Most brokers start every carrier payment with a person opening the POD. Uber Freight flipped that. Clean paperwork gets paid automatically and people only see the problems. What does your team check by hand before paying a carrier?
2
You don't need an AI team to sell AI agents to your clients.
If you run an IT services firm or an ERP practice, your clients are already asking. Can AI read our order emails? Can it answer customers who ask where their shipment is? Can it chase the invoices that are past due?
Most partners aren't ready for that question.
Nerdio's 2026 State of the MSP Business Report surveyed 153 MSPs in June and July. Only 14% run an AI service line that makes money. 17% said customers are asking about AI tools and they aren't ready to respond.
OpenText's 2025 survey of 1,019 people at MSPs found the same gap. 92% said interest in AI is driving their growth. Only about half felt prepared to guide small and midsize clients on AI tools.
When a client asks, there are two easy answers. Both cost you.
Say no, and the client finds someone who says yes. That someone now has a reason to look at the rest of their systems.
Build it yourself, and you find out an agent isn't a chatbot. It needs ERP access, rules for odd cases, a person who approves, testing on real emails and someone watching it after launch. That's a team.
The third answer is to keep the client and bring in a build partner.
1. Refer the client and let the partner deliver.
2. Sell it under your name while the partner builds it.
3. Work together, you on the ERP, they on the agent.
Then keep the first project small.
One client who trusts you. One kind of work, like turning order emails into draft sales orders. Read access to the ERP first. A person approves every action at first. Test on last month's emails. Two or three numbers agreed up front.
What usually goes wrong? Nobody owns the exceptions. The scope grows before the first part works. Access takes longer than the build. The promise is too big. Nobody watches it after launch.
If you're an IT or ERP partner and your clients keep asking about AI, DM me.
What's the first thing your clients ask AI to do?
Full breakdown is in my profile.
1
1
32
When a client asks for AI, saying no feels safe. It isn't. They'll find someone who says yes, and that someone gets a look at the rest of their systems. Bring in a build partner and start with one small project. What do you tell clients today?
12
Your suppliers keep asking "did you pay me?"
Here are 5 fixes, ranked from weakest to strongest. Fixes 2 to 4 cost almost nothing, so start there.
Reply INBOX and I'll send you the full playbook, with a remittance note you can copy.
47
Half of AP teams spend six or more hours a month answering one question from suppliers.
"Has my invoice been paid?"
That's MineralTree's 2024 State of AP Report, more than 1,100 finance people surveyed.
The usual answer is a supplier portal. It doesn't work as well as people hope. In MineralTree's 2023 survey, 88.6% of suppliers said they prefer to ask about a payment by email. Only 3.3% picked a self-service portal.
So build the portal if you like. The emails keep coming.
It's getting worse too. IOFM found the share of invoices taking more than 15 days to process went from 5% to 25% in about six months. Slower invoices, more suppliers asking.
Here's the system I'd build instead. An agent on the AP inbox.
1. It reads the email and pulls out the invoice number, PO number or amount.
2. It checks the sender's address matches the supplier on record.
3. It looks the invoice up in your ERP. Received, on hold, scheduled or paid.
4. It drafts a short reply with the status, the date and the payment reference.
5. It logs the question next to the invoice.
People approve anything unusual. Holds, disputes, short payments, angry emails.
Two rules from day one.
It never changes bank details from an email. Fake supplier emails asking to change where money goes are a common fraud.
It never promises a pay date that isn't already scheduled.
What to measure. Hours per month on status emails, time to answer, share of emails closed with no edits, repeat emails about the same invoice.
How to start small. Pull last month's supplier emails, tag the status questions, run the agent on them in draft mode and compare its replies to what your team sent.
If supplier emails eat your AP team's week, DM me.
How many "where's my payment" emails hit your AP inbox each week?
Full breakdown is in my profile.
14
A supplier portal won't empty your AP inbox. 88.6% of suppliers prefer to ask about payments by email. Only 3.3% pick a portal. Fix the inbox instead. What do your suppliers use?
11
Business Central now has an AI agent that turns order emails into quotes and orders. Microsoft's own example puts the AI bill at $16.50 a month for 100 requests, half of them with a PDF.
That number is the wrong thing to look at.
Here's what the agent does, in order.
1. It watches one Microsoft 365 mailbox.
2. It reads the email and any PDF, PNG or JPG attachment.
3. It matches the sender's email address to a contact card to find the customer.
4. It finds items by your item number, the vendor's number, the barcode or part numbers you've stored for that customer.
5. It checks stock, creates a quote and drafts a reply.
6. A person approves every email before it goes out.
The cost is tiny. Microsoft charges 2 credits to read an email, 5 to read a PDF with order data, 5 each to check stock, create a quote or create an order, and 2 per reply. At one cent per credit, a full round trip for one order with every step on is about 28 cents. That's example math from Microsoft's list prices, not a promise.
So what decides it? How many of your real orders it can finish while your team only clicks approve.
Microsoft's own pages say where it needs help.
A forwarded email. The agent sees your rep as the sender, so a person has to pick the customer.
A customer writing from an address that isn't on their contact card.
An order in a unit other than the one you stock.
An item that needs a size or color nobody named.
Don't decide from a demo. Demos use clean emails.
Pull 100 to 200 order emails from last month. Run them through the agent in a test copy of Business Central. Count the ones it finishes with no fixes. Sort the misses by cause.
Few misses, keep it and fix the data behind them.
Misses in one spot, like one big customer's PDF, solve that spot.
Misses everywhere, or most of the inbox is status questions and backorders, look at an order entry tool or a custom agent next to Business Central.
Keep people on new customers, credit holds, price exceptions and anything that changes fit or safety.
If you're weighing this for your order desk, DM me. Happy to look at your test with you.
What share of your order emails come in forwarded or as PDFs?
Full breakdown is in my profile.
2
1
69
Forward a customer's order email to Business Central's order agent and it stalls. It sees your rep as the sender, so a person has to pick the customer. How many of your order emails arrive forwarded?
19
Your software already checks every invoice against the order and the delivery.
The cost is in the ones that fail. On average 18.4% of invoices get stuck (Ardent Partners).
Put an AI agent on that pile. It finds the reason, asks the right person and runs the check again. Bank detail changes stay human.
What gets your invoices stuck most often?
Full breakdown is in my profile.
34
If you sell an AI automation pilot, agree on the finish line before you start. Pick one workflow, measure the time it takes today and count how often a person still has to step in. That gives the buyer something concrete to judge.
2
67
Before you pay for an AI security pilot, ask for one verified fix. Start with one repo. Reproduce the issue, review the patch and run the tests. Then compare the engineer time spent with your usual workflow.
2
1
112
107 chats. 16 memory tokens. Then the AI turned a question from an old chat into a fact.
If your agent remembers customers, steal these 3 tests before it acts on a bad memory.
Props to @adibvafa and the Memorilla team for sharing the demo.
xitter.gq/orbiresearch/status/21…











