The AI analysis tool landscape in 2026: assistants, spreadsheets, notebooks and MCP

0:00 The AI analysis tool landscape

Ask five analysts which AI tool they use for data, and you'll get five answers: ChatGPT, Claude, Copilot in Excel, Gemini in Sheets, or a notebook in Colab. They're all partly right. In this lecture you'll get a clear map of where AI analysis lives as of September twenty twenty-six, four questions for choosing the right tool for a task, and a hands-on exercise where you answer the same question in three tools and learn from where they disagree.

0:34 Why tool choice matters

Why does tool choice matter? Because it decides where your data goes, how big a problem you can tackle, and whether your result is a durable artifact or a disposable chat answer. A formula in a workbook can be checked by finance. A SQL query can be scheduled. A notebook can be rerun next month. A chat paragraph can't be any of those things. And policy matters: some data shouldn't leave your company's tenant at all.

1:07 Four places

Here's the map. First, general assistants with code execution: ChatGPT's data analysis, Claude with code execution and file creation, which can produce Excel, PowerPoint, Word and PDF files, and Gemini. Second, inside spreadsheets: Copilot in Excel and Python in Excel on supported Microsoft 365 plans, the Claude for Excel add-in on paid Claude plans, and Gemini in Google Sheets. Third, notebooks and warehouses: Gemini in Colab, Gemini in BigQuery, and conversational agents in Data Studio. Fourth, assistants connected to live systems through M C P servers.

1:44 The kitchen analogy

Here's an analogy for choosing. Think of a kitchen. A general assistant is like a talented chef visiting your home: brilliant, flexible, but you hand over your ingredients. An in-spreadsheet assistant is a chef who works in your kitchen, with your utensils, and leaves everything where you can see it. A notebook or warehouse is a commercial kitchen: bigger, repeatable, governed. And M C P connections are like giving the chef a key to the pantry, so give them a key that only opens the shelves they need.

2:22 Four questions

Now four questions for choosing. One: where does the data live, and may it leave? If policy says it stays in your Microsoft or Google tenant, start with the approved in-spreadsheet assistant. Two: how big is it? Uploads have size limits, for example Claude documents a thirty megabyte per-file limit in its apps, so large tables belong in a warehouse or notebook. Three: do you need a reusable artifact, like a formula, query or notebook? Four: who will review it? Choose the output your reviewer can read.

3:00 Features change

And remember, features and plans change constantly. A capability available on one license, in one region, may not exist on another. In September twenty twenty-six, Microsoft, Google, OpenAI and Anthropic all shipped meaningful changes to their data features within months of each other. So before you standardize on a tool, check the vendor's current documentation and your organization's approved list. The durable skill isn't knowing today's feature list. It's asking the four questions.

3:32 Example 1: a London consultant

First example, a simple one. A solo consultant in London serves small e-commerce clients. She uses a general assistant with code execution for fast one-off analyzes on anonymized exports. Then she asks it to produce a clean notebook and an Excel file with real formulas as the deliverables. Clients get files they can open and check, not screenshots of a chat. And her contract states which tools she uses and that client data is anonymized before upload.

4:05 Example 2: a Riyadh retail chain

Second example, a business case. A Riyadh retail chain's finance team works in Excel on Microsoft 365, with strict data-residency rules. IT approves Copilot in Excel and Python in Excel for internal data, and a general assistant only for anonymized, aggregated exports. The team standardizes. Exploratory questions happen in Excel with AI. Recurring reports become reviewed Python in Excel or SQL in the warehouse. And executive summaries are drafted by the approved assistant from aggregated tables only.

4:38 Watch me do it, part 1

Watch me answer one question in three tools. The question: which of our twelve product categories had the biggest drop in gross margin from the second to the third quarter, and how many orders is that based on? In a general assistant, I upload the CSV with a prompt that asks for Python, margin by category and quarter, the change in percentage points, order counts, a sorted table, and a flag for categories under a hundred orders. It shows its code, and I read the filters and the group by.

5:17 Watch me do it, part 2

Next, the spreadsheet. I write a LET formula with SUMIFS for revenue and cost per category for the second quarter, and compute margin, with the date boundaries from the first of April up to, but not including, the first of July. Then the warehouse. A SQL query groups by category and quarter, counting distinct orders and computing margin with safe divide. Now I compare all three. They agree to rounding, except one category. That disagreement turns out to be refunds: the SQL table excludes them, the CSV includes them. That's my most valuable finding.

5:58 Disagreement is data

Here's the key idea from that exercise. When three tools disagree, don't pick the one you like. Find out why. The disagreement is almost always a definition, a filter or a data problem, exactly the kind of thing that would otherwise silently distort your decision. Triangulating across tools is one of the cheapest, most powerful verification techniques you have.

6:24 Durable deliverables

One more practical point: deliverables. Whatever tool you use, ask it to hand you something durable. From a general assistant, ask for a notebook or script plus an Excel file with live formulas, not just a chat answer. Claude, for example, can create and edit Excel and PowerPoint files directly. From a spreadsheet assistant, keep the formulas and pivots it creates, and check them. From a warehouse, save the query with a comment describing its definitions. The deliverable is what gets reviewed, reused and trusted.

7:01 Common mistakes

Common mistakes. Choosing a tool by hype instead of data policy, size and reviewability. Uploading raw customer data to a personal account. Treating a chat answer as a deliverable instead of the formula, query or notebook behind it. And assuming a feature exists on your license or in your region because you saw it in a demo.

7:26 Success signs

How do you measure success? Your team has an approved-tools list by data type. Every recurring analysis has a durable artifact: a formula, query or notebook. Reviewers can read what they approve. And disagreements between tools are logged and explained, not ignored. When those four things are true, your AI tool choices are a system, not a habit.

7:51 Recap

Recap. AI analysis lives in general assistants, spreadsheets, notebooks and warehouses, and connected systems. Choose with four questions: where the data lives and may go, size, whether you need an artifact, and who reviews it. Check current docs because features change. And answer important questions in more than one tool, treating disagreement as information.

8:14 Try this now

Try this now. Pick a real question you answered recently. Answer it again in two different tools, one of them a spreadsheet formula or SQL query you can inspect. Compare the results. If they differ, find the reason, and write it down as a definition or data note for your team.