Which tool, and what it costs

Four shapes, not forty brands

The market is noisy and the brand names change every few months. The shapes do not. Almost everything you will be offered at work is one of four things, and most disappointment comes from using the wrong shape rather than the wrong brand.

1. The chat window. ChatGPT, Claude, Gemini, Copilot, and dozens of others. You paste text in, you get text back. Best for drafting, rewriting, explaining, restructuring and thinking out loud. Free tiers of the major ones are genuinely usable for most of the work in this course; paid tiers, typically around $20 a month, mostly buy higher limits, better models and file handling.

2. The assistant inside your files. Gemini in Google Workspace, Copilot in Microsoft 365, and the AI features now built into most document tools. The difference that matters is not the model — it is that it can see the document you are in, and it writes back into it. Convenient, and priced per seat per month, usually as an add-on your organisation buys. The free path: keep using a chat window and paste in what you need. You lose the convenience, not the capability.

3. The transcriber. Turns speech into text: meeting notes, interviews, dictated letters. A distinct technology from the chat models, with distinct failure modes. Otter, Fireflies and the recording features in Teams, Zoom and Meet are the paid versions. The free path is strong here: OpenAI's Whisper is open source and runs on an ordinary laptop through whisper.cpp or a desktop wrapper, and Google's Live Transcribe and Recorder do a decent job on a phone at no cost.

4. The one that runs code. The mode variously called data analysis, code interpreter or advanced analysis. You give it a file; it writes a Python program, runs it, and shows you the result and the program. This is the only one of the four that does arithmetic rather than predicting text that resembles arithmetic, which makes it the correct shape for anything involving counting, totalling, or a chart from real rows. Available on the paid tiers of the big chat products; the free path is to ask any chat tool for the spreadsheet formula and run it yourself, which is slower and equally reliable.

The category errors that waste people's time

• Counting in a chat window. Rows pasted into shape 1 produce a plausible total, not a total. Use shape 4, or a formula. This is the single most common mistake in office use.
• Expecting the chat window to know your files. It knows only what is in the conversation. If you did not attach the policy, it is guessing at your policy.
• Expecting a transcriber to summarise well. Most bundle a summariser and most summaries are mediocre, because the product is optimised for the transcript. Take the transcript and do the summarising yourself in shape 1, with instructions about what must appear.
• Buying seats before auditing tasks. Shape 2 is the most expensive per person and the most often unused. Run the audit first, then buy what the audit points at.

What to check before you type anything into it

Three questions, none of them about quality.

1. Whose account is this? A personal account you signed up for on your phone is not an approved work tool. This has a whole lesson later; the short version is that the difference between a consumer account and your employer's tenant is a legal difference, not a feature difference.
2. Is training on by default? On consumer tiers it often is. On business and enterprise tiers it usually is not. It is a setting you can read in ten seconds and should.
3. Where does the data go? For most enterprise deployments the answer is a specific region and a specific retention period, and your IT team knows it. For a free personal account the honest answer is: somewhere else, kept for a while, possibly read by a person.

A workable free stack

If your employer has bought nothing and you need to start today, this combination costs nothing and covers most of the course:

• A free chat tier for drafting, summarising and explaining.
• LibreOffice Calc or Google Sheets for the actual arithmetic, with formulas the chat tool wrote and you tested.
• Whisper locally for transcription, when consent and confidentiality allow recording at all.
• OpenRefine for messy data — free, and better than any chatbot at deduplication and inconsistent categories.
• A plain text file of prompts that worked. Unglamorous and the highest-return item on the list.

Two constraints on this stack are worth naming up front. Free tiers have message limits that arrive at the worst moment, so do not build a deadline around one. And a personal free account is the least appropriate place for anything confidential, which is a subject serious enough to have a module of its own.

The choice that is not about the tool

Whatever you pick, the value comes from the tasks you chose in the audit, not from the brand on the tab. A well-audited team on free tiers reliably beats a badly-audited team with enterprise licences. The tools converge; the judgment does not.