Microsoft’s cloud migration tool moves data. It does not move code. Most of what makes a move to Business Central online difficult follows from those two sentences.
The tool itself is fine: it replicates your on-premises tables into the online tenant through Azure Data Factory, table by table, tracking success and failure per table. What it cannot do is carry a C/AL modification across. Business Central online runs extensions only, and data from tables with code customisations can’t be carried forward unless those customisations are handled by extensions installed on both sides. The application layer has to exist online before the data has anywhere to land.
Microsoft documents two paths out of a modified on-premises database: a full migration, where every modification becomes AL and the history follows, or a reimplementation, which takes master data, opening balances, open documents and setup and leaves the rest. The version 14 reimplementation tool — still a preview — covers the standard tables and says plainly that legacy C/AL customisations and their extension data aren’t migrated unless you extend it.
Many companies sit closer to the second than the first, whatever the proposal calls them. So the valuable early work is not writing AL — it is deciding what not to bring. Three questions carry that decision.
Data: what has to be in the system
The distinction that matters is between data that has to be in Business Central and data that only has to be findable. Open balances, unapplied entries, inventory, open orders, the master records those point at, and enough posted history for your reporting and audit obligations belong in the system. The invoice from twelve years ago does not need to live in a posting engine.
That distinction has a price. A tenant comes with 80 GB of database capacity plus an allowance per licensed Essentials or Premium user, shared across every environment, production and sandboxes together. Online data is compressed at page level, so the on-premises size predicts little; Microsoft publishes a method for estimating the compressed size before you commit.
Ledger history is the biggest line item and the one worth arguing about. Bringing every posted G/L, VAT and item ledger entry across is possible, and every future upgrade, environment copy and restore then carries them. Ask what anyone would genuinely ask of entries older than the statutory retention window, and whether the answer has to come from an ERP.
There is a supported way to keep history without keeping it live. Date compression batch jobs summarise old ledger entries, and the Data Archive extension writes what they delete into an archive you can export to Excel or CSV — 10,000 records each, held in Tenant Media, so the pattern is export and clear. Retention policies handle the other end, giving log tables and archived document versions an expiry date and a job queue entry to apply it.
Attachments are usually the surprise. Scanned invoices and PDFs sit in the database by default and grow unwatched. Business Central can keep them outside the database in Azure Blob Storage, an Azure file share or a SharePoint library — though files kept there are yours to back up, not Microsoft’s.
Some things simply don’t come across. Cloud migration doesn’t move most system tables, users or permissions, and record links aren’t migrated at all because they’re tied to a user ID. Number series are copied only for the companies under replication. Microsoft also advises fewer companies per run, and moving one into an already-live tenant isn’t supported — so in a group, batching is a design decision, not a scheduling one.
Integrations: what survives the move
Sort every integration into one of three piles.
Already speaking the right protocol. OData V4 and REST API calls over HTTPS carry over with a base URL change and an app registration. What changed underneath is authentication: access keys and Basic auth stopped working online on 1 October 2022, so anything holding a web service access key needs an app registration and OAuth. SOAP still works for your own pages but is on its way out — since version 29.0 (October 2026) Microsoft UI pages can no longer be published as SOAP endpoints at all.
Needing a rebuild around an API page. A page published as an OData web service is a UI page with an endpoint attached, and UI pages change between releases — from Microsoft, or from any extension that modifies them. That is why REST API pages are the recommended approach, quite apart from being faster and exposing fields the page never showed. From version 30 (2027 release wave 1) Microsoft pages can’t be exposed as OData endpoints at all, so an integration on a published Microsoft page has a deadline.
Only ever possible because the server was yours. This is the pile that costs money.
- Anything reading SQL directly — Excel over ODBC, an SSRS report, a nightly job selecting from
G/L Entry. Online there is no database connection to hand out. These become API consumers, or move to Power BI over API pages, or Microsoft Fabric and OneLake for heavier work. - .NET interoperability. DotNet types require
"target": "OnPrem"and are unavailable online. The documented replacement, Azure Functions called from AL, means a subscription, a deployment story and a named owner for what was a DLL in a folder. - File-system integrations. A job queue entry that reads a folder on a server has no folder to read; Azure Blob Storage, an Azure file share through the Azure File Services API module, or SharePoint take its place.
- Printing and mail from the service tier. Printing goes through Universal Print, which needs its own subscription, or an email-enabled printer; otherwise the browser handles it. Mail goes through the Microsoft 365 connector or SMTP with OAuth 2.0.
Then there are the limits on-premises hid, because there they were configurable. OData requests time out at eight minutes and the gateway at ten, a request returns at most 20,000 entities, a $batch holds 100 operations, an environment supports 200 webhook subscriptions. A client that doesn’t handle 429, 503 and 504 with backoff will pass testing and fail on a busy Monday.
Customisations: how much of the list is still a customisation
This is where re-implementation earns its keep. Every entry on the list was written against a particular version, and the product has not stood still. Sort it into four.
Standard now. E-invoicing is the clearest case in Belgium: the E-Documents framework arrived in 2023 release wave 2, Business Central supports PEPPOL for Belgium as required for the B2B mandate in force since 1 January 2026, and connectors for several access points ship with the product. Retention policies, external attachment storage, approval workflows, item references and the Shopify connector are the same story: someone’s modification once, base functionality now.
Configuration, not code. A real share of old modifications exist because a setting didn’t exist yet, or because nobody knew it did. The Belgian localisation alone covers posting period restrictions, mandatory journal templates, enterprise and branch numbers and multiple VAT registrations without a line of AL.
Genuinely yours. The routines that encode how your business actually works, with no equivalent in the product. These get rebuilt as proper extensions — event subscribers, no base app modification, clean install and upgrade codeunits — and as per-tenant extensions the source stays the customer’s. The list is usually shorter than anyone expects.
Dead. Modifications for a process that stopped, a report nobody opens, an integration to a system replaced two suppliers ago. Nobody removes these on-premises, because it costs effort and changes nothing. A re-implementation is the one moment when not rebuilding something is free.
Two constraints while sorting. Runtime packages aren’t allowed online, so a modification whose source is lost is a rewrite either way. And AL objects not covered by an entitlement won’t be accessible in the cloud, which the compiler tells you long before a user does.
The order the questions go in
Customisations first, because they determine which tables can be migrated at all. Then data, because the volume decision depends on which tables exist. Then integrations, because their endpoints depend on the objects the first two produce. And at least one full dry run into a sandbox before cutover — Microsoft’s guidance, and the only honest way to learn how long the real one takes.
When we’re asked for a second opinion on a proposal, this is the part we read first: not the day rate, but whether anyone has written down what is being left behind. A quote that migrates everything is not a cautious quote. It is an unmade decision, priced.