A downloaded template does not make the decision about whether your level one techs should have client data in a chat window. You have to make that call.
Look honestly at your data
Automation runs on the data in your PSA and RMM. If ticket categorization is inconsistent, if time entries are guesses, if documentation is thin, that constrains what you can build. Find out in month one, not month four. Bad data is not a reason to stop. Plenty of useful work runs fine on imperfect records. But you need to know which work that is before you promise anything.
Build two or three automations inside your own operation
Not a pilot with a client. Inside your own shop, where the cost of getting it wrong is your time instead of your reputation. Pick the ones with hours attached: ticket triage and routing, documentation generated from resolved tickets, proposal and quote drafting, after hours ticket summaries for the morning handoff, onboarding runbooks, alert noise reduction. Measure what came back. Eight hours a week returned to your team is roughly $19,000 a year at a fully loaded $45 an hour, and you cannot claim that number to a client until you have measured it on yourself. Have your builder do the building. If you do it yourself, the capability leaves when you get busy.
Cost the delivery before you price the offering
This is where most MSP AI offerings fall apart. The tooling cost is the small part. The real cost is the labor to configure, monitor, tune, and support it per client. Take one of the automations you built internally and work out honestly what it would take to deliver that to a client every month. Then price it. If the margin is not there, you found out before you sold it, which is the whole point of doing it in this order.
Define the package, then sell it to two or three clients
What exactly do you sell, to which clients, at what price, with what delivery model behind it. Then have the conversation with a small number of clients who already trust you. Your salesperson or you needs to be able to explain the offering, qualify a client, and handle the obvious objections without phoning a consultant. Two or three conversations tell you more about the offering than six months of planning.
THE THING NOBODY TELLS YOU
The internal work is not a warm-up for the client offering. It is the offering. The policy you write for yourself is the policy you sell. The automations you built are the proof. The cost model you worked out on your own operation is the only honest basis for pricing someone else's.
MSPs that skip step one through five and go straight to step six end up selling a service they cannot deliver at a margin they cannot hold, to clients who will remember it.
WHO DOES THE BUILDING
One person other than you. Technical or ops, either works. Some of the best builders came out of service coordination, not engineering. What matters is that they are curious and they finish things. Four to six hours a week is enough to start.
Do not hire for this. If you have to hire to make it work, you are not ready yet, and that is a real answer rather than a failure.
WANT TO KNOW IF YOU ARE READY?
Six questions, scored, with a straight answer about where you stand.
Want someone in the room while you do it? See the AI Practice Build engagement.