This is the last part of my AI platform series. If you missed the start, read Choosing Your MSP AI Platform first. If you want help deciding between building and buying, Schedule a Call.
What building your own means
You license models directly through the OpenAI API, the Claude API, Azure OpenAI, or Amazon Bedrock. Then you build the rest yourself: multi tenant controls, audit logs, usage metering, billing, and the agents your clients pay for.
Why MSPs build
Control and margin
No platform sits between you and the model. You set your own markup on usage, and you own your agents as code.
No platform vendor risk
If a platform gets acquired, changes pricing, or shuts down, it doesn't take your practice with it. You still depend on the model makers, but you can move between them.
The right provider per client
You pick a BAA covered provider for a healthcare client and a different one for everyone else.
Why most MSPs shouldn't, yet
You carry the paperwork
No SOC 2 report covers what you built. Every client security questionnaire lands on you. Audit logs, data retention, and access controls are yours to design and prove.
Key person risk
Rewst's CEO estimates about 1,000 automation engineers serve about 80,000 MSPs. That's a vendor estimate, but the point holds. Builders are scarce. If yours leaves, your practice walks out the door with them.
Token costs swing
Usage doesn't stay flat. Without your own metering, one busy client eats your margin before you notice.
Nobody's proven it in public
I found zero firsthand accounts of an MSP running a client facing chat stack built this way. The only practitioner voices were early testers of new automation tools. You'd be early.
The math
Here's an estimate. One AI engineer at $130,000 loaded costs about $10,833 a month. Fifteen clients at $1,000 a month bring in $15,000. Take out about 10% for API costs and you keep $13,500.
That leaves about $2,667 a month after the engineer. Breakeven sits near 12 clients. The model works past about 20 clients, or with project revenue on top. Expect 2 to 4 months before first revenue.
Pricing is shifting under all of this. An Omdia poll of 255 MSPs found 60% moving off per user pricing toward hybrid, outcome, or consumption pricing. AI usage costs don't fit a flat seat price.
Where each path breaks
Every option in this series fails in its own way. Know yours before you pick.
- Copilot breaks on adoption and data readiness. People don't use it, or it surfaces files they shouldn't see. From December 1, it adds a billing risk you manage with caps.
- ChatGPT and Claude break on governance. Personal accounts, connectors left on, and no BAA on the plans small clients buy.
- Gemini breaks on Drive sharing, and on a thin track record with MSPs.
- Hatz breaks on credits you can't forecast and on clients who want a mobile or desktop app.
- Synthreo breaks on validation. You need a SOC 2 report and outside references before you commit.
- Your own stack breaks on people. One builder, one resignation, one practice gone.
Most MSP AI practices don't fail on the technology. They fail when pilots never convert and the owner is the only one who can sell it.
How to choose your path
Pick by client type
- Healthcare and dental: Copilot inside the client's tenant, after you confirm Microsoft's BAA scope.
- Defense contractors with CUI: Copilot in GCC High.
- Accounting and financial services: Copilot, with retention and audit inside the tenant.
- Construction and lower regulation SMB: Hatz, for whole company rollouts.
- Canadian clients with residency needs: Synthreo with inference locked to Canada, or Copilot.
- Nonprofits and education: whatever the client already pays for, Copilot Chat or Gemini.
- Premium custom agent work: Synthreo, or your own stack if you have the builder.
Pick by your shop
If you're Microsoft heavy and short on builders, start with a Copilot readiness practice. If you want a packaged offer fast for lower regulation clients, look at Hatz. If you have a builder and a consultative seller, look at Synthreo. If you have two builders, real documentation discipline, and 20 clients ready to buy, building your own starts to make sense.
Be honest about which one you are. The MSP Owner Reality Check helps.
Questions to ask any AI vendor
- What's my price, minimum, and term, on a signed quote?
- Will you sign a BAA, and which models and providers does it cover?
- Do you have a SOC 2 report? Which type, and what's the audit window?
- Which provider runs inference for each model, and in which region?
- What can I export when I leave, in what format, and how fast?
- What uptime do you commit to in writing, and do you pay service credits?
- Can I call three partners outside your inner circle?
What to fix in your own contracts
Talk to a lawyer before you sell AI. Your client agreement needs a clause on who owns errors in AI generated work. It can't promise more uptime, retention, or deletion than your vendor gives you. It needs a pass through clause for metered usage, including Copilot credits from December 1. And it needs consent language before you scan any client endpoint.
Do your own homework
This series is built on public information, my own hands-on testing, and written answers from vendors to my questions. It's my read as of October 25, 2026. Prices, terms, and features change fast. Don't take anything here as fact without checking it yourself. Get every claim that matters to you in writing before you sign.
That's the series. Want help picking your path and building the practice around it? Schedule a Call. Want more like this? Subscribe to my newsletter.
FAQ
Is building my own AI stack cheaper?
It has the best usage margin, but an engineer costs about $10,833 a month loaded. By my estimate, you need around 12 clients to break even and 20 to make it work.
What do I need before I build my own?
At least one dedicated builder, documentation discipline, your own security review, and enough clients ready to buy.
Which AI platform is best for MSPs?
There isn't one best platform. Pick by your clients' regulations and stack, and by the builders and sellers you have on staff.