Learn · Professional

AI for Documents, Spreadsheets, and Meetings

The paperwork trilogy: where office hours actually go, and where AI gives the most of them back.

Ask people where their working hours actually go and three answers dominate: reading and producing documents, wrestling spreadsheets, and sitting in meetings. This guide is a recipe book for exactly those three — the specific requests that work, and the specific ways each one can bite if you skip the checking step. It assumes the basics from [the working person's guide](/work/everyday-ai-at-work/): sanctioned tool, work account, data rules respected.

Documents: read with a question, draft with a skeleton

The weakest way to use AI on a long document is 'summarize this.' You get a shorter document you still have to interpret. The strong version brings the question you actually have:

*What am I being asked to decide or approve, and by when?*
*What changed from the previous version?* — paste both, ask for the differences that matter and the ones that look cosmetic but are not.
*What would a careful skeptic challenge here?*
*Where does this contradict itself, or the other document I am pasting alongside it?*

Reading with a question turns a wall of text into an answer. It also works in reverse: for producing documents, ask for a skeleton before prose — 'give me the outline a strong version of this proposal would have' — argue with the outline, then have it draft section by section under your direction. You stay the author; it does the typing.

Two cautions earn their place here. Contracts and legal documents: AI is excellent at explaining unfamiliar terms and flagging clauses worth attention, and that is genuinely valuable — but 'the AI read the contract' is not a review, and anything with obligations attached deserves the professional whose job that is. And long documents strain attention: on very long material, ask for section-by-section treatment rather than one heroic pass, and spot-check that quoted passages actually appear in the source — a fabricated quotation reads exactly as smoothly as a real one.

Spreadsheets: let it write the formula, not do the math

The single most useful rule for AI and spreadsheets: AI is a brilliant explainer and builder of spreadsheet machinery, and an unreliable calculator. Language models do arithmetic the way people recall phone numbers — usually right, confidently wrong often enough to matter. So aim it at the machinery, not the sums:

Explain this formula — paste the incomprehensible nested formula a departed colleague left behind and get a plain-English account of what it does, and where it is fragile.
Write the formula or script — describe the outcome in plain language: 'monthly totals by region, ignoring rows where status is cancelled.' Formulas, pivot-table setups, and small scripts are code, and code is home turf for these tools — especially [the terminal agents](/work/cli-coding-agents/), which can process the files directly.
Clean the mess — inconsistent date formats, names split wrong, duplicates: describe the mess, get the cleanup recipe.
Interrogate the data — 'what looks off in this table?' catches outliers and gaps human eyes skim past. Treat its answers as leads to verify, not findings to report.
Draft the narrative — paste the final table and ask for the summary paragraph the report needs. Then check every number in the prose against the table, because transcription is exactly where fabrication sneaks in.

The pattern across all five: the formula either works or it does not — that is checkable. The prose *about* numbers is where errors hide, so that is where your attention goes.

Meetings: the win is before and after, not during

Before: paste the agenda and the notes from last time. Ask what you will likely be asked, what is unresolved, and what you should be pushing for. Five minutes of this outperforms most people's meeting prep because most people's meeting prep is none.
After: notes or transcript in, then ask for the three-way split that makes follow-ups useful — decisions made, actions with owners, questions left open. Draft the recap message from that. The action items your future self will thank you for are the ones captured the same afternoon.
The recurring meeting you own gets a compounding version: keep the running notes in one place, and each week ask what was promised last time and never closed. AI is extremely good at being the institutional memory nobody wants to be.

On recording and transcription: the tools are good now, and widely built into the platforms — but recording a conversation is a consent and policy question before it is a technology one. Know your company's rule and your jurisdiction's, and when in doubt, visible notes beat silent recording.

Presentations: the deck is the last step

Bad decks come from starting in the slide software. Better: tell the AI who the audience is, what they believe now, and what you need them to agree to — ask for the storyline first, as a sequence of points. Fight about the storyline, then let it draft speaker notes and slide text per point. Design polish belongs to your template, not to prose generation; what AI moves is the thinking-and-structure hours, which were always the real cost of a deck.

One habit ties this whole guide together: match the checking to the stakes. A summary for your own orientation needs a skim; numbers going to a customer need every figure traced to source; a contract needs a professional. AI changes how fast the work gets done — [what it does not change is whose work it is](/work/ai-work-habits/).

Common questions

What is the best way to summarize a document with AI?

Bring a question instead of asking for a generic summary: what am I being asked to decide, what changed from the last version, what would a skeptic challenge. Purpose-shaped requests produce answers rather than shorter walls of text. On very long documents, work section by section and spot-check that quoted lines actually exist in the source.

Can AI do spreadsheet calculations?

Use it to build and explain the machinery — formulas, scripts, pivot setups, cleanup recipes — rather than to perform arithmetic, which language models get confidently wrong often enough to matter. A generated formula is checkable and reusable; a generated total is a guess wearing a suit. Verify any number that appears in prose against the actual table.

How can AI help with meetings?

The leverage is before and after the meeting: preparation briefs built from the agenda and prior notes, and afterwards a three-way split of the notes into decisions, actions with owners, and open questions, plus a drafted recap. For recording or transcribing, check consent rules and company policy first — that question precedes the technology.

Should I let AI review contracts?

AI is genuinely useful for explaining unfamiliar terms and flagging clauses worth a closer look, which makes you a better-prepared reader. It is not a substitute for professional review of anything with obligations attached — treat it as preparation for the conversation with the expert, not a replacement for one.