AI for business without the magic: four tasks where it pays off

In two years AI went from "this replaces everyone" to "this writes copy nobody reads". The truth sits in the middle and is rather dull: a model does a few specific things well, and if you give it exactly those, it pays for itself within a month.

1. Triaging what comes in

Email, enquiries, messenger threads. The model reads the text and answers the questions a human spends ten seconds on a hundred times a day: order or question, which product, urgent or not, happy or angry.

This pays first because mistakes are cheap: a mislabelled letter is one click to fix, and the ninety percent labelled correctly save an hour a day.

2. A draft a human edits

A reply to a client, a product description, a follow-up after a meeting. The word that matters is draft: sending without reading is how you end up apologising for a text you did not write.

The rule is simple: AI prepares, a person sends. The moment the person stops reading, you start sending polite nonsense.

3. Search across your own documents

Contracts, instructions, correspondence, internal knowledge. Ordinary search looks for words; a model looks for meaning. "What did we promise this client about deadlines" is found even in a letter that never says deadline.

4. Labelling and tidying data

• Sorting a thousand companies into niches
• Pulling contacts and addresses out of unstructured lists
• Merging duplicate clients written down differently
• Drafting a second-language version of a catalogue

Where it does not pay off

• Where mistakes are expensive and nobody checks: money, legal promises, medical advice
• Where there is no data: the model does not know your prices unless you give them
• Where a rule would do: if the condition fits in three ifs, you do not need a model

What it actually costs

Do not trust "AI costs pennies". Count it: requests per month times text size times the price per thousand tokens. Triaging mail for a small company comes to a few dollars; generating catalogue pages runs into tens. The spending cap belongs in the settings before the first run, not after the first invoice.

One more thing. Any sane AI integration has an off switch. If the process stops without the model, you have not built an assistant — you have built a single point of failure that occasionally invents things.