If you run an IT services firm or an ERP practice, your clients are probably asking you about AI agents already. You don't need to hire an AI team to say yes. You need a build partner behind you, one small first project at one client, and clear rules about who owns what. Here's how to set that up, how to scope the first project and what usually goes wrong.
Your clients are already asking
The questions sound simple. Can AI read the orders that come in by email? Can it answer customers who ask where their shipment is? Can it chase the invoices that are past due? At a distributor or a logistics company these jobs eat the week. And the person the owner asks first is often the one who already looks after their systems. That's you.
Most partners aren't ready for that question yet. Nerdio's 2026 State of the MSP Business Report surveyed 153 decision makers at managed service providers (MSPs, the firms that run IT for smaller companies) in June and July 2026. Only 14% run an AI service line that makes money. And 17% said customers are asking about Copilot or other AI tools and they aren't ready to respond.
OpenText saw the same gap in its 2025 survey of 1,019 people at MSPs in the US, UK and Canada. 92% said interest in AI is driving their growth. Only about half felt prepared to guide small and midsize clients in adopting AI tools.
So the demand is real, and most partners are watching it walk past.
The two easy answers both cost you
When a client asks for AI, the two easy answers are to say no or to build it yourself. Both cost more than they look.
Saying no feels safe. In July 2026 Channel Dive reported that Cynomi, a company that works with MSPs, went through a year of MSP discussions on Reddit and in its own community. It found MSPs looking for ways to turn down customer AI requests without losing the business. The trouble is that the client still wants it. They'll find someone who says yes, and that someone now has a reason to look at the rest of their systems.
Building it yourself sounds better, but an agent that works inside a client's business isn't a chatbot. It needs access to the ERP (the main system that holds orders, stock and invoices), rules for the odd cases, a way for a person to approve, testing on real emails and someone watching it after launch. That's a team, not a side project.
There's also research that suggests going alone is the harder road. MIT NANDA's 2025 study of AI projects, as reported by Fortune, found that buying from specialized vendors and building partnerships worked about 67% of the time. Internal builds worked only a third as often. It looked at companies building for themselves, and its method has critics, but the lesson carries over.
Three ways to partner without an AI team
The third answer is to keep the client and bring in a team that builds and runs agents. There are three common ways to set it up.
Referral
You introduce the client, and the build partner signs and delivers the project directly. You may get a referral fee. It's the lightest option and a good way to test a partner. The risk is that your client now has a second vendor, so pick one that stays in its lane.
White label
The build partner works under your name. You sell it, you own the contract and you talk to the client. The partner builds and runs the agent in the background. The relationship stays fully with you, but so does the risk when something breaks, so get support and response times in writing.
Co-delivery
You both show up. You handle the ERP, the access, the security review and the people. The build partner handles the agent, the rules, the testing and the monitoring. I think this is the best fit for most ERP consultants, because the hard part of an agent project is usually the data and the access, and that's what you already know.
Whichever you pick, settle four things before the first client meeting. Who owns the client. Who answers when the agent does something odd at 7 am. Who approves changes. How the money is split.
How to scope the first project
The first project decides whether there's a second one. Keep it small enough to show results in weeks, not quarters.
- Pick one client who trusts you and has a clear pain, like a shared inbox full of order or quote emails.
- Pick one piece of work, for example turning order emails into draft sales orders, and leave everything else out.
- Start with read access to the ERP where you can, and add write access once the drafts are right.
- Keep a person approving every action at first, then let the simplest cases go through on their own.
- Test on last month's emails before touching live ones, and compare the drafts with what the team actually did.
- Agree on two or three numbers up front, such as hours per week, time to answer and the share of drafts sent with no edits.
For a distributor, a good first project is order entry from emails and PDFs. For a third-party logistics company (3PL) or a freight broker, start with shipment status questions. On the finance side, supplier payment questions and collections emails are good candidates. Each one has clear inputs, one place where the right answer lives and a person who can approve.
What usually goes wrong
- Nobody owns the exceptions. The agent flags an odd order and it sits there, because no one at the client was named to handle it.
- The scope grows before the first part works. Someone asks for quotes, returns and invoices too, and nothing ships.
- Access takes longer than the build. ERP permissions, security reviews and a test company can take weeks, so start them on day one.
- The promise is too big. If the client hears fully automatic, every human review looks like a failure. Promise drafts that a person approves, then earn the rest.
- Nobody watches it after launch. Customers, email formats and ERP fields change, and someone has to check the results every week.
- There are no written rules. Nerdio found only 14% of MSPs consistently apply a documented AI governance framework (written rules for how AI is used) across every customer using AI tools.
- The money isn't agreed. Who pays the monthly running costs and who keeps what margin should be settled before the client sees a proposal.
Questions to ask a build partner
- Have you built agents that work with an ERP, and which ones?
- In your design, what stays with people and how do they approve?
- How do you test before launch, and on what data?
- Who watches the agent after launch, and what do you report each month?
- Will you work under our name, and will you stay away from our other services?
- Where does the client's data go, and who can see it?
If your clients keep asking
I think the partners who do well over the next few years won't be the ones with the biggest AI teams. They'll be the ones who already know the client's systems and can bring in the right builder for the agent work. That works best as a real partnership, where each side knows its part.
At OrbiResearch we're a small team that builds AI agents for logistics, sales and operations teams in the US. If you're an IT provider, an ERP or warehouse system consultant, or a fractional COO or CFO firm (part-time operations or finance leaders), and your clients keep asking about AI, I'm open to talking about any of the three setups. Work with us and tell me what your clients are asking for. You can also see the kind of system we build in our logistics operations case study.