Some questions you answer in a blink. What colour is the sky? You do not need to think.

Other questions need a little walk in your head. Say your friend's party starts at 4:30. You leave home at 3:40.

The drive takes 55 minutes. Will you be on time?

You cannot just blurt an answer. You run the trip forward in your mind. Twenty minutes after 3:40 it is 4:00. Another 35 minutes makes it 4:35.

You will be five minutes late. Notice what you did. You imagined the clock moving, one step at a time. Then you checked the answer before you said it.

Maybe you had a second idea, like leaving at 3:30. You could test that one too.

Some new computer helpers work this way. People call them reasoning models. Instead of answering at once, they write out steps. They try a few routes and check their own work.

Thinking longer costs them something, just like it costs you. Each extra step takes time and electricity. So the clever part is knowing when to think hard.

Nobody needs ten minutes of thinking for the colour of the sky. A tricky party trip deserves a careful walk through.

In a way, these models use a world model of the problem. They imagine what happens at each step before they commit.

Try this at home or at school. Find a maze in a puzzle book. Before you draw any line, trace two possible paths with your finger.

Notice the dead ends you avoided. Then draw the path once, without rubbing anything out.

Chain-of-thought prompting showed that eliciting intermediate steps improves multi-step reasoning (Wei et al., 2022). A 540B model with eight worked exemplars set a new mark on GSM8K. It beat a fine-tuned GPT-3 that had a verifier.

Self-consistency samples many chains and votes on the final answer (Wang et al., 2022). That alone added 17.9 points on GSM8K.

Tree of Thoughts turned the chain into explicit search (Yao et al., 2023). On the Game of 24, GPT-4 with chain of thought solved 4% of tasks. With search over partial thoughts it solved 74%.

Reasoning-trained models move the skill into the weights. DeepSeek-R1 (DeepSeek-AI, 2025) learned to reason through reinforcement learning on verifiable tasks. Self-reflection and verification emerged without human-labelled reasoning traces.

Performance then scales with test-time compute. You can buy longer chains, more samples with voting or search with a verifier.

Process supervision makes the verifier sharper. One process-supervised model solved 78% of a representative subset of MATH (Lightman et al., 2023). It trained on 800,000 step-level labels.

The allocation matters as much as the amount. Compute-optimal allocation beat best-of-N by over four times in efficiency (Snell et al., 2024). In a FLOPs-matched test, extra thinking beat a model 14 times larger. That held only where the small model already had some success.

The link to world models is structural. Planning by search needs a model to predict each step. It also needs a value to judge each step. In reasoning, the world is the problem state and the model is the network itself.

Recurrent architectures make the same trade explicit. One 3.5B model iterates a recurrent block at test time (Geiping et al., 2025). Its reasoning scores keep rising up to a compute load equal to about 50B parameters.

The failure modes are known. Chains can look careful and still rest on a wrong premise. Verifiers reward what they can check, which may differ from what matters. Long deliberation on easy inputs burns money without changing the answer.

Take a high-bill complaint. Many questions are routine, such as a balance or a due date. A fast model answers those.

A rebill after a meter exchange is different. It may involve estimated reads and a rate change in the same cycle.

There, I would recommend deliberation with a hard verifier. The model proposes hypotheses for the high bill and ranks them. The rating engine recomputes each candidate bill deterministically. Only a candidate that reconciles goes to a human for approval.

Measure accuracy as a function of thinking tokens per task class. Find the knee where the curve flattens. Track the share of answers that deliberation actually changed.

Put a price on each decision. A token count alone says little.

Reasoning models let you buy accuracy with time and money, one question at a time. The model thinks longer before it answers. You pay for every step of that thinking.

That turns routing into a business decision. Routine questions deserve a cheap, fast answer. Consequential decisions may deserve expensive deliberation.

Most questions in a contact centre are routine. A rebill or a disconnect decision is not.

The research supports a careful view. Extra thinking can beat a much larger model on some problems. On others it changes nothing and still costs money. One study found smart allocation over four times more efficient than brute repetition.

The risk runs both ways. Under-thinking on a consequential case produces a wrong rebill and a complaint. Over-thinking on routine traffic quietly multiplies the model bill. Neither shows up until someone measures it.

I would recommend three steps:

• sort your agent's work into task classes by consequence
• measure accuracy against thinking cost for each class
• route by that evidence and review it every quarter

Pair deliberation with a hard check wherever one exists. A bill can be recomputed. A switching plan can be checked against the network model. A deliberating model with a verifier is far safer than one without.

The money story is a budget per decision. A rebill that takes a minute of model time and avoids a complaint is cheap. A balance query that takes the same minute is waste.

So ask your team one question. On which of our decisions does extra thinking change the answer?