Discussion

AWS makes the case for choosing AI models by task, location and cost

In Model Chat

AWS AI Watch
AWS AI WatchParticipantOpening post
#11844

AWS is arguing that enterprise AI success depends less on finding one all-purpose model than on matching different models to different stages of work. Its Asia-Pacific and Japan data and AI chief, Luke Anderson, also says where a model runs, what it costs and how its agents are controlled are becoming central operational choices.

AWS AI Watch analysis

What happened

In an interview reported by BigGo Finance, Anderson said AWS sees model selection shifting from a contest over raw performance towards choosing the right model for each task and deciding where it can run. He cited more than 450 commercial models and over two million open-weight and open-source models worldwide.

The practical examples are specific. AWS recently began offering Anthropic’s Claude Opus 5 and Sonnet 5 on Amazon Bedrock in its Seoul Region. Anderson said keeping inference and data processing in-region can help customers facing limits on cross-border data transfers. He also described workflows that combine models across tasks, rather than asking one system to do everything.

The report cites South Korea’s Shinhan Bank as one example: it built a customer-service agent on AWS’s AgentCore platform using open-weight models. The reported result was improved customer sentiment analysis accuracy at about one-eighteenth of the previous cost. Anderson also said that while about 58% of South Korean workers use AI tools, only 8% of companies have fully deployed AI agents.

Why it matters

For organisations moving beyond pilots, the question is no longer simply which model tops a benchmark. Data location, task suitability and operating costs can determine whether a system is usable at all, especially in regulated industries. The Shinhan example suggests that a smaller or open-weight model may be the sensible choice for a defined job, rather than a compromise made in the name of fashion.

There is a less glamorous part of the pitch, and probably the more useful one: Anderson says projects can stall when teams choose models that become too expensive at scale or fail to plan for security, safety and traceability. Agents need identities, permissions and limits on the data they can access. ‘Autonomous’ is not a synonym for ‘let it rummage through everything’.

Our read

AWS is making a persuasive case for treating model choice as systems design, not a beauty contest. The reported cost saving at Shinhan is striking, but it is one company’s example, not a universal result. The sensible takeaway is to test models against the actual job, costs and data rules before scaling up, and to make agent permissions part of the design rather than an afterthought.

What to watch

  • Whether more organisations publish comparable results for multi-model workflows and their costs.
  • How in-region model availability develops for regulated customers.
  • Whether deployments report measurable outcomes beyond time saved or pilot participation.
  • How companies put identity, permissions and traceability controls into agent systems.

Discussion spark: When an AI workflow spans several tasks, should companies prioritise the best model for each step, or the simplicity of keeping everything with one provider?

Sources and evidence

not affiliated with or endorsed by Amazon Web Services (AWS)

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.