Sales Order Agent vs. Payables Agent vs. Expense Agent — Which One Actually Earns Its Keep?
Every customer conversation I’ve had this year eventually arrives at the same question, phrased in slightly different ways: “We’ve heard Business Central has agents now. Should we turn them on?”
That’s not one question. It’s three, and they have three different answers. The Sales Order Agent, the Payables Agent, and the Expense Agent shipped at different times, sit at different maturity levels, hit different license and geography walls, and cost wildly different amounts per unit of work. Recommending “the agents” as a bundle is how you end up in an uncomfortable meeting six weeks later.
So here they are next to each other, with the honest version of each. Challenge, what the agent actually does, and what you’re really left with.
Sales Order Agent
The challenge. A shared mailbox that receives customer requests in prose. “Could you quote us 2 Antwerp tables and 4 Berlin chairs, delivery by March 2nd, our PO is DC1011234.” Someone reads it, hunts for the items, checks stock, builds a quote, exports the PDF, writes the reply. Ten minutes if the customer was clear, considerably longer if they weren’t. Repeat all morning.
What the agent does. It monitors that mailbox, identifies the customer from the sender address, extracts items and quantities — including from PDF or image attachments — checks availability, creates the quote, and drafts the reply email with the PDF attached. If details are missing it runs a multi-turn email conversation with the customer to clarify. Once the customer confirms, it converts the quote to an order. A designated reviewer confirms every outgoing message before it leaves the building.
The architecturally interesting part: the agent doesn’t call APIs. It starts at a Role Center and navigates pages, invokes actions, and enters data the way a user would, using UI metadata like captions and tooltips to work out its next step. Nothing is hardcoded. Which means it can discover and use your custom fields and actions — and can even read a validation error message and try to fix its own input.
The reality check. That same design is the limitation. The agent finds items across the item master, variants, references, attributes, categories, translations, and extended text — and Microsoft says plainly that its effectiveness depends on how good that data is. Vague product descriptions produce a vague agent. In practice, the honest scoping question isn’t “can the agent do this,” it’s “how clean is your item master?” That’s a data-quality project wearing an AI costume, and it’s the one nobody budgets for.
Item availability, including capable-to-promise, is still flagged preview even though the agent itself is generally available. And the credit math is public: a typical flow runs 2 credits to analyse the email, 5 to check availability, 5 to create the quote, and 2 to generate the response. Microsoft’s own worked example lands at roughly 1,650 Copilot Credits per month for 100 emails.
One operational trap worth knowing: when credits run out, the agent stops processing but stays active and scheduled. When credits renew, it wakes up and processes the entire accumulated backlog — potentially burning through the new allocation immediately. If you’re piloting on free credits, deactivate the agent before you hit zero.
Verdict: the most impressive of the three and the most dependent on your customer’s housekeeping. Best fit: repetitive B2B reorders from known customers with a well-maintained catalogue.
Payables Agent
The challenge. Vendor invoices arrive as PDFs in a mailbox. Someone opens each one, keys the vendor, invoice number, date, and lines into a purchase invoice, and routes it for approval. It’s the highest-volume, lowest-judgement task in most finance departments.
What the agent does. It monitors a designated mailbox, imports each PDF into Inbound E-Documents — one entry per attachment, so the audit trail is intact from the start — extracts the data, matches the vendor, suggests account classifications, and produces a draft purchase invoice for review. It runs under its own agent user identity, so every action is attributed and traceable. When it can’t identify a vendor with confidence, it stops and asks. Vendor creation and unusual account assignments are escalated for mandatory human approval. Since version 27.3 it also handles purchase order matching, including three-way matching, which is what finally made it relevant for goods-based invoicing rather than just utilities and subscriptions.
The reality check. This one has the most published limits, and I mean that as a compliment — you can actually scope it. It only processes emails with PDF attachments. It skips emails carrying more than 10 attachments. It handles up to 100 emails per day and 50 in a single batch. Microsoft is also explicit that vendor matches and account classifications can be wrong and that the agent can’t evaluate business context the way a person can.
The dependency to flag in a scoping call is historical data: account suggestions improve with transaction history, so a customer fresh off a migration gets noticeably less value than one with three years of posted invoices behind them. That’s an awkward thing to discover after go-live.
Credits scale with document complexity rather than document count: roughly 50 credits to process an invoice and produce the draft, plus about 5 credits per invoice line. A ten-line invoice therefore costs meaningfully more than a one-line one — which is the opposite of how most people instinctively budget.
Verdict: the most boring and probably the best ROI of the three. Highest volume, clearest limits, most predictable review burden. If a customer wants exactly one agent to prove the concept, this is usually it.
Expense Agent
The challenge. Expenses are a small process that everyone hates. Receipts pile up in wallets, submissions arrive weeks late, finance chases missing documentation, and the people submitting are field engineers and drivers who have no business owning an ERP licence.
What the agent does. Employees submit through a dedicated web app, by forwarding a receipt to a monitored mailbox, or through Copilot Chat — and critically, the submitter doesn’t need a Business Central licence. The agent extracts merchant, date, amount, currency, and payment method; categorises against your configured categories; applies policy rules; itemises hotel receipts; calculates per diem from an uploaded itinerary; handles mileage; and flags duplicates. Supported formats are JPEG, PNG, and PDF. It’s paired with a new Expense Management module that supplies the underlying structure — categories, posting groups, approval workflow — so posting runs through your existing financial controls, not around them. A mobile app for iOS and Android was expected on the stores in July 2026.
The reality check. The geography story moved faster than most of us tracked. The agent launched English-only and US-only in May, which is where a lot of the commentary still sits — but the July wave added German, Danish, French, and Spanish, covering Australia, Canada, Denmark, France, Germany, New Zealand, Spain, and the UK. Localisation goes beyond the interface: receipt submission, expense review, report generation, and approval all run in the local language, on web and mobile.
So for a German customer, this is no longer a “watch this” item. It’s a “you can pilot this now” item.
What it is not is GA. The release plan lists public preview for July 2026 with no general availability date attached, and carries the standard caveat that timelines can move. That’s a fine basis for a pilot and a bad basis for a fixed-price rollout — and the distinction is worth saying out loud in the meeting, because “available in German” and “ready to commit to” are not the same sentence.
Cost is the simplest of the three and the easiest to model: 50 Copilot Credits per uploaded receipt, regardless of how many lines the resulting expense has. So a company with 40 travellers submitting 15 receipts a month is looking at 30,000 credits monthly. Do that arithmetic before the demo, not after.
Worth noting for anyone with a strict AI governance process: community reporting indicates the Expense Agent runs on Anthropic models and requires Anthropic to be explicitly enabled as an AI subprocessor in the Microsoft 365 admin centre. Verify that against your customer’s approved-subprocessor list before you promise a timeline.
Verdict: the most genuinely new capability — Business Central simply didn’t have native expense management before — and, since July, the one whose availability story changed most. Pilot-ready in German. Not yet something to sign a fixed scope against.
Where I’d Actually Start
| Sales Order Agent | Payables Agent | Expense Agent | |
|---|---|---|---|
| Status | GA (availability check still preview) | GA (PO matching added in 27.3) | Public preview |
| Geography | Broad | Broad | US, plus DE/DK/FR/ES/UK/AU/NZ/CA from July 2026 |
| Main dependency | Item master data quality | Historical transaction data | Preview status, not geography |
| Credit driver | Per email processed | Per invoice + per line | Per receipt (50) |
| Review burden | High — every outbound email | Medium — every draft | Low — exceptions only |
Three things cut across all of them.
They’re SaaS-only. On-premises and partner-hosted customers can’t have them, full stop. For some of your install base, this conversation is actually a cloud migration conversation.
Nothing here is autonomous. All three are human-in-the-loop by design. The work doesn’t disappear — it converts from data entry into review. That’s a real gain, because reviewing a populated draft is genuinely faster than keying one from scratch. But if you size the business case as “we eliminate the AP clerk,” you’ve mis-sold it. Size it as “the same team handles three times the volume.”
Consumption billing changes the shape of the project. These agents don’t sit inside the BC subscription; they draw Copilot Credits, and credits are only visible to users with SUPER or the AGENT – DIAGNOSTICS permission set. Assign that to someone with a budget interest on day one.
If a customer asks me where to start in production, I say Payables. Highest volume, clearest boundaries, most predictable cost per unit, and the review workload is easy to size honestly. Sales Order Agent second, once you’ve been through their item master with your eyes open.
Expense Agent is the one I’d put in a pilot this quarter. It’s the capability I’d most like to deploy, the German language support just landed, and preview is exactly the right place to find out how your customer’s policy rules survive contact with an AI that categorises receipts. Just don’t let “it works in German now” quietly become “it’s ready” in someone’s project plan.
The agents are real. The hype around them is doing them a disservice, because the sober version — draft generation with mandatory review, priced per action — is a good enough product that it doesn’t need the oversell.
Discover more from Lubenheimer
Subscribe to get the latest posts sent to your email.