Late last month Anthropic released Claude 3.7 Sonnet, the first mainstream model that can either answer immediately or visibly "think" through a problem first – same model, your choice, reasoning on display. OpenAI's o-series has been pushing the same direction for months. Whatever tool your organisation has standardised on, reasoning models are about to be in it, and your staff are about to start reading AI chains of thought whether anyone planned for that or not.
We've spent the past week running these models through the exercises we use in training, and the change is real. So is the new failure mode.
What actually improvedLink to this section
Reasoning models are meaningfully better at the tasks knowledge workers were already stretching chatbots toward: multi-step analysis, working through a policy against a scenario, maths that has to be right, planning with dependencies. Problems that used to need a human to break into pieces can now often be handed over whole.
That expands what people will delegate. Good – that's the productivity your organisation is paying for. But it also moves AI use up the stakes ladder, from "draft this email" toward "assess this contract clause", and the checking habits that were adequate for emails don't survive the promotion.
The new failure mode: a persuasive traceLink to this section
Here's the subtle bit. When a model shows its reasoning, people trust the answer more. I watched it think – it considered the edge cases – it double-checked itself. The trace reads like a colleague's working, so we extend it a colleague's credibility.
Two problems with that:
- The trace is also generated text. It reads as an explanation, but research keeps showing that a model's stated reasoning doesn't always reflect what actually drove the answer. It can be a plausible story about a conclusion rather than the path to it.
- Fluent working amplifies confident errors. A wrong answer with no explanation invites a check. A wrong answer wrapped in three paragraphs of tidy, self-assured reasoning invites a nod. The failure rate may drop; the detection rate of the remaining failures drops further.
The one-line version we now teach:
The working is a claim too. Verify the answer, not the vibe of the explanation.
Verification habits that match the taskLink to this section
Blanket "always double-check AI" advice fails because it's unactionable – check what, exactly? We teach verification matched to task type, and reasoning models don't change the table, they just raise the stakes:
| Task | The check that matters |
|---|---|
| Summarising or extracting | Spot-check claims against the source document |
| Analysis or recommendation | Verify the inputs it relied on; re-derive one branch yourself |
| Anything numerical | Recompute the number that will end up in front of others |
| Plans and sequences | Attack the assumptions; ask what happens if step two slips |
| Cited facts | Open the citation. Every time. |
The discipline scales with consequence, not with how impressive the output looks. A reasoning trace can help you locate an error faster – read it to find the load-bearing assumption, then test that assumption in the real world. That's the productive use of visible thinking: a map for your scepticism, not a substitute for it.
Two practical notes for rolloutLink to this section
Cost and patience. Extended thinking spends time and tokens. Staff need a rough rule for when it's worth it: hard, consequential, multi-step problems, yes; reformatting a paragraph, no. Otherwise you get the worst of both worlds – slow answers to trivial questions and quick answers to hard ones.
Policy is unchanged, and say so. A smarter model does not change what data may be pasted into it. We've already seen the reasoning "this model is better, so it must be fine to give it the client file" in a training room. Capability and data-handling are separate axes; make sure your acceptable- use guidance says which tools are approved for which data, whatever their IQ.
The skill that lastsLink to this section
Models will keep leapfrogging each other – this year's reasoning novelty is next year's default setting. The durable workforce skill isn't operating any particular model; it's calibrated trust: knowing what to delegate, what to verify, and how much checking a given consequence deserves.
That's trainable. It's most of what we do. And it's worth building now, while "the AI showed me its reasoning" is still a sentence that makes your team pause rather than one they've stopped noticing.