Is Claude Accurate? How to Check Its Work Before a Customer Sees It
10 min read
Claude is accurate enough to draft most business writing and wrong often enough that numbers, dates, quotes and legal points need a check before they leave the building. This lesson shows where the errors cluster and four checks that catch them.
Claude's own interface carries a line saying it can make mistakes and asking you to double-check responses. Take that at its word. The errors aren't random, though, and once you see where they collect, checking takes minutes.
Where Claude gets it wrong
Missing facts get filled in. Ask for a reply about your warranty without supplying the warranty, and you get a plausible one. The writing is so smooth that the invention is hard to spot. This is the most common failure and the easiest to prevent.
Arithmetic over long lists. A total of eight numbers in a sentence is usually right. A total of 300 rows worked out "in its head" may not be. With code execution on, Claude can calculate in a real spreadsheet or script, and those results are far more dependable. Ask for that.
Stale knowledge. What Claude learned in training has a cut-off date. Prices, plan limits, tax thresholds and regulations change after it. Unless it searched the web and shows you the source, treat a current fact as a guess.
Summaries that drop the clause that mattered. A 12-page lease summarized to ten bullets reads well and may omit the renewal notice period. Summaries are for orientation. Decisions need the original.
Confident tone on shaky ground. Claude doesn't sound less sure when it is less sure, unless you ask.
Three lanes for what it writes
| Lane | Examples | What you do |
|---|---|---|
| Send after a read | Internal notes, first drafts, brainstorms | Skim for tone |
| Check the facts | Customer emails, quotes, posts, reports | Verify each number, date, name and promise |
| A professional decides | Tax treatment, contracts, HR actions, safety | Claude prepares the questions. A licensed person answers |
Most trouble comes from treating lane two like lane one. Our advanced course has a full lesson on the three lanes of trust.
Four checks that take minutes
1. Make it show its sources. Add this to a request that involves facts from outside your business:
For every factual claim, give the source link. If you
are relying on memory and did not look it up, label the
claim UNVERIFIED.Then open two of the links. A source that doesn't say what Claude says it does is your signal to check the rest.
2. Make it do the math in a file.
Do not calculate this in your head. Put the figures in
a spreadsheet, use formulas, and give me the file so I
can see how each total was reached.3. Ask what it's unsure of.
Before I use this: list the three things in your answer
you are least sure about, and what I should check to
confirm each one.This one is worth memorizing. The answers are often specific and useful: "I assumed the 14-day window starts at delivery. Your policy file doesn't say."
4. Get a second reading in a fresh chat. Open a new chat, outside the Project, and paste the draft with this:
You are reviewing a draft before it goes to a customer.
Here are the source facts: [PASTE]. Here is the draft:
[PASTE]. List every statement in the draft that the
source facts do not support, every number that differs,
and every promise made. Do not rewrite it.A fresh chat has no stake in the first answer. It catches things the original conversation talked itself past.
The line that prevents most of it
You've seen it in the prompts throughout this course:
Use only the facts I have given you. If something is
missing, ask me or leave a [BLANK]. Do not guess.Put it in your account instructions and in each Project. A blank costs you ten seconds. A guess that reaches a customer costs you the gap between what Claude wrote and what you can deliver.
You said it, whoever wrote it
In 2024 a British Columbia tribunal held Air Canada responsible for what its website chatbot told a customer about bereavement fares. The airline argued the bot was responsible for its own words, and lost. The amount was about $800. The principle is the part to remember: a customer can hold you to what your business sent, and "the AI wrote it" is no defence.
For catching the errors that look most convincing, see when the AI gets it wrong. To audit replies already sent, the Support Reply Auditor grades ten of them against your policy.
Where checking stops being enough
If you find yourself verifying each line of a certain kind of output, Claude is the wrong tool for that job, or the job needs better inputs. A quote built from a rate card file is checkable in a minute. A quote built from Claude's general sense of what landscaping costs can't be rescued by checking. Fix the input.
Try it now
Take the last thing Claude wrote for you that went to a customer. Run check 3 and check 4 on it. If either turns up something you missed, add the guard line to that Project's instructions today.