What you can really do with Claude
(and what you can't)
Concrete examples: summarising documents, answering recurring emails, extracting data from PDFs, supporting customer care. Where it works well today and where it's better to wait.
9 minute read
The quickest way to lose confidence in a tool is to use it for the wrong thing. Language models like Claude are extraordinary at some tasks and unreliable at others, and the line separating the two is fairly clear. This guide draws that line with concrete examples, and without promises.
Transforming text placed in front of them,
not recalling facts from memory
Language models are excellent at transforming text that is already in front of them, and fragile when they have to recall precise facts that you have not provided. Translating, summarising, rephrasing, extracting, classifying: these are all transformations of present text, and models excel here. "What is the exact price of our product X?": that is a fact, and if you don't supply it the model may invent it with disarming confidence.
If the information is in the text you provide, trust it a great deal. If you are asking the model to retrieve it from its own memory, always verify.
Five use cases
with high returns
Summarising long documents
Tender specifications, contracts, meeting minutes, technical articles, product sheets. Provide the document and ask for a structured summary, the key points, deadlines, and unusual clauses. It works because all the material is in front of the model. It is most useful as a first filter — surfacing what deserves careful reading — rather than a replacement for reading when the stakes are high.
Answering recurring emails
Information requests, confirmations, follow-ups, standard replies in multiple languages. The model drafts a response consistent with your tone from just a few elements. The right way to use it: generate the draft, have a person review it, then send. The time saving is real and errors stay under control because there is always a human eye before anything is sent.
Extracting data from PDFs
Invoices, delivery notes, orders, transport documents, supplier quotes. The model reads a disorganised PDF and returns the fields you need in a structured form (number, date, amount, line items). This is one of the highest-return use cases for SMEs, because it replaces hours of manual copy-and-paste. Two caveats: spot-checking numerical data is essential, and results deteriorate on low-quality scanned documents.
Supporting (not replacing) customer care
Draft replies for agents, ticket routing and classification, summaries of a customer's history before a callback, translation of requests in other languages. Where it works well: as an assistant to the agent. Where to tread carefully: as a fully autonomous chatbot facing the customer, without supervision, on topics where a mistake has consequences.
Translating and adapting content
Business correspondence, marketing materials, technical data sheets across working languages. Quality is high, especially if you provide a glossary of your technical terms. For legal or contractual texts, review by someone who knows the subject matter in the target language remains necessary.
Good candidates
vs cases to defer
| Good candidates | Better to wait / do not rely on alone |
|---|---|
| Summarising, rephrasing, changing register | Precise facts not provided (prices, dates, references) |
| Extracting fields from documents you supply | Exact calculations and accounting |
| Classifying and routing text | Automated decisions without supervision |
| Translating with a glossary | Legal or tax advice as a sole source |
| Generating drafts reviewed by a person | Citations or regulatory references recalled from memory |
| Comparing two documents and flagging differences | Anything where an error is serious and unverifiable |
The limitations
to keep in mind
Calculations and exact numbers
A language model is neither a calculator nor a spreadsheet. It can get a sum or a conversion wrong and present it with complete confidence. For exact calculations, let the right software do the arithmetic: the model at most prepares the data or explains the result, it does not compute it.
"How reliable is this?" cannot be read from the response
The most insidious problem: a model can produce incorrect information in exactly the same confident tone as a correct one. There is no "traffic light" in the response. This is why the winning use cases are those where a person can verify in a few seconds, or where the cost of an error is low.
High-risk automated decisions
Approving a credit line, rejecting a candidate, sending a legal notice: these are decisions where full automation is, today, unwise and in some cases sensitive from a regulatory standpoint. The correct use is in support: the model prepares, the human decides.
Four moves
that make the difference
- Provide the context, don't ask for it from memory. Attach the document, the price list, the instructions. The more the model works on material you give it, the more reliable it is.
- Keep a person in the loop at the start. Trust is built by seeing that the outputs are correct across dozens of real cases — not assumed in advance.
- Ask for a verifiable format. If the output must contain specific fields, request them explicitly and check that they are all present before using them.
- Measure the "right first time" rate. It is the only number that tells you whether you can reduce human oversight.
In one sentence
Use Claude to transform text and data you place in front of it, with a person validating until you trust the numbers. For precise facts, exact calculations, and high-risk decisions, the tool is an assistant, not the final arbiter.
Frequently asked
questions
Is Claude better than ChatGPT?
It depends on the use case. On long-form writing and document analysis tasks, Claude tends to have a more measured, less "marketing" tone; ChatGPT is more versatile on creative tasks and has a broader plugin ecosystem. For typical SME work (summarising, extracting, classifying) they are broadly equivalent. The real difference is made by the prompt, not the model.
Can Claude read my Excel/PDF files?
Yes, both via chat (by uploading the file directly) and via API. On well-formed text-based PDFs, extraction is accurate; on low-quality scanned PDFs, results deteriorate and an upstream OCR step is needed. On Excel it works well when sheets are "tabular" and less well when there are merged cells and cross-referenced formulae.
Can I trust Claude for legal or accounting documents?
As an assistant, yes; as a sole source, no. It is perfectly suitable for summarising a contract, flagging unusual clauses, or preparing a draft. But regulatory citations recalled from memory can be wrong, and exact calculations should not be left to the model. Review by a professional who knows the subject in the target language remains indispensable.
How often does Claude make mistakes on repetitive tasks such as data extraction?
On well-defined tasks with clean inputs, in an SME context, the "right first time" quality typically ranges between 85% and 98%. The remaining 2–15% justifies human oversight in the early stages. Measurement is the only way to know whether you can reduce supervision: measure it.
Does Claude work well in Italian?
Very well, both in comprehension and generation. Italian tone is natural and the model handles formal/informal register, technical terminology, and commercial language correctly. For niche specialist terminology, it is worth providing a glossary in the prompt.
Would you like to talk through
your specific case?
A 30-minute call to get your bearings. No pre-packaged demos.
Write to us at [email protected]