We finally have AI that thinks slowly. Most workflows are built for fast.
Reasoning models that deliberate before answering have arrived. The enterprise question isn't capability — it's where deliberation is actually worth it.
🐢 We finally have AI that thinks slowly. Most workflows are built for fast.
This month OpenAI shipped something genuinely different: o1, a model that reasons before it answers. It takes its time, works through the problem, and — on hard tasks — produces markedly better results. It is also slower and more expensive per query than anything we've deployed before.
The instinctive enterprise reading is "better model, swap it in." I think that misses what actually changed.
Until now, every AI interaction we designed assumed near-instant response. Chat, autocomplete, summarise-on-click — the entire UX vocabulary of enterprise AI is built on speed. A model that deliberates for thirty seconds doesn't fit that vocabulary. It fits a different one: the considered draft, the overnight analysis, the second opinion you request and come back to.
Which reframes the leadership question. Not "should we upgrade?" but "which of our decisions deserve deliberation — and which just need an answer?"
The way I'd sort it:
Fast AI for high-frequency, low-stakes moments — lookups, summaries, first drafts. Latency matters more than brilliance; the human is the reasoning layer.
Slow AI for low-frequency, high-stakes moments — pursuit strategy, risk analysis, plan generation. Nobody minds waiting minutes for something they'd otherwise spend hours on.
And a design rule between them: the cost of deliberation should scale with the cost of being wrong.
The trap to avoid is paying o1 prices for autocomplete problems — which is exactly what happens when "use the best model" becomes policy instead of architecture.
There's a quieter implication too: models that can genuinely work through multi-step problems are the missing ingredient for agents that plan rather than react. What ships as a chat feature this month is really the engine of next year's autonomous systems. Worth positioning for now.
Try this with your leadership team: name the three decisions in your organisation where you would gladly wait five minutes for a materially better answer. Those three are where this generation of AI belongs. If you can't name them, the problem isn't the model's speed — it's that you haven't mapped your decisions by stakes.
Reading: https://openai.com/index/introducing-openai-o1-preview/
https://www.nytimes.com/2024/09/12/technology/openai-chatgpt-math.html