Microsoft presents: Behind the scenes of how we build the new Expense Agent

Overview

The session from Microsoft’s Expense Agent team covers the agent from three angles. First it shows what the agent does: it reads receipts, itemizes and categorizes them, handles several languages and currencies, and calculates mileage and per diem. Expenses reach it by shared mailbox or web app, with a mobile client coming later. Business Central remains the backend where data is stored.

It then explains how the team built it. The team worked AI-first, with 99 percent plus of the code generated and everything kept in one repository. Quality was managed with evals: 1,500 mostly hand-labeled datasets, a 95 percent accuracy requirement, a hill climbing loop, model comparison, red-team testing and online LLM-judge evals. The last part walks through the Azure service architecture, the Business Central setup (expense users, categories, rules, consumption page, reports, posting) and the current limits.

Key points

• The Expense Agent is in public preview and available in the US only. More regions were said to come the following month.
• Expense users are a new Business Central entity attached to employees. They need no BC license and can submit by web app or shared mailbox. A mobile client is coming. Existing employees can be imported in one action.
• Limits mentioned: 10 MB per expense file, 25 files and 50 MB per batch, 100 emails per company in the shared mailbox scenario, and 50 expenses per expense user per company per day. These limits are hardcoded for now.
• Category and subcategory descriptions act as prompts for the agent, so their detail affects categorization accuracy. Rules cover amounts, thresholds and required information. Natural-language policies are not yet available.
• Some steps are still manual, such as setting mileage information and posting. For now, projects are attached to expense lines or reports in Business Central. Attaching them from the web app is work in progress.
• Microsoft requires 95 percent accuracy to ship an agentic system. The team started at 65 percent and used 1,500 labeled datasets, synthetic edge cases, a hill climbing loop and model comparison to improve it.
• The service uses Azure Front Door, a Python gateway for authentication, and Azure Functions processors. Redis caching limits repeated lookups when locating a user’s Business Central environment.

Features

A status with a roadmap link comes from the Microsoft 365 roadmap feature this part of the video covers (matched by Haiku; links Opus dropped are not used); other statuses need a status word in the video itself.

AL objects mentioned

As heard in the captions. A name that matches one object page by exact type and name links to it; the others stay as named.

• other “Opus 4.5” at 18:05
• other “GitHub Copilot CLI” at 18:23
• other “LLM judge” at 24:20
• other “Azure Front Door” at 39:38
• other “Gateway” at 40:22
• other “Sales Order Agent” at 51:06
• other “Payable Agent” at 51:06
• other “Expense Agent” at 51:06
• other “Graph API” at 52:58
• page 6910 “Expense Report” at 1:06:02
• page “expense agent setup page” at 1:04:37
• page “agent consumption page” at 1:04:59

Not found in BC28-30: page “expense agent setup page”, page “agent consumption page”.

Quotes

• 0:46 “It’s still in preview will be for a little while. uh very important that we get the the feedback back.”
• 0:56 “this expense agent is not a conventional business central agent”
• 1:40 “we wanted nonBC users I mean all your employees don’t you wouldn’t like to have a to buy a license for all the employees”
• 1:58 “we are still using BC as the back end. It’s where all data goes. It doesn’t go anywhere else really.”
• 5:31 “this actually works remarkably well and uh and uh actually cuts out a lot of humans errors”

Disclaimers in the video

• 0:46 preview: It’s still in preview will be for a little while
• 19:52 preview: We’re even working on like truly policy validation from written policies. We’re not quite there yet.
• 33:58 other: the 1,500 data sets we have maybe is not enough because the as I said the universe of uh of …
• 37:12 other: there is a session in couple of hours from my colleagues Claus and Cloudu which they will talk more in …
• 45:51 subject-to-change: if you feel that this doesn’t fit your need or maybe you want them configurable because now are hardcoded please …
• 48:57 preview: as we speak right now uh we are in public preview
• 49:13 subject-to-change: there are few things that you still have to do manually. For example, setting the uh for example the mileage …
• 49:33 coming-later: this is our goal uh on our speedboard uh we want to do in in coming months
• 51:53 coming-later: pretty soon we will also have the mobile client as well
• 1:03:01 coming-later: we are working for the policies as well uh which will come uh pretty soon
• 1:08:45 coming-later: pretty soon uh the expense users will also be able to attach a project from the web app as well. …
• 1:10:05 preview: this is a new and still in public preview
• 1:12:00 coming-later: we are in US only. More regions will come next month

Presenters (as heard): Christian, Casim, Michael, Cyprin, Claus, Cloudu, Sepan, Kasim.