Automation and AI

Chain of Thought

Also called CoT, step-by-step prompting

Asking a model to work through its reasoning in steps before answering, which helps on tasks with several dependent stages.

Quick facts: Chain of Thought

Category
Automation and AI
Also called
CoT, step-by-step prompting
Level
Intermediate
Affects
Accuracy on multi-step tasks, review time, cost, response length
Where to see it
Chat assistants, reasoning models, model APIs
In this article4
  1. How chain of thought works
  2. Why chain of thought matters
  3. Common mistakes with chain of thought
  4. How to act on it

How chain of thought works

A language model builds its answer one piece at a time, and each piece is influenced by what it has already written. Ask for the conclusion immediately and it commits early, with nothing in between to keep it honest. Ask it to lay out the steps first — identify the figures, state the method, work through them, then conclude — and each step becomes context for the next, which tends to improve results on anything with several dependent stages.

In practice it is triggered by a plain instruction: work through this step by step, or explain your reasoning before answering. Some newer models do this internally by default and show you only a summary, so an explicit request adds less than it once did. It is worth checking what your tool already does before adding the instruction to everything.

Why chain of thought matters

Marketing questions that go wrong quietly are usually multi-step. Working out which campaigns to cut when budget and returns are both uneven, reconciling numbers from two reporting tools that count differently, or deciding whether a change in enquiries could plausibly be seasonal all need several linked judgements. A single-shot answer to any of those sounds authoritative and hides where it went astray.

The visible working is the practical benefit. When you can see which figure was used, which assumption was made and where the leap happened, you can correct one step instead of rejecting the whole answer. That makes the output reviewable by someone who knows the business but not the tool.

Common mistakes with chain of thought

Reading the explanation as proof is the important one. The written steps are generated text like everything else, and they are an account that fits the answer rather than a record of how it was produced. A model can present tidy, plausible reasoning and still reach a wrong conclusion, so the working helps you find errors but never guarantees there are none.

Using it everywhere is the second. Longer answers cost more and take longer, and on simple tasks — rewriting a sentence, sorting an enquiry, drafting a subject line — the extra reasoning adds expense without accuracy.

The third is leaving the working in the deliverable. Reasoning is for your review, not for the customer, and it should be stripped before anything is published or sent.

How to act on it

Reserve it for tasks with genuine steps, and say what the steps are rather than leaving the model to invent a structure. Naming the sequence you want — check the figures, state the assumptions, compare the options, then recommend — gives you something consistent to review across different questions.

Read the working the way you would read a colleague’s calculation: check the inputs first, then the assumptions, then the arithmetic. Verify any figure it introduces against your own source, because a confident wrong number inside a tidy explanation is the classic route to a hallucination reaching a client report. For recurring analysis, a fixed set of steps in a saved template is more reliable than asking freshly each time, and it fits neatly into an automated reporting process.

Do and do not

Do

  • Name the steps you want rather than leaving them open
  • Check every figure the working introduces against your own source
  • Ask for a clean final version once you have reviewed

Do not

  • Treat the stated reasoning as proof the answer is right
  • Apply it to simple rewriting or sorting tasks
  • Leave the working in a client-facing document

Questions people ask about this

Does asking a model to think step by step actually make it more accurate?

On tasks with several dependent stages it generally helps, because each step conditions the next instead of the model committing to a conclusion straight away. On simple rewriting or classification it adds length and cost without improving anything. Newer reasoning models already work this way internally, so test whether the instruction changes your results before applying it everywhere.

Can I trust the reasoning a model shows me?

Treat it as a useful account rather than a faithful record. The steps are generated in the same way as the answer, and a model can produce convincing reasoning that supports a wrong conclusion. The value is that it makes errors findable: you can check which figures were used and which assumptions were made, then correct one step rather than discarding everything.

Should the reasoning appear in what I send to a client?

No. It is working, not output. Ask for the reasoning while you review, then request a clean final version, or remove the steps yourself before sending. Leaving them in makes a report longer and harder to read, and exposes assumptions you may not have checked. Keep a copy for your own records if the decision matters.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.