Copilot in Business Central Applied to the Tasks That Repay It for a 900-employee Manufacturer
A manufacturer switched on every Copilot feature in Business Central and measured no improvement. We fixed the master data underneath it, narrowed the scope to three high-volume tasks, and got 63% of purchase invoices matching without human coding.
About the Client
The Challenge
The accounts payable team had stopped clicking the Copilot buttons. Not policy, and not objection to the idea — the suggestions were wrong often enough that checking them took longer than doing the work. A purchase invoice line for a machined component matched to the wrong purchase order. A bank statement line assigned to a general ledger account that made no sense. After a few weeks the team went back to keying.
The cause was in the master data. The item master held 21,000 items, many with descriptions that were a bare part number, an abbreviation, or the same text across four variants separated only by a suffix. The vendor list had accumulated duplicates over years of acquisitions — the same supplier under three names and two sets of payment terms. A block of general ledger accounts still carried names inherited from a previous system.
Copilot in Business Central grounds its suggestions on that data. A sales line suggestion is built from item descriptions and history. An e-document match compares document lines to purchase order lines on the descriptions, quantities and prices that are there. A bank reconciliation suggestion proposes a G/L account from what the account is called and where similar transactions went before.
There was also a scope problem. Every capability had been enabled at once, including marketing text generation for items, at a manufacturer that publishes no product copy. With no baseline and no view of which features were used, there was no way to tell whether Copilot helped anywhere — and the impression that it did not came from features that were never going to matter.
Our Approach
We did the unglamorous part first. Copilot in Business Central is a thin layer over your master data, and no prompt work compensates for an item master nobody maintained.
Baseline and Feature Audit (Weeks 1-3)
Measured what the tasks actually cost before touching anything: hours per bank reconciliation, time from invoice receipt to posting, the proportion of purchase invoices requiring manual coding, and exceptions per run. Then sorted the capabilities into generally available and preview, because the two carry different commitments. Bank account reconciliation assist, sales line suggestions and Chat with Copilot are GA, as are the autonomous Payables, Sales Order and Expense agents. E-document matching, analysis assist, summarise and autofill are preview. Everything without a business case was switched off, including marketing text generation. Availability is also gated by country and language, so we read the region matrix rather than assuming, and recorded the data movement consent settings — a compliance conversation that belongs before the feature conversation, not after.
Master Data Remediation (Weeks 2-9)
This phase was most of the work. Rewrote item descriptions to read as descriptions rather than codes, taking the 21,000 items in transaction-volume order so effort followed value. Merged duplicates across the 680 vendor records and standardised payment terms. Renamed general ledger accounts to describe what they hold. Set default dimensions and posting groups consistently, because a suggestion that gets the account right and the dimension wrong still needs a human. Added text-to-account mapping rules for recurring vendor lines with a fixed pattern.
E-Documents and Incoming Document Matching (Weeks 8-14)
Configured the e-document framework for suppliers able to send structured documents, and incoming document capture with OCR for the rest. Copilot proposes the match between e-document lines and open purchase order lines, and the clerk confirms or corrects. This is the one capability in the deployment still in preview, so it went in with a named owner, a rollback plan and no dependency in the month-end timetable. We wrote our own review rule rather than relying on a published cutoff, because none exists: a partial match or a tolerance breach sends the document to a person with nothing pre-filled. A plausible wrong suggestion costs more than no suggestion. Reached 63% of 4,200 purchase invoices a month matched without manual coding.
Bank Reconciliation and Data Queries (Weeks 12-18)
Enabled bank account reconciliation assist, which is generally available. It proposes matches for statement lines the standard rules leave open and suggests a G/L account for the leftovers — transfers, fees, interest, charges — with a short explanation. Finance reviews every proposal; nothing posts automatically. Across 3,100 statement lines a month, reconciliation went from 26 to 7 hours. Chat with Copilot went out alongside it, so staff can ask for a record or a figure and be taken to the page.
Measurement and Switching Things Off (Weeks 16-20)
Re-measured the same metrics and compared. Two features were switched off: sales line suggestions, because this manufacturer's orders arrive through EDI rather than being typed, and one where review time matched doing time. Reporting usage and outcome per feature keeps the deployment honest — the approach we use in Microsoft Copilot consulting engagements.
Solution Architecture
Grounding data: Item descriptions, vendor master, account naming, posting groups and default dimensions treated as a prerequisite, not a follow-up task
Document capture: E-document framework for structured supplier documents and incoming document OCR for the rest, with preview-stage Copilot matching to open purchase order lines
Exception handling: A review rule of our own — partial matches and tolerance breaches go to a person with nothing pre-filled, so weak guesses never reach a clerk
Bank reconciliation: Copilot proposes matches and G/L accounts for statement lines the rules leave open, with human review before posting. Deployed alongside our Business Central consulting team
Governance: Copilot capabilities managed centrally with features enabled individually, data movement consent recorded, and per-feature usage reviewed monthly
Results
Client identity withheld under confidentiality. Figures are representative of engagements of this scope.
Technologies Used
Key Takeaways
If your organization is facing a similar challenge, here's what we learned:
Copilot in Business Central is only as good as the item and vendor master. Every suggestion is grounded on descriptions, names and history. Part numbers used as descriptions, duplicate vendors and unnamed accounts produce confident wrong answers. Fix the data first.
Enable features individually, with a baseline for each. Switching everything on at once counts the features that could not help alongside the ones that did. Measure the task cost, enable one capability, measure again, and be willing to turn it back off.
Know which features are generally available and which are preview. Bank reconciliation assist, sales line suggestions and chat are GA. E-document matching is preview. That difference decides whether a feature can carry a month-end dependency, and collateral calling them all GA is checkably wrong.
Copilot does not repair a broken process. If purchase order discipline is poor, matching still fails: Copilot cannot match an invoice to an order nobody raised. Where the exception rate is structurally high, the fix is upstream.
Human review is the point, not a limitation. Bank reconciliation assist and invoice matching work because someone competent confirms them in seconds instead of researching them for minutes. See our Copilot for Dynamics work for the same pattern in the front office.
Related Case Studies
Is Copilot Earning Its Licence in Your ERP?
Most Business Central tenants have every Copilot feature enabled and no measurement. We baseline the tasks, fix the master data they depend on, and switch off what does not pay.