Business Central 2026 Wave 1: What AI Agents Mean for Your UAT Process
Business Central 2026 wave 1 is rolling out now, and this one is different. Microsoft is shipping AI agents that don't just suggest actions. They take them. The Payables Agent reads your inbox, matches invoices to vendors, and prepares them for posting. The Sales Order Agent processes routine orders. The Fulfilment Agent handles picking and shipping. These are not Copilot prompts you click “Accept” on. They are autonomous processes running inside your ERP.
If you are an implementation consultant, a delivery manager, or a project manager running a BC go-live in 2026, this changes what your UAT process needs to cover. You are no longer just testing that a user can post a purchase invoice. You are testing that an AI agent posts the right purchase invoice, to the right vendor, coded to the right GL accounts, without anyone telling it to.
1. What's Actually Landing in 2026 Wave 1
Before getting into testing implications, it helps to be specific about what Microsoft is shipping. The 2026 wave 1 release (rolling out April to September 2026) includes several categories of change that affect how implementations are tested.
AI Agents
The Payables Agent, Sales Order Agent, Expense Management Agent, and Fulfilment Agent. These operate autonomously with human oversight via a new centralised task pane. They process transactions, match documents, and prepare postings without manual intervention.
Copilot Enhancements
Expanded Copilot capabilities across the platform, including item insights with advanced KPIs and summaries. Copilot is now included at no extra cost with every Business Central licence, which means every user on every implementation will have access to it.
Developer Tooling
Improvements to AL testing frameworks, debugging tools, Copilot extensibility, and agent design capabilities. Full-text search metadata on table fields. Public NuGet feed for downloadable symbols. Rule-based scheduling for performance profiling.
MCP Server and Agent Designer
An enhanced Model Context Protocol (MCP) server that lets external AI agents (including tools built in Copilot Studio, or third-party AI like Claude) read, create, and post documents directly in BC. Plus a new Agent Designer for building custom agents inside Business Central itself, and a Troubleshooting MCP Server for AL developers.
The important detail for implementation teams: AI agents and Copilot features are only available on Business Central Online. On-premises deployments do not get these capabilities. If your client is online (which most new implementations are), these features will be available in their environment whether you planned for them or not.
The MCP Server deserves special attention. It means BC is no longer a closed system where only the BC client and Power Automate flows touch the data. Third-party AI agents can now interact with your ERP through a standardised protocol. For consultants, this opens a new category of testing: what happens when an agent you did not build starts creating transactions in your client's system?
2. Why Agents Are Not Just Automation
BC consultants have been building automation for years. Power Automate flows, scheduled reports, approval workflows. Those are rule-based: if X happens, do Y. You configure the trigger, you define the action, and the system follows the rules. If it does something wrong, you can trace the issue back to a configuration decision.
AI agents are different in a way that matters for testing. The Payables Agent does not follow a fixed rule that says “match invoice line 1 to PO line 1.” It reads the invoice, interprets the content, identifies the vendor, matches it to the most likely purchase order, and codes it to GL accounts based on patterns. It makes judgement calls. And judgement calls can be wrong in ways that rule-based automation cannot.
Consider the difference:
Traditional Automation
- Follows configured rules exactly
- Fails predictably when rules don't match
- Errors are traceable to configuration
- Testing validates the rule works
AI Agents
- Interprets content and makes decisions
- Can produce plausible but incorrect results
- Errors may not be immediately obvious
- Testing must validate decision quality
This distinction is critical for anyone writing test plans. A traditional test case says “Post purchase invoice, expected result: invoice posted to GL.” An agent-aware test case needs to ask: “Did the agent match this invoice to the correct vendor? Did it pick the right GL account? What did it do with the ambiguous line item? Did it escalate the one it was not confident about?”
3. The UAT Gaps This Creates
Most ERP UAT processes were designed for a world where the system does what it is told. Here are the specific gaps that AI agents introduce.
Gap 1: No test coverage for agent decisions
Standard BC test plans cover posting transactions, running reports, and validating data flows. They do not cover whether an AI agent correctly interprets an ambiguous invoice, handles a vendor with multiple trading names, or codes an expense to the right cost centre. If your test plan does not explicitly cover agent behaviour, it will not be tested.
Gap 2: The “it looked right” problem
When an AI agent prepares a purchase invoice for posting, it will look correct. The vendor name will be right, the amounts will match, the GL coding will be plausible. The risk is that testers approve agent output because it looks reasonable, without checking whether the agent's decision was actually correct. This is especially dangerous for payables and expenses where incorrect coding may not surface until month-end close.
Gap 3: No boundary testing for agent confidence
Every AI agent has a confidence threshold. Below that threshold, it should escalate to a human. But what is that threshold? Does it behave correctly at the boundary? What does it do with invoices from a brand new vendor it has never seen? These edge cases are exactly the scenarios where agents are most likely to get it wrong, and they are exactly the scenarios most UAT plans skip.
Gap 4: Override and escalation paths untested
The new centralised task pane lets users review and override agent decisions. But has anyone tested whether overriding an agent decision actually stops the downstream process? If you reject the Payables Agent's invoice match, does it re-queue the invoice, or does it sit in limbo? Override paths are the safety net. If they do not work, the safety net has holes.
Testing AI Agents on Business Central?
Track Agent Decisions Alongside Traditional UAT
LogicHive gives you centralised test management with structured sign-off and full audit trails. Track every test, every decision, every sign-off in one place.
4. Testing Each Agent: What to Cover
Here is a practical breakdown of what UAT should cover for each of the first-party agents shipping in 2026 wave 1.
Payables Agent
The Payables Agent scans a shared mailbox for invoices, reads the content, matches vendors, matches purchase orders, and prepares invoices for posting. It is the most impactful agent for most BC implementations because nearly every client processes purchase invoices.
- Vendor matching accuracy: Send test invoices from vendors with similar names, vendors who trade under a different name than their legal entity, and vendors with multiple remittance addresses. Does the agent pick the right vendor card?
- PO matching: Test with invoices that partially match a PO (different quantities, different prices, additional lines). How does the agent handle discrepancies?
- GL coding: Validate that the agent codes to the correct GL accounts, especially for invoices without a PO where it must infer the coding. Check that it respects dimension rules and posting groups.
- Duplicate detection: Send the same invoice twice. Does the agent catch it, or does it create a duplicate?
- Escalation: Send an invoice the agent should not be able to process confidently (a new vendor, an unusual format, a foreign currency invoice). Verify it escalates to the task pane for human review.
Sales Order Agent
The Sales Order Agent processes routine sales orders, reducing manual data entry for high-volume, repetitive order patterns.
- Item resolution: Test with orders referencing items by description, partial item number, and customer-specific item references. Does the agent select the correct item?
- Pricing: Verify the agent applies the correct price list, customer-specific discounts, and campaign pricing. A plausible but incorrect price is worse than an obvious error because it may not get caught.
- Credit limits: Submit an order that would push the customer over their credit limit. Does the agent respect the limit, or does it process the order anyway?
- Location and warehouse rules: For clients using multiple locations or warehouse management, verify the agent selects the correct shipping location and respects warehouse policies.
Expense Management Agent
- Policy compliance: Submit expenses that violate company policy (over-limit meals, unapproved categories, missing receipts). Does the agent flag these or let them through?
- Duplicate claims: Submit the same expense receipt twice across different claim periods. Does the agent detect the duplication?
- Approval routing: Verify the agent routes to the correct approver based on amount, department, and expense type.
Fulfilment Agent
- Pick accuracy: Verify the agent generates correct pick instructions, especially for items with lot/serial tracking, multiple bin locations, or FEFO requirements.
- Partial fulfilment: Test with orders where stock is insufficient. Does the agent handle backorders correctly? Does it split shipments appropriately?
- Shipping method selection: For clients with multiple carriers or shipping rules, verify the agent selects the correct method based on order weight, destination, and service level.
Custom Agents via MCP Server and Agent Designer
Beyond the first-party agents, 2026 wave 1 opens the door for custom agents. The enhanced MCP Server lets external AI tools interact with BC data through a standardised protocol. The Agent Designer lets users build agents directly inside Business Central. If your client plans to use either of these, your UAT scope expands significantly.
- Permission boundaries: What can the custom agent access? If it has permission to post purchase invoices, can it also post credit memos? Test that MCP Server permissions are configured to the minimum necessary scope. An agent with overly broad access is a risk that UAT should catch before go-live.
- Data integrity: If a third-party agent creates a sales order via MCP, does it populate all mandatory fields? Does it respect number series, posting groups, and dimension rules? Or does it create technically valid but business-invalid records?
- Audit trail: When an external agent posts a transaction, what does the audit trail show? Can your client's finance team identify which transactions were created by agents versus humans? This matters for compliance and for debugging when something goes wrong.
- Regression after updates: When BC updates (which happens monthly for online tenants), do custom agents still work correctly? The MCP Server interface may change between versions. Build regression test cases that validate agent connectivity and behaviour after each update.
5. Copilot Features Still Need UAT
Alongside the autonomous agents, Copilot in Business Central continues to expand. Every BC Online user now has access to Copilot at no additional cost. This is worth calling out because it means Copilot features will be active in every implementation, regardless of whether you planned for them.
Copilot features that need UAT attention:
- Bank reconciliation suggestions: Copilot suggests matches between bank statement lines and ledger entries. Test with ambiguous matches (similar amounts, same-day transactions to different vendors) to validate the suggestions are trustworthy.
- Item descriptions and marketing text: Copilot can generate product descriptions. If your client uses these in customer-facing documents or e-commerce integrations, the generated text needs review as part of UAT.
- Analysis assist: Copilot helps users build ad-hoc analysis views. Validate that the data returned is accurate, especially for financial reporting where an incorrect filter could produce misleading numbers.
- Chat with Copilot: Users can ask Copilot questions about their data. Test that it returns accurate answers and does not hallucinate figures, especially for questions about balances, stock levels, and order status.
The risk with Copilot is subtler than with agents. Agents take action. Copilot influences decisions. A finance manager who trusts a Copilot-suggested bank reconciliation match without checking it is making a decision based on AI output. That decision needs to be made with an understanding of when Copilot gets it right and when it does not.
6. A Practical Approach for Consultants
If you are building a test plan for a BC implementation going live on 2026 wave 1, here is how to approach the AI and agent features without overcomplicating your existing process.
Step 1: Decide which agents to enable
Not every client needs every agent on day one. Have the conversation early about which agents are in scope for go-live and which will be enabled later. An agent that is not enabled does not need UAT. An agent that is enabled without UAT is a risk.
Step 2: Add agent-specific test cases to your existing plan
You do not need a separate test plan for agents. Add test cases to the relevant process area. Payables Agent tests sit in your purchase-to-pay cycle. Sales Order Agent tests sit in your order-to-cash cycle. Test the agent as part of the business process it supports, not as a standalone feature.
Step 3: Focus on edge cases, not happy paths
The agent will probably handle a clean, standard invoice from a known vendor correctly. Your testing time is better spent on the exceptions: new vendors, partial matches, unusual formats, foreign currencies, and items the agent has not seen before. These are the scenarios where agent errors are most likely and most costly.
Step 4: Test the task pane and override workflow
The centralised task pane is where users review agent actions and handle escalations. Make sure your testers know it exists, know how to use it, and have tested the full override workflow: reject an agent decision, correct it, and confirm the correction flows through correctly.
Step 5: Document agent behaviour as part of sign-off
When users sign off on a process area that includes an agent, make sure the sign-off explicitly covers agent behaviour. “I confirm the purchase-to-pay process works” is not the same as “I confirm the Payables Agent correctly matches invoices from our top 20 vendors and escalates anything it is not confident about.” Be specific.
For consultants managing UAT across multiple BC projects, the good news is that agent test cases are reusable. Once you have built a solid set of Payables Agent test scenarios, you can adapt them for every client. The vendor names and GL accounts change, but the test patterns stay the same.
Running UAT on Business Central?
Stop Managing Test Cases in Spreadsheets
LogicHive gives BC implementation teams centralised test management, real-time progress tracking, and structured sign-off. Built for the way consultants actually work.
Frequently Asked Questions
What are the new AI agents in Business Central 2026 wave 1?
Business Central 2026 wave 1 introduces several AI-powered agents including the Payables Agent (automates invoice matching and payment preparation), the Sales Order Agent (handles routine order processing), the Expense Management Agent (processes expense claims), and the Fulfilment Agent (manages order picking and shipping). These agents can take action autonomously with human oversight through a central task pane.
How do AI agents in Business Central affect UAT?
AI agents change UAT fundamentally because you are no longer just testing whether the system follows a process correctly. You are testing whether the system makes acceptable decisions on its own. UAT must now cover agent decision boundaries, escalation paths, override mechanisms, and what happens when the agent encounters data it has not seen before. Traditional pass/fail test cases need to be supplemented with scenario-based validation of agent behaviour.
Should we test Copilot features during Business Central UAT?
Yes. Copilot features should be tested with your actual data in a sandbox environment, not assumed to work because they are standard functionality. Test that Copilot suggestions are accurate for your chart of accounts, your item catalogue, and your business processes. Validate that users understand when they are acting on a Copilot suggestion versus their own judgement, and confirm that incorrect suggestions can be easily overridden.
Are Business Central AI agents available on-premises?
No. Copilot and AI agent features are only available to Business Central Online customers. On-premises deployments do not have access to these capabilities. If your client is on-premises, AI agents are not part of your UAT scope, but this may become a factor in cloud migration conversations.
What is the MCP Server in Business Central and why does it matter for testing?
The Model Context Protocol (MCP) Server in Business Central allows external AI agents to read, create, edit, and post documents directly in your ERP through a standardised interface. This matters for testing because it means transactions can now be created by tools outside of Business Central itself. UAT must validate that MCP-connected agents respect permission boundaries, populate mandatory fields correctly, follow posting rules, and produce clear audit trails.
What is the timeline for Business Central 2026 wave 1?
Business Central 2026 wave 1 features begin rolling out in April 2026, with general availability of planned features through September 2026. Early access for some features started in February 2026. Implementation teams should plan sandbox updates and UAT cycles accordingly.