AI
-
Business Central 28.2: everything that shipped and the line you shouldn't skip
A complete walkthrough of Business Central update 28.2: Copilot and agents, financials, e-documents, the new auditor APIs, quality management, sustainability. Plus the one good-to-know line that quietly changes data residency: from July 1, 2026, Copilot and agent prompts can be processed in another Azure geography.
-
Claude Fable 5: What the New Top Tier Changes for Your Code
Anthropic's Fable 5 opened a new model tier above Opus, then the US government pulled it four days after launch. Twice the price, a much slower first token and a lead that grows with task length: why Opus 4.8 stays the better daily driver, what breaks in your code and why a model vanishing overnight is the best argument for routing through a gateway.
-
Your Business Central MCP server should be read-only until it can't be
The Business Central MCP server is GA-adjacent and the pitch writes itself: connect Claude or Copilot Studio, read and write ERP data in natural language, no custom API. The connection is the boring part. You've just handed a promptable, nondeterministic actor a write path to your ledger and the only thing standing between it and a posted document is how you configured it.
-
When NOT to use an AI agent in Business Central
The built-in agents are GA and every partner is asking 'can an agent do this?' That's the wrong question. For a large class of BC work, a deterministic codeunit beats an agent on cost, latency and auditability and picking the agent anyway is a downgrade dressed as innovation.
-
Your AI feature must run on a fresh tenant, or it doesn't run
The line between a demo and a product is whether it works on a clean Business Central environment with none of your developer setup. Five concrete things that separate 'works on my box' from 'a customer can actually use it'.
-
Shipping a Copilot feature in Business Central that survives real users
A PromptDialog demo takes an afternoon. A Copilot feature real users won't switch off takes the other 90%: scoping, grounding, the generate, review, accept loop and treating the model's output as a proposal that never silently touches data.
-
A 502 in your PDF-to-LLM pipeline is the gateway, not the model
Business Central sends a PDF through an AI gateway and gets a flat 502. The instinct is to blame the prompt. Don't, the model never saw your document. Here's how to bisect the pipeline and force a text-extraction path that works.
-
Route, don't call: why a model gateway belongs between you and the LLM
Calling a provider SDK directly is the fastest way to start and the slowest way to scale. A thin gateway turns the model into config and buys you fallback, cost control and observability you'll need the week after launch.
-
Calling AI from Business Central: the platform realities nobody warns you about
Your AL HttpClient call to an AI endpoint returns send=false and no error. It's not your code, it's three platform guardrails (the HttpClient flag, the anti-SSRF allowlist and loopback hostnames) that fail silently. Here's the checklist.
-
Treat LLM output as untrusted input
The single mental shift that prevents a whole class of AI bugs and vulnerabilities: a model's response is not a value you computed, it's user input from a stranger and you already know how to handle that.
-
Two crafts, one blog: why I write about AI and Business Central
I do two jobs that are quietly merging, ERP engineering on Business Central and building with LLMs. This is the notebook where I work out loud on both and on the seam between them.
