Amazon Quick drops token pricing for ‘agent hours’ — but does one hour really mean an hour?

Amazon Web Services (AWS), the cloud arm of Amazon (Nasdaq: AMZN), announced earlier this month that the desktop version of its AI workplace assistant Amazon Quick was officially available for macOS and Windows. Quick does not charge by tokens. Instead, it measures usage in “AI agent hours” — the time a task spends invoking models in the background, calculated by the second, according to Guo Ren, AWS’s director of AI technology in its product division, who spoke to mainland Chinese outlet Jiemian News.

AWS has also added an Activity Feed to the iOS and Android versions. It combines email, calendars, customer relationship management (CRM) systems and messaging into a prioritised view. Routine matters that the AI agent can handle automatically disappear from the feed, leaving only items requiring human judgement. The desktop version can read local files and connect to the workplace applications users rely on daily. Jiemian News reported that AWS said the product integrates more than 60 officially configured tools and supports integrations with more than 1,000 applications.

Pricing structure

According to AWS’s pricing page, individual plans are divided into Free, Plus and Max tiers. Plus costs US$20 per user per month when billed annually, or US$25 on a monthly basis, and supports up to 300 users. Max costs US$100 per user per month, or US$125 when billed monthly.

Organisations must register through an AWS account. The Professional plan costs US$20 per user per month, while the Enterprise plan costs US$40. Both also incur a basic infrastructure fee of US$250 per account per month. The included allowance is eight agent hours per month for Professional and 18 hours for Enterprise.

The two plans differ when users exceed their allowance. Professional customers currently cannot purchase additional usage. Enterprise customers can do so at US$3 per agent hour, with administrators able to set a usage limit. If the organisation exceeds its pooled allowance for index storage, the charge is US$5 per GB per month.

How is an “hour” measured?

AWS’s frequently asked questions page says time is counted only while Quick is actively processing a task; idle time is excluded. Actual usage, however, depends on the complexity of the task and the selected mode — Fast, Balanced or Smart. The deeper the reasoning, the faster the consumption. A complex task that takes 15 minutes could use more than 15 minutes of agent time. Time spent waiting for an external tool to respond, such as when writing to Salesforce, will generally also be counted. Scheduled tasks reduce the allowance each time they run.

In other words, an agent hour is not the same as an hour on the clock. Guo said the calculation is based on the time models are actually invoked in the background, rather than the time experienced by the user on the front end.

As for actual usage, Guo said all 340,000 Amazon employees use Quick, averaging about 10 agent hours per person each month. He said this could reflect usage at most enterprises. On that basis, a company with similar usage would exceed the Professional plan’s monthly allowance of eight hours, while that plan does not permit additional usage. The Enterprise allowance of 18 hours would provide more headroom. This is an estimate based on two publicly disclosed figures; AWS has not made this comparison itself.

Customers and security

Justin Bundick, vice-president of the technology intelligence platform at Southwest Airlines, which has more than 70,000 employees, said Quick enabled teams to ask complex questions in natural language and receive evidence-based answers within seconds. It is already being used for tasks including market analysis. Chuck Metturdharma, vice-president and chief AI officer at medical diagnostics group Labcorp, said Quick helped teams focus their efforts on higher-value work. Both comments were published by AWS as customer testimonials.

On security, AWS said Quick runs on AWS infrastructure, with data remaining in the customer’s own environment. It also includes compliance certifications covering HIPAA for medical data, FedRAMP for the US federal government, SOC 2 and ISO 27001, in an effort to address the risk of employees privately using unapproved AI tools — known as “shadow AI”. These are all vendor self-declarations.

Questions remain

Guo said Quick’s usage-based pricing makes it cheaper than other comparable products on the market. Jiemian News did not provide specific comparative figures. Three issues will be worth watching: whether enterprise usage generally exceeds the included allowance; whether rivals such as Microsoft and Google follow with time-based pricing; and how deeply Quick integrates with companies’ existing ERP systems and private code repositories.