Amazon Unveils Q, Its Generative AI Assistant for Businesses

On November 28, 2023, at its re:Invent conference, Amazon unveiled Q, a corporate chatbot built into AWS that answers questions, drafts documents, and helps developers using each company's internal data.

Amazon wants every employee at every AWS customer to have an AI assistant that knows their documents, their systems, and how they work. That's the pitch behind Amazon Q, unveiled today by AWS CEO Adam Selipsky on the re:Invent stage, Amazon's annual cloud conference in Las Vegas.

A Chatbot That Knows the Company's Data

Amazon Q is a conversational assistant built for the corporate world, not the home user. It connects to the tools an organization already relies on — Amazon has cited dozens of integrations with systems like Salesforce, ServiceNow, Slack, Google Drive, Gmail, Microsoft 365, and Zendesk, among others — to answer questions, summarize documents, draft text, and generate content based on internal information rather than generic knowledge pulled from the internet.

The core promise is personalization: instead of a model that answers with general knowledge, Q adapts to each AWS customer's data, processes, and policies, while respecting whatever access permissions the company already has in place. An employee without access to a given document won't be able to ask Q to summarize it.

Woven Into All of AWS, Not Just a Standalone Product

Amazon Q doesn't arrive as a separate app — it's woven throughout the AWS catalog. It shows up inside QuickSight, the data analytics service, to generate visualizations and plain-language explanations; inside Connect, the customer service platform, to assist agents in real time; and inside the AWS management console itself, to help answer technical questions about cloud infrastructure.

For developers, Amazon showed off a capability that expands on what CodeWhisperer, its coding assistant, already offered: the ability to automate code migrations, such as updating applications written in older versions of Java to more recent ones — a task that traditionally takes weeks of manual work.

Amazon Q is built on Amazon Bedrock, the company's platform that provides access to multiple foundation models — including its own Titan models and third-party ones like Anthropic's — rather than relying on a single closed model.

Amazon's Answer to Copilot and Gemini for Business

The launch puts Amazon in the same race Microsoft has been running for months with Copilot, built into Office 365 and GitHub, and that Google is pursuing with Workspace's generative AI features. Amazon's angle is different because of its position as an infrastructure provider: AWS doesn't sell its own office suite or search engine, so its edge lies in connecting Q to the business systems its customers already run in the cloud, whatever those systems may be.

It's also a defensive move. AWS dominates the cloud computing market, but it arrived later than Microsoft and Google to the public conversation around generative AI. With Q, the company is trying to show that its infrastructure advantage — the same servers where thousands of companies' data already lives — translates into a more useful assistant than the competition's, precisely because it starts from private data rather than a generic model.

Pricing and Availability

Amazon announced a starting price of $20 per user per month for Amazon Q Business, the version aimed at non-technical employees. The service is available in preview starting today for AWS customers, with plans to progressively expand integrations and add features geared toward developers. Source.

It remains to be seen whether companies will trust an AI assistant with broad access to internal documents, contracts, or support conversations, and how Amazon will manage the risk of the model generating wrong answers with the same apparent confidence as correct ones — a problem no language-model-based assistant has fully solved yet.

The triangle that determines whether an enterprise assistant is useful

A corporate chatbot combines three systems that should be audited separately. The first retrieves documents: it must find the relevant passage and preserve provenance. The second writes an answer from that material. The third, when enabled, takes an action. Each layer introduces a different failure. An answer may sound right while relying on the wrong document; a correctly specified action may still lack authorisation.

AWS says in the preview announcement that Q provides more than 40 connectors, honours the user's access controls and displays references. Those are testable properties, not a universal guarantee. A customer should verify that a connector synchronises deletions and permission changes, that a reference leads to the passage actually used and that the system abstains when documents conflict.

A sensible test uses accounts with different privileges and a small set of questions whose answers the team already knows. Include a current document, an obsolete one and a file the test user cannot open. The assistant should cite the current source, acknowledge uncertainty and neither reveal nor summarise the restricted file. Then revoke a permission and repeat the query: an index that updates slowly can turn a correct control into a temporary leak.

Answering and acting carry different risks

When the assistant only proposes text, a person can still correct it. When it creates a ticket, changes infrastructure or sends a message, the cost of error changes. Sensitive actions need a preview, confirmation, scope limits and a log that reconstructs which user asked for what, which source the system consulted and which operation it performed.

The transferable skill is to evaluate an enterprise assistant as a chain of permissions, retrieval, generation and action. The useful question is not merely “does it answer well?” but “can I trace the answer, prevent it from seeing forbidden material and stop an action before it causes harm?” That test remains useful when the product name or underlying model changes.

What the log should retain

A useful enterprise answer should make its path reconstructable. The minimum log includes the user's identity and role, query, retrieved documents and versions, displayed answer, proposed action, confirmation and external-system outcome. Sensitive data should be protected, but not at the cost of the traceability needed to investigate an error. Without a document version, a citation that is correct today may look false after an update.

Evaluation must include abstention. Prepare questions that have no answer in the repository and others whose sources conflict. A quality system should not reward answering everything; it should reward recognising missing evidence, exposing disagreement and escalating to a person when appropriate. In an organisation, a justified refusal is often safer than an invented but fluent synthesis.

Finally, assign an owner to every connector and action. Someone must review synchronisation, permissions, incidents and retirement. AI does not remove that responsibility; it adds a new interface over systems that already had owners, controls and obligations.

Before deployment, the team should agree which queries remain out of scope, such as final legal advice, personnel decisions or irreversible changes. That boundary belongs in permissions and review, not only a warning. The most important control may be withholding a tool the assistant should never use.