Memox: AI agent development, from customer widget to usage billing

This is AI agent development on a product we own: Memox connects a business’s own documents and website to an AI agent, a website widget, a human handover path and usage billing. Memox is Autonomous Technologies’ own product. We build and operate it for businesses that want an AI agent to answer questions from their documents and website, capture a lead and bring a human into the conversation when needed.

That job extends beyond the answer a visitor sees. A sales representative needs the conversation in a usable dashboard. The agent needs access to the right business information. The product needs to record consumption and charge for it. Our engineering work connected those responsibilities across the customer widget, backend, operator tools and billing.

We stopped renting the part of the product that decides what to say, and built it ourselves so we could change it. The earlier product called a third-party no-code flow tool to run conversations. As the roadmap grew, we needed direct control over the agent’s tools, streaming behaviour, state and instrumentation. The change was a product decision about what Memox needed to own.

An in-house agent layer landed in August 2025. It uses LangGraph and Azure OpenAI, with retrieval over ingested documents and tools for tasks such as pricing lookup, calendar booking and human handover. Tracing makes an agent’s behaviour available for investigation when an answer or tool call needs attention. The old flow tool was replaced as the reasoning layer, although references to it remain in the codebase rather than being fully removed.

Workflow: Business knowledge → AI agent → Human handover → Usage billing.

The agent is only half the product. The other half is the widget a visitor sees, the dashboard a sales representative works in and the billing that charges for use. The website widget was rebuilt in TypeScript and placed inside a Shadow DOM to isolate its styles from the customer’s page. It supports lead capture and an ongoing conversation without asking each customer to redesign their site around the component. Bundle-size checks keep the cost of embedding it visible during development.

On the operator side, the dashboard brings together live conversations, assignment, presence and handover. A Stripe-backed wallet records usage and checks balances around billable features, followed by subscription support. Duplicate payment notifications are handled deliberately so a repeated webhook does not simply credit the wallet twice.

The same engineers have carried this product from March 2024 through September 2026, so a change crossing the agent, billing and dashboard has someone who understands all three. Usama, lead backend engineer, and Abdullah, founding engineer, carried the core engineering continuity, with founder-led product work and contributions from the wider Autonomous team. The work began in March 2024 and continued through September 2026.

That continuity matters when a change crosses boundaries. Adding a billable capability can touch the agent, usage records, customer permissions and dashboard at once. Pull-request review, backend coverage checks and preview environments give the team places to inspect those interactions before release. Voice and mobile work then extend the same product foundation rather than starting unrelated systems.

As of September 10, 2026, Memox has two paying customers and roughly $50k ARR. These figures describe the product’s scale as of September 10, 2026.

The relevant lesson for a store considering automation is how much sits around the AI itself. Useful answers depend on controlled access to business information, a clear human handover and reliable connections to the systems that finish the job. Memox gives us direct operating experience with those decisions.