The public preview for Business Central version 29 has been available for a few days now, and there’s already a lot to go through a long list of new capabilities across development, agents, reporting, and supply chain. Most of it is worth knowing about individually.
Having gone through the full preview list, I believe the meaningful changes cluster into four themes
Theme One: The Platform Is Forcing a Deadline, Not Suggesting One
SOAP web service support is permanently removed in v29, arriving with general availability in October. Not deprecated removed, on a fixed date, with no fallback. Any integration still depending on a SOAP endpoint stops working the day this update lands, and the organizations most exposed are the ones who don’t know they still have one running.
Subcontracting is getting the same kind of treatment in the other direction not a removal, but a full overhaul, with assisted setup, routing instructions carried into purchase orders, warehouse put-away and pick support, and dedicated migration tooling for Italian legacy subcontracting data. It signals that Microsoft is willing to invest heavily in modernizing an entire workflow at once rather than shipping incremental patches to it indefinitely.
Both of these say the same thing from opposite directions: Microsoft is no longer treating architectural change as optional or gradual. Some things get a hard deadline. Others get a complete rebuild. Neither leaves room for partners to wait and see.
Theme Two: AI Agents Are Taking Over Judgment Work, Not Just Generative Work
AI agents can now debug a recorded Business Central failure directly describe the scenario, record the matching session, and hand the agent an offline snapshot to diagnose. The Expense Agent now supports a full approval process, with AI-based policy validation flagging suspicious or noncompliant items across lines and reports rather than simply capturing receipts. Shopify’s new Tax Matching Agent reviews suggested tax jurisdictions and confidence levels before a sales document is processed with sales tax.
None of these are code-generation or content-generation features. They are judgment and diagnostic tasks debugging, compliance review, tax classification that used to require a person evaluating the situation directly. It can flag a noncompliant expense line, but who is expected to catch the pattern it misses? It can suggest a tax jurisdiction with a confidence score, but who decides what confidence threshold is actually acceptable for a given customer? The agents are taking on the work. The judgment about when to trust that work is quietly becoming the more important skill.
Theme Three: AI Agents Are Also Taking Over the AL Developer’s Operational Work
Earlier this year I argued that AI coding agents would not replace the AL developer, but would increasingly participate in the development lifecycle itself. The development tooling section of v29 is the clearest evidence yet of how far that participation has gone, and it is easy to miss because it reads like routine platform maintenance rather than a headline feature.
An AI agent can now profile a slow Business Central session directly, collecting CPU and memory data and summarizing SQL and HTTP activity without a developer manually attaching a profiler. An AI agent can now allocate free AL object IDs across project ranges, a small task, but one that used to require a developer to manually check ranges and avoid collisions every time. It can query connected Business Central environments to discover installed objects and owning-app metadata, but that capability exists primarily to let an agent add dependencies and resolve symbols without a developer walking it through the project structure by hand. It can run a standalone AL language server outside Visual Studio Code entirely, which only matters if something other than a human sitting in VS Code needs completion, navigation, and rename capabilities.
Underneath this, the platform itself is being restructured to support it. The new table extension data model stores all fields base table and extensions in the same underlying table, improving performance on every database operation involving them, and finally allowing indexes that span both. MCP Server can now run custom data queries against data that has no existing API. A new AL test framework supports parameterized, data-driven test suites with full CI/CD integration through AL Tool. Each of these, read individually, looks like a developer-experience improvement. Read together, they describe infrastructure being built for a consumer that is not a person typing in an editor.
Theme Four: Governance Is Finally Catching Up to Capability
The least exciting features in this release may be the most important ones. A new review bar now surfaces documents and lists where an agent requires human interaction before proceeding. Agent permissions can now be granted individually as agents request them, rather than through a blanket setup applied once and forgotten.
These features exist because the first three themes created a problem that had to be solved eventually: once agents are debugging sessions, approving expenses, classifying tax jurisdictions, and allocating object IDs, someone has to be able to see where an agent got stuck, and control precisely what it’s allowed to touch. This is the same pattern I described with AI code review effort levels a few weeks ago: capability arrives first, and governance arrives once organizations have enough exposure to realize they need it. Version 29 is the point where Business Central’s own platform team is visibly building that governance layer into the product itself, rather than leaving it entirely to partners to bolt on afterward.
Finally
Most of what will be written about BC v29 in the coming weeks will treat it as a features list — evaluate each item, decide what to enable, move on. I think the more accurate way to read this release is as four simultaneous signals: some architectural change is no longer optional, agents are taking over judgment work that used to require a person, agents are taking over operational development work that used to require a developer, and the platform is finally building the governance layer that all of that requires.
The AL developer’s job was never simply writing code, and this release confirms that the mechanical parts of the job debugging, profiling, ID allocation, symbol discovery, dependency resolution are the parts leaving fastest. What remains is judgment: deciding whether an agent’s diagnosis is correct, whether its permission request is reasonable, whether its architectural choice belongs in this specific Business Central environment. That is not a smaller job than the one Business Central partners and developers had a year ago. It is a different one, and version 29 is where the platform itself started being built to assume that shift has already happened.
Leave a Reply