AI in finance: the month-end close, cash forecasts and margins
AI in finance, for a company that buys, makes or sells physical goods, means software that reads the general ledger, the ERP and the spreadsheets your finance team already keeps, and drafts the routine work: bank matches at month-end, cash and sales forecasts, margin reports by product and customer, and the notes that explain why a number moved. A controller or owner approves each draft. This guide covers what AI for finance does at a manufacturer, a distributor or a food producer, the data it needs, how to start with one report, and where it goes wrong.
What AI in finance means for a company that sells goods
The phrase covers two different worlds. In banks and insurers, AI scores credit, flags fraud and trades. This guide covers the other one: the finance team inside an operating company, meaning the controller, the bookkeeper and the owner or CFO at a business that buys, makes or sells physical goods. Their work runs on the general ledger, the ERP that holds orders, inventory and purchasing, and the spreadsheets that connect the two.
Finance teams have been adopting AI for a few years. In Gartner’s 2025 AI in Finance Survey of 183 CFOs and senior finance leaders, 59% reported using AI in their finance function, up from 58% in 2024 and 37% in 2023. Among the teams that had put AI to work, the most common uses were knowledge management (49%), accounts payable automation (37%) and error and anomaly detection (34%). Gartner also found that 91% of respondents reported low or moderate impact at first, and that larger gains came once projects reached production.
Canadian wholesalers are among the slowest adopters. In Statistics Canada’s survey for the second quarter of 2026, 19.2% of all businesses had used AI to produce goods or deliver services in the previous 12 months, against 7.9% in wholesale trade. Among the businesses that did use AI, the most common application was data analytics, at 36.6%.
How the finance work is done by hand
At a manufacturer, a distributor or a food producer, the monthly cycle looks much the same. Someone exports the trial balance and the bank statements, matches deposits to invoices and payments to bills, chases receipts that have no supplier invoice yet, posts accruals and inventory adjustments, and rolls the numbers into a management report. Each step lives in a spreadsheet or a checklist, and much of it depends on the one or two people who know where the exceptions are.
Benchmarks show how long the cycle takes. APQC, which runs a large cross-industry benchmarking database, measures the calendar days from running the first monthly trial balance to the finished result. The median is 6 days to complete the monthly consolidated financial statements, across 11,223 companies, and 10 days to produce the period-end management report for senior management, across 3,287 companies.
Forecasts get rebuilt by hand too. BDC’s cash flow template includes a 13-week view, described as a way to “Forecast your business’s cash inflows and outflows over a quarter to better plan your short-term needs.” Where that forecast lives in a spreadsheet, someone updates it each week from the receivables aging, the bills due and the payroll calendar. A margin report by product or customer starts from a sales export and a standard cost that may have been set at the start of the year. The notes that explain why a number moved get written last, by whoever has time left.
Physical goods add work that a services firm does not have. Inventory has to be counted and valued, freight and duty have to land in the cost of the right items, and a plant has to estimate its work in progress at month-end. Customer rebates and supplier price changes arrive by email and PDF, outside the ERP, so someone keys them in before the numbers are right.
What AI changes in the finance work
AI in finance does three kinds of work. It matches and sorts records, it forecasts from the company’s own history, and it drafts text that explains the numbers. In each case a person approves the result before anything posts to the ledger or leaves the building.
The month-end close
Matching comes first. In Microsoft Dynamics 365 Business Central, bank account reconciliation with Copilot runs after the standard automatic matching and uses AI to inspect the transactions left over, finding more matches “based on the dates, amounts, and descriptions.” Microsoft’s example is a customer who paid several invoices in one lump sum, which Copilot matches to the separate ledger entries. For lines that match nothing, it suggests the general ledger account to post to, and the accountant reviews the proposed matches line by line. The feature is in preview.
Accounting software for smaller companies is adding the same kind of help. In July 2025, Intuit announced an Accounting Agent for QuickBooks Online, rolling out first to customers in the US, that “automates bookkeeping and transaction categorization, and assists in reconciliation.”
Checking comes second. Software that has learned what normal journal entries look like for your company can list the unusual ones, such as a vendor posted to an account it has never used, a round amount booked on the last day of the month, or an invoice number entered twice, so that the controller reviews a short list instead of every entry. Accruals follow the same pattern: goods received against a purchase order with no supplier invoice yet are listed with their value at the order price, ready for the accrual entry.
Supplier invoices and the three-way match
Accounts payable automation was the second most common use in Gartner’s survey. Language and vision models read a supplier invoice from a PDF or a scan, match it to the purchase order and the receipt, and hold anything that disagrees for a person to resolve. AI invoice processing covers that match in detail, and AI in procurement covers the purchase orders behind it.
Cash forecasts
A cash forecast improves when each line comes from how people actually pay. Software can measure the days each customer really takes to pay, from the dates on past invoices and payments, and project open receivables on that basis instead of on the due date. Payments come from the bills already entered and the open purchase orders, and payroll from its calendar. The result is a weekly view of the next 13 weeks that updates when the books change, with each line traceable to the invoice or order behind it.
Sales are the hardest input. A forecast of orders by product family comes from the order history, which AI demand forecasting covers, and stock purchases follow from the reorder settings in inventory optimization. Stock bought this month is cash spent this month, so the sales forecast, the purchasing plan and the cash forecast belong in the same model.
Margins by product, customer and order
A margin report uses the cost on file, and that cost can be months old. When AI recomputes each item’s cost from the latest purchase receipts, freight bills and bills of materials, it can list the products, customers and orders whose margin has slipped below the floor you set, with the amount in dollars. Margin analysis covers the method. Cost-plus pricing covers prices built from cost, and price optimization covers prices set from sales and quote history.
Variance notes and questions about the numbers
Writing the explanation is a natural job for a language model. Microsoft’s finance solution in Copilot includes a variance analysis feature for Excel, in preview, built to “highlight critical variances, key contributors, and trends” and to write an executive summary of the findings. A controller still checks each explanation against the ledger, because a fluent paragraph can be wrong.
Knowledge management, the most common use in Gartner’s survey, looks like this in a finance team: someone asks what the company paid for freight on one customer’s orders last quarter, and the answer comes back with the invoices it was built from. That depends on a system that searches your own records and cites them, which retrieval-augmented generation explains. Budgets and the three-statement model are covered in AI financial modeling.
AI for finance at a glance
| Job | What the software drafts | Data it reads | Who approves |
|---|---|---|---|
| Bank reconciliation | Proposed matches, and a suggested account for lines that match nothing | Bank statements, the ledger, open invoices and bills | Accountant, line by line |
| Journal entry review | A short list of unusual entries, each with the reason it was flagged | Journal entries and their history | Controller |
| Accruals | Goods received with no invoice yet, valued at the order price | Purchase orders, receipts, supplier invoices | Controller |
| Supplier invoices | Invoice fields matched to the purchase order and the receipt | Invoices, purchase orders, receipts | Payables clerk, then the budget owner |
| 13-week cash forecast | Weekly receipts and payments, each traced to its source | Receivables with payment history, payables, payroll, open orders | Owner or CFO |
| Margin report | Margin by product, customer and order, with items below the floor listed | Invoice lines, purchase receipts, freight, bills of materials | Controller or sales lead |
| Variance notes | A draft explanation of each large variance | Budget, actuals, transaction detail | Controller |
Pick the finance job to start with
Tell Derik which accounting system and ERP you run, and which month-end task takes the most hours. He will tell you what your data can support.
Start a conversationWhat data AI for finance needs
Every job above reads records the business already keeps. The work is getting them out cleanly and fixing the places where they disagree.
- The general ledger. A monthly trial balance and the transaction detail behind it for two to three years, with the chart of accounts and a record of any changes to it.
- Receivables and payables. Invoices and bills with their dates, due dates and payment dates, which show how each customer really pays.
- Bank statements. Exported files or a bank feed for every account.
- ERP sales and purchasing. Invoice lines with item, customer, quantity, price and discount, plus purchase orders and receipts at the prices actually paid.
- Inventory and costs. On-hand quantities and standard costs, and for a plant the bills of materials and routings.
- Payroll totals. Totals by department and pay date. Individual pay records hold personal information, and most finance jobs need only the totals.
- Budgets. The spreadsheets the team already uses, which show how the business thinks about its own numbers.
Three data problems come up often. A chart of accounts that changed mid-year splits one history in two. Freight, rebates or inventory adjustments posted as one lump entry hide which products and customers they belong to. Manual journal entries with no description give the software nothing to explain. If your ERP has no export or API, legacy ERP automation covers older systems, and AI for ERP covers the other routes.
How to start with one report
- Pick the report that costs the most hours. Bank reconciliation, the 13-week cash forecast and the monthly margin report each have a clear right answer to test against.
- Pull the history. Export two to three years of the records that report uses, and fix the gaps the export reveals.
- Define the check. Decide how you will know the output is right, for example that the forecast’s opening cash ties to the bank and the margin report’s revenue ties to the income statement to the dollar.
- Run it beside the current process. For two month-ends, compare what the software drafts with what your team produces. AI evals explains how to test on past work.
- Keep a person approving. Nothing posts to the ledger or goes to the bank without a named approver. Human in the loop covers designing an approval step people actually read.
- Measure the result. Track the hours spent and the days to close before and after.
Cost of AI explains how each part of a build is charged, and AI strategy covers choosing what to do second.
Risks and limits
Language models make mistakes that read well. OpenAI’s help centre says ChatGPT “can produce incorrect or misleading outputs.” In a finance system, the arithmetic belongs in the ledger, a spreadsheet or code, and the model matches, drafts and explains around it.
- A source for every number. A forecast line links to the invoices or orders behind it, and an explanation links to the transactions it describes.
- Separate duties. The software proposes matches, accruals and notes, and an accountant approves them before anything reaches the ledger or the bank.
- Limits of history. A forecast built from the past cannot know about a customer you lost last week or a price increase announced yesterday, so someone adjusts for what the records cannot show.
- Records kept. The Canada Revenue Agency says you must generally keep records and supporting documents “for a period of six years from the end of the last tax year they relate to.” Keep the source documents an AI step reads, and a log of what it proposed and who approved it.
- Data location. Ledgers, payroll totals and customer records are sensitive, so know where each AI step sends them and whether the provider keeps a copy. Private AI for business lists the questions to put in writing, and data sovereignty in Canada covers where data may sit.
How ThriveAI helps
ThriveAI is an AI engineering company in Ottawa that builds private AI systems on a company’s own data, for businesses that make, move or sell physical goods. For a finance team, that means reading the general ledger, the ERP and the spreadsheets you already keep, reconciling them so each number means the same thing everywhere, and drafting close work, forecasts and reports for a named person to approve. Every answer shows where it came from.
ThriveAI’s systems read your ERP and accounting system as they are, and every connection only reads data. They are designed to keep each client’s data on its own server in Canada. You choose a model on that server or a hosted model under a written zero data retention agreement, and a hosted model may process requests outside Canada. Derik Lawlis, the founder, leads every project and stays close to the build. About ThriveAI covers the company.