Four days of visiting a chatbot. Today the model comes to you, inside the documents, sheets, inboxes and meetings.
No copying, no pasting, no re-explaining context the file already contains. Embedded assistants can see your material: Copilot in Excel sees the sheet, Gemini in Docs sees the document. That is why summarise the decisions from last week’s meetings works there and cannot work in a bare chatbot.
The price is that you are handing organisational data to a system. Before you roll anything out, know three answers: is this the consumer tier or the business tier, is our data used for training, and does our client contract permit it. In our market, with government and semi government clients, this is not paperwork. This is the deal.
The strongest use is formula authoring and explanation. Describe the calculation in words and get the formula, or paste a monstrous nested formula you inherited and get it explained line by line. The next strongest is cleaning and structuring: splitting inconsistent columns, standardising dates, flagging duplicates, converting a text blob into a table.
Analysis is where care is required. Ask for trends and you will get confident prose about your numbers. Some of it is a correct reading, some is a plausible sounding pattern the data does not support. Treat every generated insight as a hypothesis to check against the sheet, not a finding.
Run every generated formula on a row where you already know the answer before it goes anywhere near a client. This takes thirty seconds and has saved more careers than any prompt technique in this course.
For documents, the AI first draft is a scaffold, not a draft. It gets you past the blank page and gives you a structure to argue with, and the rewrite is where the value is: change the tone, tighten this section, make this argument stronger, cut a third.
For email, the leverage is triage and reply drafting, with two cautions from practice. Long AI written emails read as AI written to anyone paying attention, and in a relationship driven market that costs you. A short human reply almost always beats a polished paragraph of nothing. Draft with it, then cut it in half and put your own voice back in.
Automatic transcription and summary changes what people do rather than what they produce: nobody is taking notes, so everybody is actually in the conversation, and action items surface reliably. Consent first. Recording without telling the room is a legal and relationship problem, and in the Gulf it is a trust problem, which is worse.
Embedded assistants are good at the mechanical and the structural. They are weak at judgement, at knowing what your organisation actually cares about, and at anything where being wrong is expensive and undetectable. Route the drudgery to them and keep the decisions.
NOVA reacts, nothing is scored, nothing is stored against you.
Take a spreadsheet you already maintain and use Copilot or Gemini to write three formulas you would otherwise have looked up. Validate every one against rows where you know the answer. Then have it explain the most complicated formula already in that sheet.
Take a real recurring report and rebuild the production process with embedded AI at every stage: gathering, calculation, draft, formatting. Time the old way and the new way, and document what you had to check by hand.
Day 5 in progress
Tomorrow, Day 6, Phase 2 begins: pictures. Words become images, and you become picky about light.