Chain of thought

What is chain of thought?

If someone asks you to work out in your head the price of an order with a volume discount, VAT and delivery, you’ll get it right more often going step by step than blurting out a number. Language models are similar. They produce text one token at a time, and every word they write supports the next one. If they write out the steps first, the conclusion rests on them.

That’s chain of thought: getting the model to reason in writing before it answers. You can ask for it in the prompt (“think step by step”) or show it with examples that include the reasoning, combining it with few-shot prompting. Today’s reasoning models do it by default, without being asked.

How it works

1. The model receives the question and the instruction to reason.
2. It writes out the intermediate steps: data, relevant rules, calculations.
3. It gives the final answer, kept separate from the reasoning.

In an application, step 2 usually stays behind the scenes: it’s kept for review, but the user only gets step 3.

It comes at a cost. The reasoning takes up tokens, which you pay for, and takes time to generate. In a written chat, a few extra seconds don’t matter. On a phone call they do: silence is noticeable. So in voice it’s kept for what really needs it.

And it isn’t foolproof. Well-written reasoning can start from a wrong fact. It reduces errors; it doesn’t eliminate them.

Example

A van hire firm in Glasgow uses AI to prepare quotes that a member of staff then checks. Rates vary by number of days, mileage and season. Without reasoning, the model sometimes gets the summer surcharge wrong. With this prompt it makes fewer mistakes, and when it does, they’re easy to spot:

Work out the quote using the attached rate table.
Before the total, write out the steps:
1) days and season, 2) daily rate, 3) extra miles, 4) surcharges.
Then write "TOTAL:" and the amount.
If any information is missing, don't assume it: say what's missing.

The person checking doesn’t have to redo the sums: they just check the steps.

What it means for your business

For tasks with several rules in a chain, asking for the steps improves the result and makes checking easier. For quick replies to customers, it’s unnecessary.

With UTIO’s agents, replies to customers are direct and, especially with the voice agent, they need to arrive without delay. Anything unclear isn’t worked out blindly: the agent starts in draft mode and can hand the conversation to your team under rules you set.