AI in Excel for Accountants: What Actually Works in 2026

Guide · Updated July 2026 · ~11 min read

You have a 40,000-row ERP export, a bank statement PDF and a month-end deadline. Somebody said AI could do this now, so you paste a few hundred rows into a chatbot, get a plausible-looking answer, then spend an hour checking whether it's true — longer than doing it by hand.

The four routes, and why they're not interchangeable

1. A general chatbot (ChatGPT, Claude, Gemini) — the AI never touches Excel.
2. Microsoft 365 Copilot in Excel — built in by Microsoft, licence-gated.
3. Formula and analysis web tools (GPTExcel, Ajelix, Formula Bot) — a browser app you paste into.
4. In-Excel AI add-ins that write into the live workbook and can reach your accounting system.

What separates them isn't "how smart is the model". It's three things: does it see my whole workbook, does it apply the change or only describe it, and can it reach the ledger the numbers came from?

Route 1: Copy-pasting into ChatGPT or Claude

What it's genuinely good at

Free or near-free, and better than most people credit — provided you use it for language about spreadsheets rather than the spreadsheet itself:

• Explaining a formula you inherited. Paste =IFERROR(INDEX($D$2:$D$5000,MATCH(1,($A$2:$A$5000=$G2)*($B$2:$B$5000=$H2),0)),0) and ask what it does.
• Writing a formula from a description, e.g. "return the invoice date if the invoice is over 60 days old and unpaid, otherwise blank".
• Debugging a #VALUE! or a circular reference, if you describe the layout accurately.
• Drafting Power Query M or a VBA routine you paste in yourself.

Why it breaks on real workbooks

• Truncation. A month of transactions doesn't paste cleanly, and when it does you can't tell whether the model read all of it or the first few hundred rows.
• It can't see the rest of the file. Your figure depends on a lookup table on another sheet, a named range and a filter someone left on. The chatbot sees flat text.
• Nothing is applied. You still copy the answer back, place it, fill it down and check it. Fine once; on a monthly reconciliation it's the whole job again.

The confidentiality problem

Pasting a client's transaction listing, payroll register or bank statement into a consumer chatbot is a data-handling decision: identifiable financial data leaving your environment for a processor you never named in your engagement letter. If you do it at all: a business or enterprise tier with no-training-on-your-data terms; anonymise first (strip names and account numbers, keep dates, amounts and a row key); send structure and samples, not the population; and cover it in your firm's AI policy beforehand.

Route 2: Microsoft 365 Copilot in Excel

You open it by selecting the Copilot icon in the lower-right corner of Excel — a task pane, not a ribbon tab.

What it actually does

Microsoft's capability list for Copilot in Excel is broader than most people realise: edit worksheets (add, rename and delete sheets; insert and modify cell values and ranges; conditional formatting, data validation, borders and styles), generate data and formulas, create charts, PivotTables and shapes, identify insights, search the web and import data from other workbooks. You can also pick the model — it "supports the latest Anthropic and OpenAI models", with Auto the default.

There's also a native worksheet function, =COPILOT():

=COPILOT(prompt_part1, [context1], [prompt_part2], [context2], ...)

So =COPILOT("Classify this expense as Travel, Software or Professional Fees", A2) returns a category into the cell — but read the limits below first.

What it costs — name the SKU

There is no single "Copilot price" (figures from Microsoft's Copilot pricing page):

The $18.00 figure is a promotion running between 1 July 2026 and 30 September 2026, annual commitment, first year only — budget on $21.00. Not every plan carries Copilot into Excel; the licence list is in the Copilot in Excel FAQ. Microsoft lists it for Excel for Microsoft 365 (Windows and Mac), Excel for iPad and the Excel Web App.

The limits you'll hit

• Calculation mode. "Copilot editing is only supported when Calculation Options are set to Automatic." If your model runs on Manual because a recalc takes a minute, Copilot editing is out.
• SharePoint check-out. "If your organization requires check-out for editing in SharePoint, Copilot in Excel on Windows or Mac won't work for files from that site."
• =COPILOT() is rate-limited and blind. Up to 100 COPILOT functions every 10 minutes, and it "only has access to the prompt and context provided" — no other workbook data, external files or enterprise data. It won't calculate in Confidential or Highly Confidential workbooks, and it needs an internet connection. Still gated behind the Frontier and Microsoft 365 Insider programs, so treat it as preview.
• Autosave is not a gate, despite the widely repeated claim: "Copilot in Excel runs in workbooks with autosave enabled or disabled."
• Microsoft's own warning. The FAQ advises users to "Avoid using Copilot for decisions in sensitive areas such as finance, legal, or medical topics."

Each of these is unpacked in what Microsoft Copilot in Excel can't do for accountants.

Can it reach your accounting system?

Partly, and probably not the part you want. Copilot in Excel pulls external data through federated connectors — real-time retrieval that isn't stored in the workbook, needing consent re-confirmed in each new chat, admin-controlled, and only on a commercial Microsoft 365 Copilot subscription. The published list skews to market data and productivity tools: Canva, HubSpot, Google Calendar, Notion, Linear, FactSet, Morningstar, S&P Global. QuickBooks and Xero are not on it.

That may shift: Xero announced a Microsoft 365 integration on 1 July 2026 that surfaces live Xero data inside Microsoft 365 Copilot via its JAX agent (media release). But today, "ask Copilot for my aged payables from QuickBooks" isn't a thing.

Route 3: Formula and analysis web tools

A stale assumption worth correcting: two of the three best-known names are no longer cheap formula generators.

• GPTExcel (gptexcel.uk) is still primarily a formula/script/SQL/regex generator plus spreadsheet chat. Free $0 (10 AI chat messages per 30 days); Pro $9.00/month or $6.30 billed annually; Pro Plus $18.00 or $12.60 annually.
• Ajelix (ajelix.com) now markets an agentic AI platform with 17 AI-powered tools — formula generators, Excel and Google Sheets agents, analytics, workflow automation. Lite $39 user/month ($32 yearly), Pro $99 ($82), Max $199 ($165). Free tier: 10 requests a month.
• Formula Bot (formulabot.com) positions as "The AI data analyst for Business Owners" — analytics, dashboards and connectors for Google Analytics, Snowflake, MySQL and PostgreSQL. Free $0 (10 messages a month); Starter $18/mo ($12.58 annual); Max $29/seat/mo ($20.33); Enterprise $149/seat/mo ($99).

Where they fit: one formula, one regex, one M snippet, without opening a chat — the free tiers cover an occasional user. Where they don't: none sits inside your workbook. They see what you paste, not the eleven other tabs the number depends on, and uploading a client file raises the same questions as route 1.

Route 4: In-Excel AI add-ins that apply the work

The last category runs inside desktop Excel, reads the live workbook and writes the result into it — and, in the accounting-specific ones, connects to the ledger the data came from. It's the only route that closes the loop from ledger to sheet and back — see connect QuickBooks and Xero to Excel.

HISAB 360 is one of these. It docks an AI chat panel inside Excel; you describe what you want in plain English and it writes and applies the work to the open workbook — formulas, Power Query (M) queries, PivotTables, Power Pivot data-model tables, relationships and DAX measures, charts and VBA macros. It reads and writes six accounting systems (QuickBooks Online, Xero, Zoho Books, Odoo, FreshBooks, Sage), pulling invoices, bills, payments, journals, ledgers and reports into the sheet and posting or updating them back. It also does AI bank reconciliation (amount and date tolerance, many-to-one deposits, fuzzy descriptions, duplicate detection), data cleaning, PDF table extraction and double-click drill-down on a total.

The limits matter. It's Windows desktop Excel only (Excel 365 / 2019 and later) — no Mac, no Excel on the web, which is exactly where Copilot does work. It's paid, with no free plan: $7/month or $60/year to bring your own AI key, $10 or $20/month hosted, $199 perpetual, plus a limited free trial allowance of credits (breakdown in the FAQ). Connecting a ledger requires you to authorise access via OAuth, and every write back should be reviewed before you approve it and reconciled afterwards — you are the control, not the model.

The comparison table

Five accounting jobs AI genuinely does well in Excel today

The real test is whether the output is checkable in under a minute. These pass it.

1. Reconciliation matching — statement to ledger on amount plus a date tolerance, with many-to-one deposits and fuzzy descriptions. You still review the exceptions. Walkthrough: reconcile a bank statement in Excel with AI.
2. Categorising transactions. Ask for a description-to-category mapping table rather than per-row labels, then apply it with XLOOKUP — you review 60 rules instead of 3,000 rows.
3. Cleaning an ERP export — text-stored-as-numbers, trailing minus signs, merged header rows, Total lines mixed into the detail — and, because QuickBooks Online exports values rather than formulas, no live calculations to inherit. See clean messy data in Excel.
4. Writing M and DAX, e.g. "merge the sales export into the customer master on Customer ID, keep all sales rows", or Revenue LY = CALCULATE([Revenue], SAMEPERIODLASTYEAR('Calendar'[Date])). Unforgiving syntax, simple logic, easy to verify. Background: Power Query for beginners.
5. Extracting a PDF statement into a structured range, or bucketing an aging analysis into 0–30 / 31–60 / 61–90 / 90+ and pivoting by customer. Foot the result back to the source before you use it.

What AI is still bad at

• Judgement calls. Capital or revenue, provision or not, related-party balance recoverable or not. AI gives a confident answer with no accountability attached.
• Materiality. A model has no sense of what a 0.4% variance means in your context. It flags noise and skips the one line that matters.
• Evidence. Nothing an AI produces is audit evidence. The evidence is your check of it and your documentation of that check.
• Silent failure. A wrong formula returning a plausible number is more dangerous than one returning #N/A. Foot totals, agree control counts, reconcile to a known figure.
• Ambiguous input. If two humans would read the source differently, the model picks one reading and doesn't say it chose.

Use AI for the mechanical part; spend the saved hour on the part that needs an accountant. See also Excel skills for accountants and Excel automation for finance teams.

Frequently asked questions

Try HISAB 360 on your own workbook

HISAB 360 is an AI assistant inside Excel for accountants and finance teams — it writes macros, Power Query and formulas from plain English, and connects two-way to QuickBooks, Xero, Zoho Books, Odoo, FreshBooks and Sage. The 15-day trial is the full product, no card required.