Let AI test-drive the spreadsheet before it costs you money

The most dangerous spreadsheet in a small business is rarely the biggest one. It is the file that has "always worked". The price list copied from last year. The staffing sheet where one formula skipped a school holiday. The budget workbook where a cost sits on the wrong row but looks reasonable at a glance.

That is why this week's practical AI signal is not another giant model launch. It is more ordinary: Microsoft keeps moving Copilot deeper into Excel, and the useful part is not that AI can write formulas for you. The useful part is that AI is starting to show its plan before it touches the workbook.

For small Nordic teams, this is a better starting point than "automate everything". Start with one file where an error would cost time, money, or trust. Let AI read, suggest, question, and explain. Do not let it press publish for you.

What Microsoft is changing in Excel and Copilot

Microsoft's April update for Microsoft 365 Copilot, updated on May 13, 2026, calls out Plan mode in Copilot in Excel. Plan mode means Copilot should outline a step-by-step plan before anything in the workbook is updated. The same update also says Python-powered editing in Copilot in Excel is rolling out in May, so more advanced analysis can happen without leaving the workbook.

It is easy to dismiss this as a feature for finance teams. I think the real use for smaller teams is simpler: AI becomes a second reader for files that only one person currently dares to touch.

Source: Microsoft 365 Copilot April 2026

Microsoft's support page for "Edit with Copilot in Excel" also says Copilot can build and change workbooks using Excel's own features: tables, charts, PivotTables and formulas. The important word is change. We are no longer only talking about chat advice beside the file. We are moving toward assistants that can suggest real edits inside the spreadsheet.

That needs a different habit. When AI only replies in a chat, you can ignore a bad answer. When AI changes a budget file, the team needs to know what changed, why it changed and who approved it.

Source: Edit with Copilot in Excel

Why this matters for small teams

A restaurant with ten employees may not have a BI team. A small school may not have a data analyst. A solo consultant rarely has time to rebuild a pricing workbook from scratch. Still, many decisions live in spreadsheets:

• How much margin remains after wages and materials
• which customers should be followed up this week
• Whether a class, course, or workshop pays for itself
• Which products are stuck in inventory
• How much time is really going into administration

This is exactly where AI can help, but only if you treat it as a reviewer, not as the answer key.

A spreadsheet audit with AI means letting the model read structure, headings, formulas and patterns, then look for oddities, weak assumptions and missing checks. It should not replace accounting, legal review or financial responsibility. It should help you ask better questions before you decide.

It is also a strong first exercise for Tool Forge: take a real tool the team already uses, wrap AI around it and build a review habit before the automation goes further.

Good AI help versus risky AI help in Excel

Microsoft's COPILOT function in Excel is a useful example of the boundary. The function lets users call AI directly from a cell, using a prompt and optional cell ranges as context. Microsoft describes it for semantic and generative tasks, such as summarizing text, classifying feedback or writing product descriptions.

The same support page is clear about where it does not belong: numerical calculations, deterministic answers, legal work, compliance and other situations where errors must not slip through. Standard Excel formulas, lookup functions and human review still need to do that work.

Source: COPILOT Function

That principle is useful even if you never use that exact function.

Good AI help in a spreadsheet sounds like this:

• "These three formulas seem to break the pattern. Do you want me to explain them?"
• "This cost category is missing in April but exists in March and May. Was that intentional?"
• "I suggest a check sheet that compares totals across tabs before the report is sent."

Risky AI help sounds like this:

• "I optimized the budget. Use this number."
• "I classified all customer comments and calculated the safe decision."
• "I changed the model so it looks cleaner."

The difference is not technical. The difference is whether AI makes the work traceable.

A 35-minute exercise: test-drive a spreadsheet without changing it

Choose a spreadsheet that is used often but is not life-critical. A quote template, a simple monthly budget, an inventory list, a schedule, or a follow-up sheet works well. Make a copy. Then use this prompt in Copilot, ChatGPT, Claude, or another AI tool that is allowed to see the file under your data rules.

You are my spreadsheet reviewer. Do not change anything in the file yet.

Goal: help me find risks, unclear assumptions and manual steps in this spreadsheet before we use it for a decision.

Do this in order:
1. Describe in five sentences what the workbook appears to be used for.
2. List the sheets, columns or areas that seem most important for the decision.
3. Look for formulas, blank cells, duplicates or unusual values that should be checked manually.
4. Mark which parts are safe for AI to suggest changes to, and which parts must be reviewed by a human.
5. Suggest a small control sheet with 5-8 checks we can add without changing existing formulas.
6. End with three questions you need answered before you are allowed to suggest any edit.

Important: If you are uncertain, say so. Do not invent formulas, regulations or financial conclusions.

When the answer comes back, choose one thing to do. Not five. Not a full automation. One control sheet, one clearer instruction or one manual checklist is enough.

This is deliberately slow. Small teams rarely win by jumping from "we have never reviewed this file" to "AI updates everything for us". The win is the new habit: AI finds friction, the human chooses the next step.

What should be forbidden in week one

Set a few plain rules before anyone lets AI loose on files with money or personal data.

• AI may not edit the original file. Work in a copy.
• The system may not decide prices, wages, grades, hiring, or customer terms.
• Do not use AI on sensitive personal data unless the tool, contract, and permissions allow it.
• AI suggestions must be reviewable in cells, formulas or separate comments.
• At least one human must sign off on changes that are actually used.

That sounds boring. Good. Boring rules are often what make AI useful in a week full of invoices, schedule changes and customer promises that cannot become experiments.

What to watch the next time Microsoft, OpenAI or Google shows spreadsheet AI

Do not only ask: "Can it create a report?" Most tools can make that demo look nice.

Ask instead:

• Does the tool show a plan before it edits the file?
• Can you see which cells, formulas, and data sources were used?
• Can you undo or compare changes?
• Does it work in Swedish, English, and the languages your team actually uses?
• Is there a chat-only mode when you only want to think through the file?
• Can an admin disable features for certain users or files?

Those answers say more about practical value than an impressive demo.

Thoughts on how this affects the future

AI in spreadsheets will not start with giant transformation projects. It starts when someone asks, "Why did April look so strange?" and gets help finding the cell that would otherwise take an hour to spot.

That is good, as long as we do not confuse faster analysis with safer decisions.

For Hammer Automation customers, the next step is often to build a simple AI routine around existing files: which spreadsheets may AI read, which ones must it never edit, which checks are required, and how are decisions documented? It is not glamorous. But it is where many small businesses and schools can start using AI without creating more mess than they solve.