Lesson 6 of 8 · 9 min read
So far you’ve mostly been asking AI to produce something from nothing. This lesson flips it: you bring the material, and it helps you make sense of it.
This is where AI earns its keep for most office work. Not writing your emails — reading the forty things you don’t have time to read.
It’s also where it fails in the most dangerous way, because a wrong number looks exactly like a right one.
Bring the material
The shift is simple: instead of “write me a summary of industry trends,” you paste in the actual report and say “summarise this.”
Everything gets better immediately. You’re no longer relying on what it absorbed during training — with the material in front of it, it’s working from your document rather than from the average of everything ever written on the topic. Hallucination risk drops sharply, because there’s much less need to invent.
So paste more than feels necessary. The whole email thread, not your description of it. The full feedback, not the highlights. It reads instantly and it doesn’t get bored.
What it’s genuinely good at
Summarising. Long into short, reliably, at whatever length you specify. The most common use and the most consistently useful.
Finding themes. Give it eighty pieces of customer feedback and ask what keeps coming up. This is genuinely hard for a person — you lose the thread by item twenty — and it’s what pattern-matching is built for.
Restructuring. Messy notes into a table. A rambling thread into a decision log. A transcript into action points with owners.
Explaining. Paste something you don’t understand and ask for it in plain English. Contracts, technical docs, a spreadsheet someone else built.
Spotting what’s missing. “Here’s our proposal. What questions would a sceptical client ask that we haven’t answered?” Useful precisely because it has no ego invested in the document.
Notice the pattern: these are all language tasks, applied to information. Back to Lesson 1 — its strength is words. Data made of words is where it shines.
The arithmetic problem
Now the part that catches people out.
AI cannot reliably do maths. Not because it’s a hard problem, but because — from Lesson 2 — it doesn’t calculate. It predicts. Ask it to add a column of numbers and it produces text that looks like what the answer would look like.
Often that’s correct. Sometimes it’s confidently, quietly wrong. And nothing in the output tells you which.
This is the single most dangerous thing in this lesson, because a wrong total doesn’t look wrong. Bad prose you’d catch instantly. £47,283 looks exactly as trustworthy as £47,832.
Some tools now recognise a sum and run actual code behind the scenes. When that happens the maths is genuinely reliable — it’s a calculator doing the work, not prediction. But you often can’t tell from the interface whether that’s happening, so the safe rule stands:
Never trust a number that came out of a chat box without checking it.
Formulas, not sums
Here’s the distinction that makes it useful anyway.
It’s poor at doing the calculation. It’s excellent at writing the thing that does the calculation.
❌ “Here are my sales figures. What’s the year-on-year growth?”
✅ “I have monthly sales in column B, dates in column A. Give me an Excel formula for year-on-year growth.”
The second is a language task — describing a formula. Formulas are written down all over the internet; that’s a pattern it knows deeply. And crucially, Excel then does the actual sum, so the number is right.
Same principle everywhere:
- Not “what’s the average?” but “what’s the formula for a weighted average here?”
- Not “analyse this data” but “write me a pivot table setup that would show this”
You keep the calculation in the tool built for calculating. It handles the part that’s really about language. That one habit removes most of the risk in this lesson.
Where it breaks
Big spreadsheets. Paste a thousand rows and it may work from a fraction of them without telling you. It’ll still produce a confident summary — of data it half-read.
Anything precise. Exact counts, totals, percentages. Ballpark reasoning is fine. Precision isn’t.
Recent data. Lesson 2’s cut-off. It doesn’t know your Q2 numbers, market prices, or last month’s figures unless you paste them in.
Charts and screenshots of tables. It can read them, but numbers get misread — a smudged 8 becomes a 3 and the whole analysis is quietly built on it.
Causation. It’ll happily tell you sales dropped because of your price change. It has no idea. It’s pattern-matching a plausible story, and plausible stories are exactly what you should be most suspicious of.
That last one is worth sitting with. It’s most seductive where it’s least reliable — in exactly the “why did this happen” questions you most want answered.
How to check its work
Ask for the reasoning, not the answer. Lesson 5’s habit applies doubly here. “Show me how you got that” turns an unauditable number into something you can inspect.
Spot-check. Take two claims and verify them against the source. If both hold, the rest is probably fine. If either doesn’t, throw the whole thing out — one fabrication means the process failed, not that one item slipped.
Ask it to quote. “For each point, quote the sentence it came from.” Invented claims have nothing to quote, and this flushes them out fast.
Use it to find, not to conclude. Its job is pointing you at the interesting bit. Yours is deciding what it means.
And before you paste anything: think about what’s in it. Customer names, salary data, anything confidential — that’s a real question with a real answer, and it’s Lesson 8.
Key takeaways
- Give it the actual material — it’s far more reliable working from your document than from memory
- Strongest on language tasks applied to information: summarising, themes, restructuring, explaining
- It predicts rather than calculates, so numbers from a chat box are never trustworthy without checking
- Ask for the formula, not the answer — let Excel do the sum
- It’s least reliable exactly where it’s most persuasive: explaining why something happened
- Use it to find things; decide what they mean yourself
Next: Lesson 7 – Automating Repetitive Work
You can point it at information. Next: finding the parts of your own job worth handing over.