What AI can and cannot do in a company
6 min read
Short answer
Language models are reliable at reading and sorting text: pulling data out of an invoice or an order, triaging the inbox, drafting a reply, answering questions from your own documents. They are unreliable anywhere that requires arithmetic, a rule kept without exception, or a decision whose consequence nobody will read.
01Where it works well
Effectively anywhere the model reads text and a person sees the result before anything happens.
- 01Extracting data from documents — invoices, delivery notes, orders, contracts. Accuracy is high and checking is cheap.
- 02Sorting what comes in — which message is an enquiry, which a complaint, which nothing.
- 03Drafting a reply that a person reads before it goes out.
- 04Answering from your own material — policies, manuals, price lists — with a link to the source.
- 05Summarising long text while the original is still at hand.
02Where it must not run alone
Anywhere a mistake is expensive and nobody sees it until it lands.
Arithmetic is the most common error. A model should not calculate prices, margins or discounts — it should pull the numbers out and leave the calculation to code that returns the same answer every time.
The second is unsupervised outbound communication. A reply to a customer that nobody read is a commitment made by something that does not know what it promised.
The third is decisions with legal or financial consequences. Approvals, cancellations, payments. The preparation yes, the decision no.
03Assistant versus automation
An assistant helps somebody and the person stays in the loop — writing, correcting, sending. Low risk, immediate benefit.
Automation runs on its own and a person steps in only for exceptions. That is the bigger saving, and it demands what cannot be avoided: rules for what happens when the system is unsure, and a record of every step.
Most companies should start with an assistant. It is cheaper, it proves itself in weeks, and it also shows whether the data exists in a usable form at all.
04What to check before it goes live
Where the data is and who may see it. A model answering from internal policies needs somewhere to read them — and must not answer someone who has no right to them.
Where the data goes. If you process personal data, you need to know which provider sees it and whether it is used for training. That is a question for the contract, not the sales deck.
How you will know it stopped working. Automation does not break loudly; it breaks quietly and nobody notices for a month. Measure how many cases reach a person, and when that number jumps, something changed.
Frequently asked
- What is AI actually good for in a business?
- Reading and sorting text — extracting data from invoices and orders, triaging the inbox, drafting replies, answering from your own documents. Anywhere a person still sees the result.
- Can AI calculate prices or margins?
- It should not. The model should extract the numbers from the document and leave the calculation to code that returns the same result every time.
- Should AI reply to customers directly?
- Only if a person reads the reply before it goes out, or if the questions are tightly bounded. An unchecked reply is a commitment nobody made.